Paper deep dive

Decoding the Disaster: Multi-Task Geospatial Reasoning with Vision-Language Models and Crowdsourced Imagery for Disaster Mapping

Wenping Yin, Fabian Desuer, Ziqi Liu, Naixia Mou, Weijia Li, Pedram Ghamisi, Xiao Xiang Zhu, Hao Li

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 100

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 10/2/2026, 8:29:35 AM

Summary

The paper introduces GRDisaster, a multi-task framework leveraging Vision-Language Models (VLMs) for disaster mapping using crowdsourced imagery. It addresses the challenge of geolocalizing unstructured Volunteered Geographic Information (VGI) by combining deterministic and probabilistic cross-view geolocalization with remote sensing imagery (RSI) and street-view imagery (SVI). The framework utilizes a new benchmark dataset, PhotoMappers, containing 26,340 images. It employs spatial reasoning indicators to validate cross-view matches and assess disaster damage severity, providing interpretable outputs for emergency response.

Entities (9)

GRDisaster · framework · 98%PhotoMappers · dataset · 97%VLMs · technology · 96%RSI · data-type · 95%SVI · data-type · 95%VGI · data-type · 95%Hao Li · person · 92%Wenping Yin · person · 92%GeoAI · field · 90%

Relation Signals (8)

GRDisaster → builton → PhotoMappers

confidence 97% · GRDisaster is built on a newly curated benchmark dataset derived from PhotoMappers

GRDisaster → performs → damage assessment

confidence 96% · It introduces two sets of spatial reasoning indicators for... disaster damage assessment.

GRDisaster → performs → Cross-view geolocalization

confidence 96% · The framework combines deterministic and probabilistic cross-view geolocalization

PhotoMappers → contains → VGI

confidence 95% · comprising 26,340 images organized into human-validated volunteered geographic information (VGI)

PhotoMappers → contains → RSI

confidence 95% · RSI cross-view triplets covering multiple disaster events

PhotoMappers → contains → SVI

confidence 95% · street-view imagery (SVI)–RSI cross-view triplets

GRDisaster → uses → VLMs

confidence 95% · GRDisaster is built on a newly curated benchmark dataset... to examine the potential of vision-language models (VLMs)

VGI → complements → RSI

confidence 90% · Crowdsourced imagery provides timely, fine-grained, street-level observations for disaster mapping, complementing conventional remote sensing imagery (RSI)

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Crowdsourced imagery provides timely, fine-grained, street-level observations for disaster mapping, complementing conventional remote sensing imagery (RSI) during emergency response. However, such imagery is often unstructured, spatially ambiguous, and lacks reliable geographic metadata, making manual geolocalization and interpretation labor-intensive and difficult to scale. This work proposes a multi-task Geospatial Reasoning Disaster mapping framework, namely GRDisaster, to examine the potential of vision-language models (VLMs) in understanding, geolocalizing, and reasoning over crowdsourced disaster imagery. GRDisaster is built on a newly curated benchmark dataset derived from PhotoMappers, comprising 26,340 images organized into human-validated volunteered geographic information (VGI), street-view imagery (SVI), RSI cross-view triplets covering multiple disaster events from 2018 to 2024. The framework combines deterministic and probabilistic cross-view geolocalization with multi-view fusion to associate VGI images with georeferenced SVI and RSI. It introduces two sets of spatial reasoning indicators for cross-view geolocalization validation and disaster damage assessment. These indicators use structural, environmental, and global-scene cues to validate cross-view correspondences and visually observable damage evidence with expert-verified annotations to assess disaster severity, improving the interpretability of VLM outputs. To our knowledge, this study provides the first systematic investigation and unified evaluation framework for examining how VLM-based spatial reasoning can transform crowdsourced disaster imagery into actionable geospatial artificial intelligence (GeoAI) through cross-view geolocalization validation, interpretable spatial reasoning, and damage-aware severity assessment.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

96,978 characters extracted from source content.

Expand or collapse full text

SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE1 Decoding the Disaster: Multi-Task Geospatial Reasoning with Vision-Language Models and Crowdsourced Imagery for Disaster Mapping Wenping Yin, Fabian Deuser, Ziqi Liu, Naixia Mou, Weijia Li, Member, IEEE, Pedram Ghamisi, Senior Member, IEEE, Xiao Xiang Zhu, Fellow, IEEE, Hao Li, Member, IEEE, Abstract—Crowdsourcedimageryprovidestimely,fine- grained, street-level observations for disaster mapping, comple- menting conventional remote sensing imagery (RSI) during emer- gency response. However, such imagery is often unstructured, spatially ambiguous, and lacks reliable geographic metadata, making manual geolocalization and interpretation labor-intensive and difficult to scale. This work proposes a multi-task Geospatial Reasoning Disaster mapping framework, namely GRDisaster, to examine the potential of vision–language models (VLMs) in understanding, geolocalizing, and reasoning over crowdsourced disaster imagery. GRDisaster is built on a newly curated bench- mark dataset derived from PhotoMappers, comprising 26,340 images organized into human-validated volunteered geographic information (VGI)–street-view imagery (SVI)–RSI cross-view triplets covering multiple disaster events from 2018 to 2024. The framework combines deterministic and probabilistic cross-view geolocalization with multi-view fusion to associate VGI images with georeferenced SVI and RSI. It introduces two sets of spatial reasoning indicators for cross-view geolocalization validation and disaster damage assessment. These indicators use structural, environmental, and global-scene cues to validate cross-view correspondences and visually observable damage evidence with expert-verified annotations to assess disaster severity, improving the interpretability of VLM outputs. To our knowledge, this study provides the first systematic investigation and unified evaluation framework for examining how VLM-based spatial reasoning can transform crowdsourced disaster imagery into actionable geospatial artificial intelligence (GeoAI) through cross- view geolocalization validation, interpretable spatial reasoning, This work was supported by the Start-Up Grant (SUG) project “Geospatial Artificial Intelligence for Climate Resilient Urban Environment” from the National University of Singapore (E-109-00-0036-01), and the MoE Tier1 project ”Assessing Urban Flood Resilience against Climate Extreme with GeoAI in Southeast Asia”. (Corresponding author: Hao Li) Wenping Yin is with the College of Geodesy and Geomatics, Shandong University of Science and Technology, Qingdao 266590, China (email: wenpingyin2019@gmail.com). Fabian Deuser is with the Professorship of Big Geospatial Data Manage- ment, School of Engineering and Design, Technical University of Munich, Munich, 85521, Bavaria, Germany (email: fabian.deuser@tum.de). Ziqi Liu is with the State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China (email: lzq677@whu.edu.cn). Naixia Mou is with the College of Geodesy and Geomatics, Shandong University of Science and Technology, Qingdao 266590, China (email: mounaixia@163.com). Weijia Li is with the Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China. (email: liwei- jia@sz.tsinghua.edu.cn). Pedram Ghamisi is with Helmholtz-Zentrum Dresden-Rossendorf, Respon- sible AI Group, 09599 Freiberg, Germany and also with the University of Iceland, Faculty of Electrical and Computer Engineering, Reykjavik 102, Iceland (e-mail: p.ghamisi@hzdr.de). Xiao Xiang Zhu is with the Chair of Data Science in Earth Observation, Technical University of Munich, 80333 Munich, Germany and with the Munich Center for Machine Learning, 80333 Munich, Germany (email: xiaoxiang.zhu@tum.de). Hao Li is with the Department of Geography, National University of Singapore, Singapore 117568, Singapore (email: hao.li@nus.edu.sg). and damage-aware severity assessment. The findings provide timely guidance for developing scalable, interpretable, and re- sponsible GeoAI-assisted disaster mapping systems. Index Terms—Vision-language model, geolocalization, crowd- sourcing, geospatial reasoning, disaster response, GeoAI I. INTRODUCTION T IMELY and accurate disaster mapping is essential for emergency response, damage assessment, and resource allocation [1–3]. During major disasters, crowdsourced im- agery such as volunteered geographic information (VGI) pro- vides rich street-level observations that complement remote sensing imagery (RSI), particularly when satellite observations are delayed, obstructed by clouds, or insufficient to capture fine-grained local damage [4–6]. However, these images are often unstructured, noisy, spatially ambiguous, and lacking reliable geographic metadata, making manual geolocalization and interpretation labor-intensive and difficult to scale [7, 8]. This creates a major bottleneck in disaster response, where timely geolocalization, interpretation, and validation of visual evidence are critical for transforming scattered observations into actionable geospatial artificial intelligence (GeoAI). Effective disaster response requires georeferenced situa- tional awareness and timely damage assessment, but crowd- sourced imagery collected during early response windows often lacks reliable coordinates due to degraded global nav- igation satellite system signals, missing metadata, or inaccu- rate user-provided geolocations [9–11]. Although satellite and aerial imagery provides broad spatial coverage, it often cannot capture fine-grained street-level damage to buildings, roads, in- frastructure, and vegetation. Cross-view geolocalization helps bridge this gap by anchoring ground-level VGI observations to georeferenced SVI and RSI references, recovering coordinates needed for mapping, routing, and operational decision-making [12, 13]. However, disaster scenes are difficult to match because post-event imagery may differ substantially from reference data due to collapse, flooding, debris, vegetation loss, and road blockage. Reliable analysis must go beyond appearance similarity and reason over persistent spatial cues, damage indicators, and changes in road topology, building layout, landmarks, and vegetation patterns. Vision–language models (VLMs) provide a promising path- way to address these challenges in disaster mapping [14–16]. High-quality post-event annotations are often scarce, delayed, and unevenly distributed across regions and disaster types, limiting the scalability of supervised training. VLMs can inter- pret heterogeneous visual and textual evidence through open- vocabulary and instruction-guided reasoning, making them arXiv:2610.00302v1 [cs.CV] 28 Sep 2026 SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE2 (a) (b) Fig. 1. Examples from the PhotoMappers dataset: (a) crowdsourced VGI and (b) corresponding RSI across different disaster types. suitable for unfamiliar disaster scenes [17, 18]. For geolo- calization, they can compare candidate matches, identify per- sistent features, and reason about spatial relationships across ground-level and overhead views [19]. For damage assessment, they can recognize affected objects, estimate damage severity, and describe observed changes in their geographic context [20, 21]. Despite these capabilities, the potential of VLMs for interpretable geolocalization and damage assessment in disaster mapping remains largely underexplored. This paper proposes GRDisaster, a multitask VLM-based Geospatial Reasoning framework for Disaster mapping. GRDisaster is built on a newly curated PhotoMappers dataset comprising human-validated crowdsourced disaster VGI from multiple events between 2018 and 2024. Representative VGI images and corresponding RSI across different disaster types are shown in Fig. 1. For cross-view geolocalization, GRDisas- ter matches VGI images with candidate RSI and uses VLM- based reasoning to evaluate visual appearance, spatial layout, and landmark cues, producing geolocalization decisions and explanations. It also classifies damage severity and explains the supporting visual evidence. The major contributions of GRDisaster are summarized as follows: • To establish a human-validated VGI–SVI–RSI benchmark dataset covering multiple disaster events for cross-view geolocalization and damage assessment. • To formulate cross-view validation as a structured spatial- reasoning task using appearance, spatial-layout, and land- mark evidence while providing interpretable rationales. • To integrate cross-view geolocalization with explainable VLM-based reasoning for match verification and damage severity classification within a unified workflow. The remainder of this paper is organized as follows. Section 2 reviews related work on GeoAI and multimodal remote sensing for disaster response, VLM-based geospatial reason- ing, and crowdsourced multitask disaster mapping. Section 3 presents the design of GRDisaster, including cross-view geolocalization, VLM-based spatial validation of VGI–SVI– RSI correspondences, and VGI-based disaster damage severity assessment. Section 4 presents experiments on the curated PhotoMappers dataset, covering cross-view geolocalization performance and VLM-based spatial reasoning for geolocal- ization validation and disaster damage severity classification. Section 5 discusses the main findings, limitations, and broader implications for disaster mapping and GeoAI. Section 6 con- cludes the paper and summarizes the potential of GRDisaster for scalable, trustworthy, and interpretable disaster mapping. I. RELATED WORK A. GeoAI and multimodal RS in disaster response GeoAI has become a key foundation for disaster response by enabling spatially explicit learning, prediction, and decision support from heterogeneous geospatial data [22]. Compared with manual mapping or rule-based image analysis, GeoAI methods can extract disaster-related information from RSI, including inundated areas, damaged buildings, blocked roads, burned regions, and disrupted infrastructure. Recent deep learning methods, including convolutional neural networks, object-based semantic change detection, and transformer- based architectures, have improved rapid damage assessment and emergency response. Benchmarks such as xBD have promoted automated building geolocalization and damage classification using pre- and post-disaster satellite imagery [23]. Deep object-based change detection methods further demonstrate the potential of integrating image segmentation, object geolocalization, and semantic damage interpretation in post-disaster scenarios [24]. These studies show that GeoAI can transform Earth observation data into actionable disas- ter information, but single-source overhead imagery remains limited in capturing fine-grained, street-level, and context- dependent damage conditions. Multimodal RS has become important for disaster response because different sensing modalities provide complementary spatial, spectral, temporal, and semantic information. Optical SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE3 satellite imagery offers intuitive visual interpretation and high- resolution land-surface details, whereas synthetic aperture radar (SAR) provides observations under cloud cover and adverse weather conditions, which is valuable during floods, hurricanes, and storms. Unmanned aerial vehicle imagery provides flexible and fine-grained local observations, while street-view imagery and crowdsourced VGI capture human- scale disaster impacts that may not be visible from over- head views. Crowdsourcing has been widely discussed as a complement to conventional RS for collecting timely and locally detailed geospatial observations [8, 25]. Multi-source multimodal disaster-response studies have also shown that combining satellite, aerial, ground-level, and social media data can improve situational awareness across preparedness, response, recovery, and resilience planning [26, 27]. Social media-based geoinformation harvesting further highlights the role of community RS in complementing traditional Earth observation systems [4]. However, multimodal integration remains challenging because disaster data are often noisy, incomplete, temporally misaligned, and unevenly distributed across geolocations and events. Crowdsourced images may contain rich local evidence but often lack reliable geographic metadata, making them difficult to integrate directly into operational disaster maps. Recent GeoAI-assisted disaster mapping studies have moved from single-task image interpretation toward integrated frame- works that combine geolocation recovery, cross-view match- ing, damage severity assessment, and scene understanding. This direction is particularly relevant for multimodal disaster response, where overhead RSI, street-view references, and crowdsourced ground-level observations need to be interpreted jointly. Cross-view geolocalization methods can anchor un- structured disaster photos to georeferenced RS and street-view references [1, 7, 28, 29]. Language and multimodal models have also been used to exploit implicit geoinformation from disaster-related multimodal data [6, 12]. By linking visual evidence with semantic and spatial cues, these models can support interpretable analysis of buildings, roads, vegetation, landmarks, and disaster-related damage patterns. B. VLMs and geospatial reasoning VLMs have recently become an important direction for con- necting visual perception with natural language understanding [30, 31]. Early models such as CLIP learn transferable visual representations by aligning images and text at large scale, enabling zero-shot recognition and image–text retrieval with- out task-specific labels [32]. Later multimodal models further extended this idea from representation learning to instruction- following and visual dialogue. For example, Flamingo intro- duced few-shot visual language learning across interleaved image-text inputs [33], while BLIP-2 connected frozen image encoders with large language models through a lightweight querying transformer. These models show strong potential for open-vocabulary visual interpretation, but their original design is mainly based on natural images and general web data, which makes direct transfer to geospatial applications non-trivial. Remote sensing and geospatial scenes introduce additional challenges for VLMs [34]. Unlike natural images, RS imagery often contains large spatial extents, small and densely dis- tributed objects, multi-scale patterns, nadir-view geometry, and domain-specific land-cover semantics. A model must not only recognize objects, but also reason about spatial layout, relative position, scene function, and cross-view consistency. To ad- dress this gap, recent studies have developed remote-sensing- specific VLMs and datasets. RemoteCLIP adapts vision– language pretraining to RS imagery and improves zero-shot classification, retrieval, and object counting in remote sens- ing scenarios [35]. RS5M and GeoRSCLIP further construct a large-scale RS image-text dataset and domain-pretrained vision–language model for remote sensing tasks [36]. Recent work such as DOFA-CLIP further explores multimodal vision– language foundation models for Earth observation by aligning heterogeneous EO modalities with language [37]. These efforts indicate that domain-specific image-text alignment is impor- tant for transferring VLMs from generic visual recognition to geospatial understanding. More recent work has begun to move from image-level recognition toward geospatial reasoning and grounded in- terpretation. GeoChat, for example, supports conversational understanding, region-level reasoning, visual grounding, and referring tasks for high-resolution RS images [38]. EarthGPT further explores multimodal large language models for multi- sensor RS image comprehension, including optical, SAR, and infrared imagery [39]. TerraScope extends this direction toward pixel-grounded visual reasoning for Earth observa- tion, emphasizing fine-grained spatial evidence in geospatial reasoning tasks [40]. These models suggest that VLMs can support more flexible geospatial interpretation by linking vi- sual evidence with textual prompts, spatial cues, and task- specific reasoning requirements. However, disaster scenarios remain particularly challenging because post-event imagery often contains occlusion, debris, damage-induced appearance changes, and missing or unreliable geographic metadata. C. Crowdsourcing and multi-task disaster mapping Crowdsourced data complement conventional remote sens- ing by providing timely, street-level information from disaster- affected areas [12, 41]. Social media posts and VGI imagery can reveal local flooding, structural damage, road accessi- bility, and infrastructure disruption that may be difficult to observe from overhead imagery alone. Crisis-mapping efforts following events such as the Haiti earthquake demonstrated the value of volunteered data when official information is incomplete or delayed [42]. Their operational use, however, re- mains constrained by uneven coverage, uncertain data quality, duplicate content, weak annotations, and missing or inaccurate geographic metadata. Addressing these limitations is essential for converting crowdsourced imagery into reliable information for disaster preparedness, response, and recovery. Recent disaster mapping research has also moved toward multi-task frameworks that jointly address several operational needs rather than treating each task separately. This trend is consistent with broader Earth observation benchmark efforts, where diverse datasets, tasks, and foundation models are organized for systematic evaluation [14]. Recent extreme-event SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE4 D a t a e x a m p l e s VGI (Ground view) Task 1 Cross-view geolocalization Query: VGI Image Reference: RSIOptional mid-view: SVI Deterministic matching Probabilistic matching s (VGI, SVIi) s (VGI, RSIi) s (VGI, RSIi) Cross-view similarity Scorei=αs(VGI, SVIi)+βs(VGI, RSIi)+γs(SVIi, RSIi) Multi-source joint scoring Top-k candidate ranking Scorei=αs(VGI, SVIi)+βs(VGI, RSIi)+γs(SVIi, RSIi) p(lᵢ|x)=e^Scoreᵢ / Σⱼ e^Scoreⱼ Multi-source joint scoring Probability estimation Output: Cross-view geolocalization results (Top-1/ Top-k) Score Fusion & Decision I n p u t M e t h o d s Task 2 VLM-based spatial reasoning I n p u t VGI-SVI-RSI candidate pairs from task 1 Building and structureRoad and transport Vegetation and terrain Objects and facilities Global scene AppearanceSpatial layoutLandmark cue M e t h o d s Sreason = w1·Sapp + w2·Slayout + w3·Slandmark where Sapp, Slayout,Slandmark Task 3 VLM-based damage severity classification I n p u t Spatial reasoning score P(match|VGI,SVI,RSI) = softmax(Sreason) Matching Probability VGI images with different disaster damage severity levels Flooding and waterBuilding and structural Road and access Vegetation and natural M e t h o d s Infrastructure and facilityDebris and scene Indicator-level and overall severity Flooding and water impact The ground is wet from rain, but there is no evidence of standing water or flooding. SVI (Intermediate)RSI (Aerial view) Output: Reasoning-supported VGI–SVI–RSI matching pairs VGI (Ground view)SVI (Intermediate)RSI (Aerial view) D a t a e x a m p l e s VGI (Ground view) Task 1 Cross-view geolocalization Query: VGI Image Reference: RSIOptional mid-view: SVI Deterministic matching Probabilistic matching s (VGI, SVIi) s (VGI, RSIi) s (VGI, RSIi) Cross-view similarity Multi-source joint scoring Top-k candidate ranking p(lᵢ|x)= e^Scoreᵢ /Σⱼ e^Scoreⱼ Multi-source joint scoring Probability estimation Output: Cross-view geolocalization results (Top-1/ Top-k) Score fusion & decision I n p u t M e t h o d s Task 2 VLM-based spatial reasoning I n p u t VGI-SVI-RSI candidate pairs from task 1 Building and structure Road and transport Vegetation and terrain Objects and facilities Global scene Appearance Spatial layoutLandmark cue M e t h o d s Sreason = w1·Sapp + w2·Slayout + w3·Slandmark where Sapp, Slayout, Slandmark Task 3 VLM-based damage severity classification I n p u t Spatial reasoning score P(match|VGI, SVI, RSI) = softmax (Sreason) Matching Probability VGI images with different disaster damage severity levels Flooding and water Building and structural Road and access Vegetation and natural M e t h o d s Infrastructure and facility Debris and scene SVI (Intermediate)RSI (Aerial view)VGI (Ground view)SVI (Intermediate)RSI (Aerial view) Scorei =αs(VGI, SVIi)+βs(VGI, RSIi)+γs(SVIi, RSIi) Scorei =αs(VGI, SVIi)+βs(VGI, RSIi)+γs(SVIi, RSIi) Output: Reasoning-supported matching pairs Flooding and water impact The ground is wet from rain, but there is no evidence of standing water or flooding. D a t a e x a m p l e s VGI (Ground view) Task 1 Cross-view geolocalization Query: VGI Image Reference: RSIOptional mid-view: SVI Deterministic matching Probabilistic matching s (VGI, SVIi) s (VGI, RSIi) s (VGI, RSIi) Cross-view similarity Multi-source joint scoring Top-k candidate ranking p(lᵢ|x)= e^Scoreᵢ /Σⱼ e^Scoreⱼ Multi-source joint scoring Probability estimation Output: Cross-view geolocalization results (Top-1/ Top-k) Score fusion & decision I n p u t M e t h o d s Task 2 VLM-based spatial reasoning I n p u t VGI-SVI-RSI candidate pairs from task 1 Building and structure Road and transport Vegetation and terrain Objects and facilities Global scene Appearance Spatial layoutLandmark cue M e t h o d s Sreason = w1·Sapp + w2·Slayout + w3·Slandmark where Sapp, Slayout, Slandmark Task 3 VLM-based damage severity classification I n p u t Spatial reasoning score P(match|VGI, SVI, RSI) = softmax (Sreason) Matching Probability VGI images with different disaster damage severity levels Flooding and water Building and structure Road and access Vegetation and nature M e t h o d s Infrastructure and facility Debris and scene SVI (Intermediate)RSI (Aerial view)VGI (Ground view)SVI (Intermediate)RSI (Aerial view) Scorei =αs(VGI, SVIi)+βs(VGI, RSIi)+γs(SVIi, RSIi) Scorei =αs(VGI, SVIi)+βs(VGI, RSIi)+γs(SVIi, RSIi) Output: Reasoning-supported matching pairs Flooding and water impact The ground is wet from rain, but there is no evidence of standing water or flooding. VGI (Ground view)SVI (Intermediate)RSI (Aerial view) D a t a e x a m p l e s VGI (Ground view) Task 2 VLM-based spatial reasoning I n p u t VGI-SVI-RSI candidate pairs from task 1 Building and structure Road and transport Vegetation and terrain Objects and facilities Global scene Appearance Spatial layoutLandmark cue M e t h o d s Task 3 VLM-based damage severity classification I n p u t Spatial reasoning score Matching probability VGI images with different damage severity levels Flooding and water Building and structure Road and access Vegetation and nature M e t h o d s Infrastructure and facility Debris and scene SVI (Intermediate)RSI (Aerial view)VGI (Ground view)SVI (Intermediate)RSI (Aerial view) Output: Reasoning-supported matching pairs Flooding and water impact The ground is wet from rain, but there is no evidence of standing water or flooding. Non-matchMatch P Reasoning scor e w₃w₂w₁ Landmar kLayoutAppear ance D a t a e x a m p l e s Task 1 Cross-view geolocalization Query: VGI Image Reference: RSIOptional mid-view: SVI Deterministic matching Probabilistic matching s (VGI, SVIi) s (VGI, RSIi) s (VGI, RSIi) Cross-view similarity Top-k candidate ranking Multi-source joint scoring Probability estimation Output: Cross-view geolocalization results (Top-1/ Top-k) Score fusion & decision I n p u t M e t h o d s Fused score Merged evidence Multi-source joint scoring s (VGI, SVIi)s (VGI, RSIi)s (VGI, RSIi) Score1 Score2Score3 Task 1 Cross-view geolocalization Query: VGI Image Reference: RSIOptional mid-view: SVI Deterministic matching Probabilistic matching s (VGI, SVIi) s (VGI, RSIi) s (VGI, RSIi) Cross-view similarity Top-k candidate ranking Multi-source joint scoring Probability estimation Output: Cross-view geolocalization results (Top-1/ Top-k) Score fusion & decision I n p u t M e t h o d s Fused score Merged evidence Multi-source joint scoring s (VGI, SVIi)s (VGI, RSIi)s (VGI, RSIi) Score1 Score2Score3 Task 2 VLM-based cross-view geolocalization reasoning I n p u t VGI-SVI-RSI candidate pairs from Task 1 M e t h o d s Spatial reasoning score Matching probability Output: Matching results and reasoning rationale VLM-based disaster damage classification and reasoning I n p u tVGI images with different severity levels M e h o d s Cross-view geolocalization: deterministic vs. probabilistic Query: VGI Image Optional mid-view: SVI Deterministic matching s (VGI, SVIi) s (VGI, RSIi) s (VGI, RSIi) Cross-view similarity Top-k candidate ranking Multi-source joint scoring Probability estimation Output: Cross-view geolocalization results (Top-1/ Top-k) Score fusion & decision I n p u t M e t h o d s Fused score Multi-source joint scoring s (VGI, SVIi) s (VGI, RSIi) s (VGI, RSIi) Score1 Score2Score3 Reference: RSI Probabilistic matching VGI (Ground view)SVI (Intermediate)RSI (Aerial view) Merged evidence Task 3Task 1 D a t a e x a m p l e s VGI (ground view) RSI (a SVI (intermediate) erial view) VGI (ground view) RSI (a SVI (intermediate) erial view) Building / structureRoad / transport Vegetation / terrain Objects / facilities Global scene Appearance Spatial layoutLandmark cue Flooding and water Building and structure Road and access Vegetation and nature Infrastructure and facility Debris and scene P MatchNon-match Appearance LayoutLandmark w1w3w2 Reasoning score Fig. 2. Overview of the three interconnected tasks in the GRDisaster framework using the VGI–SVI–RSI benchmark: deterministic and probabilistic cross-view geolocalization, and explainable VLM-based reasoning for match verification and damage severity classification. benchmarks further highlight the need to evaluate foundation models across multiple disaster-related tasks, data sources, and event types [43]. In practical response workflows, geoloca- tion recovery, damage assessment, scene interpretation, and evidence validation are closely connected: an image is more useful when its geolocation is known, its damage level is estimated, and its visual evidence can be explained in relation to surrounding geospatial context. Large-scale datasets such as xBD have supported building geolocalization and damage classification from satellite imagery [23], while cross-view disaster geolocalization studies have explored how ground- level disaster images can be matched with street-view and remote sensing references [1, 6]. For crowdsourced data, this multi-task perspective is particularly important because the same image may simultaneously contain clues for ge- olocalization, damage severity classification, and scene-level reasoning. Therefore, combining crowdsourced imagery with multi-task disaster mapping can help transform scattered visual observations into more reliable and interpretable geospatial intelligence for emergency response. I. METHOD A. Overview As shown in Fig. 2, GRDisaster consists of three intercon- nected tasks: cross-view geolocalization, VLM-based cross- view reasoning and match validation, and VLM-based disaster damage severity classification. Task 1 retrieves the most likely georeferenced SVI–RSI candidates for each VGI query image through deterministic or probabilistic cross-view matching and multi-source score fusion. Task 2 uses VLMs to validate and interpret retrieved VGI–SVI–RSI candidate pairs through structured spatial reasoning based on visual appearance, spa- tial layout, and landmark evidence. It produces reasoning- supported matching decisions and interpretable spatial reason- ing scores. Task 3 assesses disaster severity from VGI images based on six visually observable damage indicators, producing indicator-level scores, an overall severity level, and a brief rationale. The three tasks provide a unified framework for geolocalization, explainable match verification, and damage- aware disaster scene understanding. B. Deterministic and probabilistic cross-view geolocalization We evaluate deterministic and probabilistic cross-view ge- olocalization methods on the disaster-oriented PhotoMappers benchmark. Each geolocation contains co-located multi-view observations, including ground-level disaster imagery (VGI), street-view panoramas (SVI), and RSI. Given a VGI query, the objective is to recover its geolocation by exploiting com- plementary visual evidence across these views. As shown in Fig. 3, deterministic methods retrieve geolocations from a georeferenced reference gallery, whereas probabilistic meth- ods estimate a conditional geographic distribution and its SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE5 Deterministic cross-view geolocalization Probabilistic cross-view geolocalization Cross-view imagery SVI RSI VGI Shared embeddingSimilarity scoring Candidate ranking (Top-k) Retrieved geolocation Spherical noise prior Conditional RFM on S 2 Predicted geolocation and uncertainty Distance metrics (both) Recall@K (retrieval only) Evaluation Conditional location density image condition c image samples p ( x0 | c ) x 1 = ε , t = 1 t = 1 t = 0.9 t = 0 Fig. 3. Overview of deterministic and probabilistic cross-view geolocalization. spatial uncertainty. The retrieval settings include VGI→RSI, VGI→SVI, SVI→RSI, and multi-view configurations that jointly use VGI, SVI, and RSI. For deterministic geolocalization, we evaluate SAIG-D [44], TransGeo [45], and Sample4Geo [7] with ConvNeXt-B [46], DINOv2 [47], and DINOv3 [48] visual backbones. The DINO- based models include base and large variants, denoted as DINOv2-B/L and DINOv3-B/L. Each image is encoded into a shared embedding space, and the similarity between a query embedding L c and a candidate reference embedding L i r is measured using cosine similarity: S(L c ,L i r ) = L c · L i r ∥L c ∥L i r ∥ (1) Candidate references are ranked by their similarity scores, and the highest-ranked reference provides the predicted geolo- cation. Sinkhorn-based re-ranking is optionally applied to the DINOv3-B and DINOv3-L fusion models under the one-to- one matching setting. Independent ranking may produce hub references that receive high similarity scores from multiple queries. Sinkhorn–Knopp normalization [49, 50] converts the query–reference similarity matrix into a doubly stochastic assignment matrix, suppressing hub references and promoting globally consistent matches without additional training. For probabilistic geolocalization, we evaluate Riemannian flow-matching (RFM) conditioned on DINOv2-L and Street- CLIP representations, following the ProbGLC [51, 52]. Unlike retrieval-based methods that select a geolocation from a finite reference gallery, RFM estimates a conditional distribution p φ (x| c), where x∈ S 2 denotes a geolocation on the spherical Earth surface and c is the visual condition. Both the ground- truth geolocation x 0 and the noise ε lie on S 2 [53]. The intermediate geolocation is defined along a spherical geodesic through the exponential and logarithmic maps: x t = exp x t−1 β(t) log x t−1 (ε) (2) where log x t−1 maps ε to the tangent space at x t−1 , exp x t−1 maps the tangent vector back to S 2 , and β(t) controls the interpolation. The conditional vector field φ(x t | c) is trained to approximate the target velocity field v(x t ) by minimizing L RFM = E x 0 ,c,ε,t h ∥φ(x t | c)− v(x t )∥ 2 x t i (3) where ∥·∥ x t is the norm induced by the Riemannian metric on the tangent space at x t . An ordinary differential equation solver is initialized at a random geolocation x 1 = ε and integrates the learned vector field backward from t = 1 to t = 0. Exponential-map updates keep the trajectory on the spherical manifold, while repeated sampling produces multiple plausible geolocations and approximates the conditional geographic distribution. The spatial ambiguity of this distribution is summarized by the geolocalizability score [51, 52]: Localizability(c) = Z S 2 p(x n | c) log 2 P(x n | c)dx n (4) where x n denotes a sampled geolocation and the score cor- responds to the negative entropy of the predicted distribution. A more concentrated distribution indicates stronger geographic evidence and lower spatial ambiguity, whereas a dispersed distribution reflects limited, ambiguous, or visually degraded cues in the disaster image. Deterministic and probabilistic methods are compared us- ing mean and median geolocalization errors and accuracy at distance thresholds of 1, 25, 200, 750, and 2,500 km. Retrieval- based methods are additionally evaluated using Recall@K at several ranking levels. These metrics provide a consistent basis SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE6 TABLE I EVALUATION INDICATORS FOR VLM-BASED INTERPRETIVE REASONING OF CROSS-VIEW GEOLOCALIZATION RESULTS. No. Indicator nameEvaluation focus Dimension Appearance Spatial layout Landmark cue 1 Building type and scaleBuilding and structure: Building appearance consistency✓ 2 Overall building appearanceBuilding and structure: Building appearance consistency✓ 3 Building position relative to nearby roads or open spaceBuilding and structure: Building spatial-position consistency✓ 4 Building position and shape correspondence in SVI–RSIBuilding and structure: Building spatial-position consistency✓ 5 Recognizable building outlineBuilding and structure: Corresponding building landmark✓ 6 Distinctive building appearance or visible nameBuilding and structure: Corresponding building landmark✓ 7 Road type and widthRoad and transport: Road appearance consistency✓ 8 Road markings and boundary featuresRoad and transport: Road appearance consistency✓ 9 Road direction and connectivityRoad and transport: Road spatial-structure consistency✓ 10 Road position relative to surrounding elementsRoad and transport: Road spatial-structure consistency✓ 11 Distinctive intersection, bridge, or road curveRoad and transport: Corresponding road landmark✓ 12 Traffic lights, road signs, or road markersRoad and transport: Corresponding road landmark✓ 13 Vegetation coverage and typeVegetation and terrain: Natural-surface appearance consistency✓ 14 Ground surface characteristicsVegetation and terrain: Natural-surface appearance consistency✓ 15 Vegetation position relative to roads or buildingsVegetation and terrain: Natural-element spatial consistency✓ 16 Waterbody, shoreline, or open-space positionVegetation and terrain: Natural-element spatial consistency✓ 17 Prominent trees, forest belts, or distinctive vegetationVegetation and terrain: Corresponding natural-terrain landmark✓ 18 Coastline, river, lake, or slope outlineVegetation and terrain: Corresponding natural-terrain landmark✓ 19 Utility poles, streetlights, and cablesObjects and facilities: Stable facility appearance consistency✓ 20 Road signs, fences, and guardrailsObjects and facilities: Stable facility appearance consistency✓ 21 Facility position relative to roads or buildingsObjects and facilities: Facility spatial-distribution consistency✓ 22 Local facility grouping and arrangementObjects and facilities: Facility spatial-distribution consistency✓ 23 Large facilities such as antennas, water towers, or pylonsObjects and facilities: Corresponding facility landmark✓ 24 Recognizable signs such as storefront signs, billboards, or gas-station signs Objects and facilities: Corresponding facility landmark✓ 25 Scene functional typeGlobal scene: Overall scene-type consistency✓ 26 Spatial density and opennessGlobal scene: Overall scene-type consistency✓ 27 Overall building–road–vegetation relationshipGlobal scene: Overall scene-structure consistency✓ 28 Consistency between the VGI visible scene and the SVI–RSI overall scene Global scene: Overall scene-structure consistency✓ 29 Large building complex, community, or commercial areaGlobal scene: Corresponding regional background cue✓ 30 Coastline, river, bridge, or regional backgroundGlobal scene: Corresponding regional background cue✓ for comparing methods that rank georeferenced candidates with those that directly estimate geographic coordinates. C. VLM-based cross-view geolocalization reasoning 1) Design of cross-view geolocalization reasoning indica- tors: We design an indicator-based spatial reasoning frame- work to evaluate the quality and validity of VLM-generated explanations for VGI–SVI–RSI cross-view geolocalization re- sults. The indicators were developed through iterative visual inspection of disaster images across different disaster types and scene categories. Domain experts discussed and organized recurring visual characteristics and spatial relationships into three dimensions: appearance, spatial layout, and landmark cue (Table I). Repeated reviews were conducted to identify missing cues and improve indicator coverage. These dimensions are further applied to five semantic classes, including building and structure, road and transport, vegetation and terrain, objects and facilities, and global scene. The resulting 30 indicators provide an initial, extensible set of stable visual and spatial cues. They reflect cues commonly used in human cross-view interpretation and support the assessment of spatial consis- tency between a VGI image and its candidate reference pair. We further design a structured prompting scheme to guide VLMs in reasoning about and validating VGI–SVI–RSI cross- view geolocalization results. Following recent studies on struc- tured prompting and visual reasoning evaluation [12, 54], the prompt consists of four main components: task definition, output format, scoring scheme, and evaluation indicators. It defines the matching task between a post-disaster VGI image and its corresponding SVI–RSI reference pair, specifies the required structured output format, and introduces a three- level scoring scheme to quantify indicator-level reasoning quality. The 30 indicators are incorporated into the prompt to guide structured evaluation. To reduce temporal inconsistency, disaster-related damage information is excluded from the matching evidence, and the evaluation focuses on relatively stable geographic and visual characteristics shared across views. A containment-based matching criterion is used, where the VGI scene only needs to be reasonably supported by part of the SVI–RSI reference pair rather than requiring com- plete one-to-one correspondence. This design better reflects practical cross-view geolocalization scenarios with viewpoint differences and incomplete visual overlap. 2) Reasoning and validating geolocalization results: Spa- tial reasoning quality is evaluated using a criterion-based scoring scheme. For each spatial indicator, the generated reasoning is assigned a score of correct (2), partially correct (1), or incorrect (0), based on whether the indicator provides consistent visual evidence of a match between the VGI image and the corresponding SVI–RSI reference pair. A score of 2 indicates that the reasoning is directly supported by visual evidence from the VGI image and at least one reference view, without clear contradictions from other views. A score of 1 represents reasoning that is partially supported but affected by incomplete, weak, or ambiguous cross-view evidence. A score of 0 indicates that the relevant feature is not visible or matched across the compared views, or that the reasoning is unsupported, contradicted by visual evidence, spatially inconsistent, or hallucinated. All VLM-based annotation and evaluation are conducted in a zero-shot setting without task-specific examples. Gemini- 3.6-Flash [55] is used as the scalable annotation and VLM-as- SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE7 TABLE I IMAGE-BASED INDICATORS FOR VLM-BASED DISASTER DAMAGE SEVERITY REASONING AND CLASSIFICATION. Indicator name Damage severity LowMediumHigh Flooding and water impact Shallow or geolocalized water with limited impact on visible objects. Clear flooding affects roads, buildings, vehicles, or public spaces. Deep water, strong flow, storm surge, or widespread inundation. Building and structural damage Buildings are mostly intact, with no clear or only minor damage. Partial damage is visible, such as doors, windows, walls, roofs, or water intrusion. Severe structural damage, collapse, heavy inundation, or loss of building function. Road and access disruption Roads or paths remain mostly usable, with minor debris or shallow water. Roads are partially blocked or difficult to access due to water, debris, vehicles, or fallen trees. Roads are impassable or heavily disrupted by deep water, large debris, collapse, or severe surface damage. Vegetation and natural damage Vegetation remains intact, with only small branches or geolocalized disturbance. Fallen trees, broken branches, or damaged vegetation partially affect roads, buildings, or open spaces. Severe vegetation damage, large fallen trees, or widespread natural disturbance blocks access or damages objects. Infrastructure and facility damage Facilities are mostly intact, with only minor or geolocalized disturbance. Some facilities are damaged or disrupted, such as poles, signs, fences, utilities, or public equipment. Critical infrastructure or multiple facilities are severely damaged, submerged, collapsed, or nonfunctional. Debris and scene disruption Limited debris or scattered small objects, with the scene mostly organized. Noticeable debris or displaced objects cause local disorder across the scene. Extensive debris, widespread destruction, or chaotic conditions affect scene elements. a-Judge model for its strong performance and cost efficiency, ranking second on the English-language Vision Arena Leader- board [56]. The generated annotations serve as a scalable alternative to exhaustive manual labeling. For human-in-the- loop validation, 20% of the samples are stratified according to predicted damage severity levels and independently reviewed by two human experts. The experts examine whether the pre- dicted spatial matching decisions and corresponding reasoning are visually grounded in the VGI, SVI, and RSI evidence. Disagreements are resolved through discussion, providing ad- ditional quality control for identifying unsupported explana- tions, ambiguous cases, and systematic reasoning errors across VLMs. LLM/VLM-as-a-Judge frameworks with human-in- the-loop verification have been increasingly explored in mul- timodal understanding tasks [57, 58]. The evaluation of both spatial reasoning in cross-view ge- olocalization and disaster damage severity assessment follows the same evaluation framework. Qwen3.6-27B [59], GPT-4o [60], and InternVL3-8B [61] are evaluated based solely on the visual evidence contained in each VGI image. Indicator-level scores and overall predictions are assessed using agreement rate and quadratic weighted kappa (κ w ). Compared with con- ventional Cohen’s kappa, quadratic weighted kappa considers the ordinal relationships among scoring categories, assign- ing smaller penalties to near-miss disagreements and larger penalties to more substantial deviations [62]. The quadratic weighted kappa is defined as follows [62]: κ w = 1− P K i=1 P K j=1 w ij O ij P K i=1 P K j=1 w ij E ij (5) where K is the number of scoring categories, O ij denotes the observed agreement matrix between VLM and human scores, and E ij represents the expected agreement matrix esti- mated from the marginal distributions. The quadratic weight- ing matrix w ij is defined as: w ij = (i− j) 2 (K− 1) 2 (6) where i and j represent the assigned score categories. This weighting scheme penalizes larger disagreements more severely while assigning lower penalties to adjacent categories. The observed agreement matrix is calculated as: O ij = N X n=1 1(ˆs n = i)1(s n = j),(7) where N is the number of evaluated samples, ˆs n is the VLM-generated score for the n-th sample, s n is the cor- responding human-verified score, and 1(·) is an indicator function. The expected agreement matrix is obtained from the marginal distributions: E ij = n VLM i · n Human j N (8) where n VLM i and n Human j denote the numbers of samples assigned to categories i and j by the VLM and human annotators, respectively. The scoring categories are defined as C = 0, 1, 2, corresponding to the three-level damage severity or reason- ing quality scores. The quadratic weighted kappa therefore measures the agreement between VLM predictions and human annotations while accounting for both chance agreement and the ordinal distance between different scoring categories. D. VLM-based disaster damage reasoning and classification We classify the damage severity of collected VGI images into three levels: low, medium, and high, using six visually observable indicators: flooding and water impact, building and structural damage, road and access disruption, vegetation and natural damage, infrastructure and facility damage, and debris and scene disruption. As shown in Table I, these indicators capture different aspects of disaster impacts and enable both indicator-level assessment and overall severity classification. All VLM-based annotation generation and evaluation are conducted in a zero-shot setting without task-specific ex- amples. We evaluate Qwen3.6-27B [59], GPT-4o [60], and InternVL3-8B [61] using only the visual information in each VGI image. For each image, the models generate scores for individual damage indicators, an overall severity level, and reasoning explanations. As described above, Gemini-3.6-Flash [55] serves as the scalable VLM-as-a-Judge for annotation, while 20% of samples are stratified by predicted severity and SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE8 SAIG-D ViT TransGeo DeiT-S Sample4Geo ConvNeXt-B Sample4Geo DINOv2-B Sample4Geo DINOv3-B Sample4Geo DINOv2-L Sample4Geo DINOv3-L 0 10 20 30 40 50 60 70 Score (%) 3.13 3.82 6.32 15.09 17.26 22.84 26.59 8.26 10.88 15.32 33.71 35.48 42.77 50.06 12.07 16.06 22.38 41.97 44.31 52.45 58.43 41.00 21.81 29.39 50.46 52.51 60.25 66.00 (a) VGI - RSI R@1R@5R@10R@1% SAIG-D ViT TransGeo DeiT-S Sample4Geo ConvNeXt-B Sample4Geo DINOv2-B Sample4Geo DINOv3-B Sample4Geo DINOv2-L Sample4Geo DINOv3-L 0 10 20 30 40 50 60 70 Score (%) 3.13 5.18 8.26 20.39 23.18 26.59 29.04 9.57 14.35 19.65 37.98 40.43 44.59 48.75 13.50 19.59 26.94 45.79 49.43 52.96 58.49 39.29 25.34 33.20 53.64 55.64 59.85 64.64 (b) VGI - SVI R@1R@5R@10R@1% DINOv2-BDINOv3-BDINOv2-LDINOv3-L 0 20 40 60 80 100 Score (%) 83.60 81.49 89.12 89.01 94.76 94.82 96.92 97.32 97.21 96.75 98.18 98.35 98.18 97.95 98.58 98.80 (c) SVI - RSI R@1R@5R@10R@1% DINOv2-BDINOv3-BDINOv2-LDINOv3-L 0 20 40 60 80 100 Score (%) +4.44 +3.36 +4.44 +4.73 +2.85 +0.00 +0.57 +3.25 +3.36 +0.00 +0.00 +2.34 +2.74 +0.00 +0.00 +2.45 (d) Fusion vs Fusion + Sinkhorn R@1R@5R@10R@1%FusionFusion + Sinkhorn Fig. 4. Deterministic cross-view geolocalization performance under different retrieval and fusion settings: (a) VGI–RSI retrieval, (b) VGI–SVI retrieval, (c) SVI–RSI retrieval, and (d) VGI–SVI–RSI fusion with and without Sinkhorn re-ranking. independently verified by two human experts in a human-in- the-loop process. The experts assess the same indicators and determine whether the predicted scores, severity levels, and reasoning are supported by visible image evidence. Follow- ing the human-aligned verification strategy used for spatial reasoning evaluation, two experts independently assess VLM outputs and resolve disagreements through discussion to obtain final verified labels [58, 63]. Agreement rate and quadratic weighted kappa (κ w ) [62] measure consistency between VLM predictions and human verification while accounting for the ordinal relationships among severity categories. IV. EXPERIMENTS A. Dataset overview and experimental settings The experiments are conducted on a curated cross-view PhotoMappers dataset. The original collection contained more than 27,575 crowdsourced disaster images, including duplicate copies of the same image with identical latitude and longitude coordinates but different image dimensions. After retaining one record from each duplicate group, approximately 13,513 candidate VGI records remained. Georeferenced SVI and RSI were subsequently collected for these locations, producing more than 9,000 candidate cross-view groups. Irrelevant, uninformative, and visually unusable samples were further removed through manual inspection. The final dataset contains 26,340 images organized into 8,780 validated VGI–SVI–RSI triplets for evaluating cross-view geolocalization, spatial rea- soning, and disaster damage assessment, as shown in Fig. 1. For Task 1, cross-view geolocalization, we adopt a standard 8:2 train–test split. The training set contains 7,024 geolocation groups, corresponding to 21,072 images in the three-view VGI–SVI–RSI setting or 14,048 images in each two-view setting. The test set contains 1,756 groups, corresponding to 5,268 three-view images or 3,512 two-view images. Its reference gallery comprises 1,756 distinct geolocations for each reference-view configuration. Samples are sorted by latitude and longitude and then uniformly assigned to the two subsets, ensuring that no geolocation is shared between the training and test sets. The VLM-based spatial reasoning and validation tasks are evaluated zero-shot, without task-specific examples or model fine-tuning. B. Cross-view geolocalization performance We evaluate cross-view geolocalization performance on the proposed PhotoMappers dataset under different retrieval settings, including VGI-to-RSI, VGI-to-SVI, and multi-view fusion. Recall at different ranks is used for evaluation, where R@K denotes the proportion of queries whose correct refer- ence appears among the top-K retrieved candidates. As shown in Fig. 4, existing methods exhibit relatively limited geolocal- ization capability when directly handling the large visual do- main gap between ground-level and aerial/satellite views. Con- ventional retrieval-based approaches, including SAIG-D [44], TransGeo [45], and Sample4Geo [7] with ConvNeXt-B [46], achieve relatively low R@1 performance in the VGI→RSI setting, obtaining only 3.13%, 3.82%, and 6.32%, respectively. This indicates that their visual representations struggle to capture discriminative cross-view semantic correspondences under substantial viewpoint and appearance variations. Replacing the backbone with stronger foundation models brings consistent improvements across all evaluation settings. With DINOv2 [47] and DINOv3 [48] features, Sample4Geo achieves significant gains compared with previous baselines. In the VGI→RSI task, DINOv3-L improves R@1 from 6.32% SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE9 TABLE I COMPARISON OF GEOLOCALIZATION PERFORMANCE BETWEEN GENERATIVE AND RETRIEVAL-BASED APPROACHES. MEDIAN AND MEAN ARE ERRORS IN KILOMETERS, WHILE @1,KM–@2500,KM ARE THE PERCENTAGES OF PREDICTIONS WITHIN EACH DISTANCE THRESHOLD. MethodTypeView Input Median (km) Mean (km) @1km@25km @200km @750km @2500km DINOv3-B×VGI–RSI381.47747.3519.1323.2937.8168.1793.96 DINOv3-L×VGI–RSI218.67525.3528.2533.2048.3576.2596.18 DINOv3-B×VGI–SVI340.37631.4225.1128.5940.4370.3995.10 DINOv3-L×VGI–SVI193.99545.4631.4435.7150.3475.4096.58 DINOv3-B×VGI–SVI–RSI206.96547.8331.7835.9949.4374.9495.90 DINOv3-L×VGI–SVI–RSI65.31390.0841.4045.6258.6081.4498.35 DINOv3-B×VGI–SVI–RSI + Sinkhorn204.06567.1534.4037.9849.6674.9495.73 DINOv3-L×VGI–SVI–RSI + Sinkhorn27.29359.4945.9649.7262.1982.1298.12 RFM DINOv2-L✓None324.17651.500.009.4539.7571.5895.50 RFM StreetCLIP✓None129.34370.660.0610.7660.9386.5698.23 Note:× indicates retrieval-based approaches, while✓ indicates generative approaches. 363, 274, 416, 147 query pred query pred RSI (a) VGI - GT (b) (c)(d) VGI RSI UAV SVI (a) VGI - GT (b) (c) (d) VGI RSI SVI Fig. 5. Heatmaps for (a) and (b) correctly and (c) and (d) incorrectly geolocalized imagery pairs generated by our best-performing DINOv3-L model with VGI–SVI–RSI fusion and Sinkhorn re-ranking, with the corresponding raw VGI ground-truth images included for reference. The first column shows the ground-truth images, and the second column shows the predicted results. to 26.59% and increases R@1% to 66.00%. Similar trends can be observed in the VGI→SVI scenario, where DINOv3- L achieves 29.04% R@1 and 64.64% R@1%. The gains across different retrieval levels indicate that stronger visual representations capture more reliable semantic and structural features for matching disaster images across viewpoints. The SVI→RSI evaluation further highlights the effective- ness of the cross-view representations. Compared with the VGI-based retrieval tasks, SVI and RSI preserve more con- sistent scene structures and geographic patterns, resulting in substantially higher retrieval accuracy. All DINO-based variants perform strongly, with DINOv3-L reaching 89.01% R@1, 97.32% R@5, 98.35% R@10, and 98.80% R@1%. These results show that roads, building configurations, and surrounding spatial layouts can be matched more reliably when the cross-view discrepancy is reduced. The substantial performance gap between VGI-based retrieval and SVI→RSI retrieval further identifies disaster VGI as a key bottleneck in the geolocalization pipeline. Even with DINOv3-L, VGI→RSI and VGI→SVI achieve only 26.59% and 29.04% R@1, re- spectively, compared with 89.01% for SVI→RSI. This gap can be attributed to disaster-induced appearance changes, un- constrained viewpoints, occlusion, limited fields of view, and temporal differences between VGI and reference imagery. For practical deployment, improving the initial linkage between disaster VGI and georeferenced references is therefore critical, motivating multi-view fusion, uncertainty-aware prediction, and additional verification for low-confidence matches. The VGI–SVI–RSI multi-view fusion strategy further im- proves geolocalization accuracy by integrating complemen- tary information from different observation perspectives. As shown in Table I, the fusion model consistently achieves SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE10 higher retrieval performance than individual-view matching. For DINOv3-L, the fusion of VGI, SVI, and RSI improves R@1 over the single-view results to 39.12%, while adding the Sinkhorn-based alignment module further increases R@1 to 43.85%. Similar improvements are observed across other retrieval thresholds, with R@1% increasing from 74.66% to 77.11%. These results suggest that multi-view feature inter- action provides more robust and complementary geographic cues while reducing ambiguity caused by individual image sources. The heatmaps in Fig. 5 show that correct matches rely on consistent structural cues across views, whereas incorrect matches are affected by ambiguous scene features. We further compare retrieval-based methods with generative geolocalization methods on the newly curated PhotoMappers dataset. As shown in Table I, retrieval-based models achieve more stable performance across different distance thresholds. The deterministic DINOv3-L fusion model with Sinkhorn alignment obtains a median geolocalization error of only 27.29 km and reaches 98.12% accuracy within 2500 km. Compared with the generative RFM DINOv2-L and RFM StreetCLIP baselines, the retrieval-based DINOv3-L model provides sub- stantially better fine-grained geolocalization, particularly at strict distance thresholds. It achieves 45.96% accuracy within 1 km, whereas the two RFM-based approaches obtain only 0.00% and 0.06%, respectively. These results indicate that deterministic cross-view retrieval with multi-view fusion and feature alignment is more effective for precise geolocalization on our newly curated disaster dataset. C. Cross-view spatial reasoning and validation results The expert verification results in Table IV show high agreement between the Gemini-generated annotations and the expert assessments. The verification protocol was relatively permissive: experts examined whether the reasoning scores were reasonable and supported by visible cross-view evidence, rather than requiring the generated explanations to repro- duce expert interpretations exactly. Differences in wording or emphasis were accepted when they did not alter the main spatial judgment. Under this criterion, Gemini achieved overall indicator-level agreement values of 0.938 and 0.962 with Experts A and B, with corresponding quadratic weighted kappa values of 0.953 and 0.960. The agreement between the two experts was 0.934, with a κ w of 0.960. Although the evaluation protocol partly accounts for these high values, all expert-validation agreement and κ w results exceeded 0.8, indicating that the annotations were sufficiently reliable for model evaluation after expert quality control. For the best-performing Task 1 results obtained using DINOv3-L, each VLM was required to determine whether the retrieved VGI–SVI–RSI triplet was correctly matched while performing spatial reasoning. GPT-4o achieved the highest matching accuracy at 69.00%, followed by Qwen3.6-27B at 65.15% and InternVL3-8B at 64.20%. These results were obtained through zero-shot annotation without task-specific training and were numerically higher than those of the evalu- ated deterministic and probabilistic cross-view geolocalization methods. GPT-4o also achieved the strongest aggregate spatial reasoning performance, with an overall indicator-level agree- ment of 0.646 and a κ w of 0.489, compared with 0.609 and 0.396 for Qwen3.6-27B and 0.465 and 0.284 for InternVL3- 8B. GPT-4o performed particularly well on building-related indicators, with agreement values ranging from 0.672 to 0.724. Its κ w reached 0.630 for recognizable building outlines, 0.613 for overall building appearance, and 0.610 for building type and scale. It also achieved agreement values of 0.751 and 0.731 for scene functional type and spatial density or openness. Its weaker results were concentrated in waterbody position, terrain outline, large facility structures, and regional background, where κ w was generally below 0.35. Qwen3.6-27B performed relatively well on relational and scene-level indicators. Its highest agreement was obtained for VGI–SVI–RSI scene consistency at 0.721, followed by the scene functional type at 0.708, building–road–vegetation relationship at 0.700, facility grouping arrangement at 0.683, and spatial density and openness at 0.682. It also produced rel- atively high κ w values for traffic markers, utility poles and ca- bles, fences and guardrails, and relationships among buildings, roads, and vegetation. Performance declined for coastline and terrain outline, waterbody and open-space position, and large facility structures. For large facility structures, an agreement of 0.589 was accompanied by a κ w of only 0.112, indicating limited discrimination among less frequent reasoning levels. InternVL3-8B produced lower results across most semantic groups. Its stronger indicators were recognizable building outlines and commercial signs, with agreement values of 0.556 and 0.531 and κ w values of 0.396 and 0.367. Its κ w values for vegetation coverage, waterbody position, terrain outline, scene functional type, and regional background ranged from 0.126 to 0.173. Overall VGI–SVI–RSI scene consistency reached an agreement of 0.407 and a κ w of 0.186, reflecting difficulty in integrating evidence across viewpoints and spatial scales. Structural and relational cues generally provided the most reliable and consistent evidence for validating and understand- ing the cross-view geolocalization results. Building geometry, road organization, distinctive intersections, facility arrange- ments, and recognizable signs remain comparatively stable across ground, street, and aerial views, whereas vegetation, terrain, waterbodies, coastlines, and regional backgrounds are more sensitive to temporal change, viewpoint occlusion, disas- ter disturbance, and scale differences. GPT-4o showed the most balanced performance across appearance, spatial layout, and global-scene indicators, while Qwen3.6-27B remained com- petitive in interpreting relational and scene-level cues. Spatial reasoning therefore provides a useful complementary signal for geolocalization validation, although its reliability depends on the visibility, stability, and distinctiveness of the available cross-view evidence. As shown in Fig. 6, the models generally agree on structural and road cues, whereas vegetation, terrain, and global-scene reasoning produces more partial or incorrect results. Final geolocalization judgments vary across models, also reflecting differences in cross-view evidence integration. D. VLM-based damage reasoning and classification results The expert verification results in Table V support the use of Gemini-generated damage severity annotations. At the SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE11 F l o o d Qwen3.6-27BGPT-4oInternVL3-8B H u r r i c a n e W i l d f i r e E a r t h q u a k e a b c d VGI–SVI–RSI image pair 0 correct 1 correct 2 correct 3 correct 4 correct 5 correct 6 correct Building and structureRoad and transportVegetation and terrain Objects and facilitiesGlobal sceneCVGL results Fig. 6. Spatial reasoning for cross-view geolocalization across disaster types. Three VLMs are compared across building and structure, road and transport, vegetation and terrain, objects and facilities, global scene (circles), and cross-view geolocalization (diamond), with different colors indicating reasoning results. SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE12 TABLE IV AGREEMENT AND QUADRATIC WEIGHTED KAPPA (κ w ) EVALUATION FOR CROSS-VIEW GEOLOCALIZATION REASONING. BOLD THE BEST-PERFORMING RESULTS IN EACH CATEGORY. BOLD THE BEST-PERFORMING RESULTS IN EACH CATEGORY. No. Indicator Abbr. VLM vs Exp. A VLM vs Exp. B Exp. A vs Exp. BQwenGPTInternVL Agr.κ w Agr.κ w Agr.κ w Agr.κ w Agr.κ w Agr.κ w 1Building type and scale0.9540.9640.9910.9850.9570.9750.644 0.434 0.708 0.610 0.537 0.376 2Overall building appearance0.9830.9900.8740.9210.8620.9140.632 0.429 0.702 0.613 0.504 0.364 3Building-road-open space position0.9400.9650.9910.9950.9480.9700.637 0.407 0.694 0.585 0.517 0.355 4Building shape correspondence 0.9890.9890.9680.9770.9740.9850.649 0.415 0.710 0.578 0.522 0.351 5Recognizable building outline0.9910.9900.9910.9900.9940.9970.633 0.426 0.724 0.630 0.556 0.396 6Distinctive building appearance0.9660.9300.9430.9200.9740.9840.587 0.426 0.672 0.558 0.529 0.375 7Road type and width0.8850.9170.9710.9780.8620.8980.604 0.445 0.661 0.591 0.438 0.268 8Road markings and boundaries0.9280.9230.9710.9530.9570.9690.552 0.459 0.578 0.531 0.453 0.292 9Road direction and connectivity 0.9230.9520.9600.9740.8830.9260.596 0.416 0.634 0.534 0.472 0.308 10 Road-element position relationship 0.9600.9730.9970.9980.9630.9750.669 0.443 0.708 0.555 0.462 0.267 11 Distinctive intersections or curves0.9890.9790.9940.9830.9940.9960.530 0.393 0.620 0.545 0.473 0.256 12 Traffic signs and markers0.9800.9540.9830.9560.9910.9940.604 0.481 0.613 0.531 0.498 0.348 13 Vegetation coverage and type0.8220.8210.9940.9940.8170.8140.626 0.379 0.686 0.502 0.407 0.173 14 Ground surface characteristics 0.8370.8450.9680.9710.8220.8430.571 0.407 0.604 0.504 0.418 0.223 15 Vegetation-road-building position0.9680.9770.9970.9980.9710.9790.652 0.432 0.677 0.493 0.437 0.211 16 Waterbody and open-space position0.9430.9600.8740.9030.8170.8660.513 0.324 0.518 0.345 0.399 0.126 17 Prominent vegetation features0.9460.9480.9540.9530.9400.9560.563 0.408 0.576 0.451 0.414 0.212 18 Coastline and terrain outline0.9050.9220.9000.8300.9540.9640.534 0.240 0.573 0.277 0.450 0.142 19 Utility poles and cables0.9340.9520.9970.9920.9370.9590.571 0.472 0.616 0.595 0.448 0.277 20 Road signs, fences, and guardrails0.9890.9930.9970.9980.9910.9950.580 0.465 0.585 0.524 0.474 0.311 21 Facility-road-building position0.9740.9850.9680.9810.9430.9650.637 0.410 0.614 0.414 0.474 0.287 22 Facility grouping arrangement0.9710.9790.9940.9920.9770.9870.683 0.438 0.635 0.401 0.486 0.296 23 Large facility structures0.8710.8550.8620.8420.9910.9900.589 0.112 0.643 0.181 0.510 0.111 24 Recognizable commercial signs0.9340.8370.9310.8340.9910.9950.621 0.380 0.695 0.514 0.531 0.367 25 Scene functional type0.9030.9050.9310.9250.8340.8120.708 0.393 0.751 0.464 0.399 0.159 26 Spatial density and openness0.8850.8890.9970.9970.8830.8850.682 0.385 0.731 0.454 0.392 0.166 27 Building-road-vegetation relationship0.9800.9800.9890.9860.9910.9940.700 0.450 0.687 0.481 0.449 0.228 28 VGI-SVI-RSI scene consistency0.9890.9930.9830.9890.9710.9820.721 0.419 0.681 0.450 0.407 0.186 29 Large building complexes0.8480.9030.9140.8230.8600.9110.517 0.337 0.593 0.487 0.498 0.334 30 Coastline and regional background0.9460.9360.9770.9550.9630.9760.478 0.241 0.489 0.262 0.409 0.128 Building and structure average0.9710.9710.9600.9650.9520.9710.630 0.423 0.702 0.596 0.528 0.370 Road and transport average0.9440.9500.9790.9740.9420.9600.593 0.440 0.636 0.548 0.466 0.290 Vegetation and terrain average 0.9040.9120.9480.9420.8870.9040.577 0.365 0.606 0.429 0.421 0.181 Objects and facilities average0.9460.9340.9580.9400.9720.9820.614 0.380 0.631 0.438 0.487 0.275 Global scene average0.9250.9340.9650.9460.9170.9270.634 0.371 0.655 0.433 0.426 0.200 Appearance dimension average0.9120.9200.9690.9710.8920.9060.617 0.427 0.662 0.539 0.447 0.261 Spatial layout dimension average0.9640.9750.9720.9790.9440.9630.646 0.415 0.656 0.484 0.463 0.262 Landmark cue dimension average0.9380.9250.9450.9090.9650.9760.566 0.344 0.620 0.444 0.487 0.267 Overall indicator-level average0.9380.9400.9620.9530.9340.9490.609 0.396 0.646 0.489 0.465 0.263 indicator level, Gemini achieved an agreement of 0.950 with both Experts A and B, with corresponding κ w values of 0.962 and 0.954. These values are comparable to the expert-to-expert agreement of 0.943 and κ w of 0.951. For overall severity, Gemini achieved agreement values of 0.915 and 0.949 with the two experts and κ w values of 0.924 and 0.954, respectively. The corresponding expert-to-expert values were 0.954 and 0.960. Gemini showed limited directional bias, with under- estimation rates of 0.066 and 0.040 and overestimation rates of 0.020 and 0.011. Agreement across individual indicators ranged from 0.934 to 0.974. The consistency with both experts indicates that the generated annotations are sufficiently reliable for large-scale evaluation after expert quality control. Table VI compares Qwen3.6-27B, GPT-4o, and InternVL3- 8B against the verified Gemini-based annotations. Qwen3.6- 27B and GPT-4o produced similar aggregate indicator-level performance, although their strengths differed across damage types. Qwen3.6-27B obtained an overall indicator-level agree- ment of 0.695 and a κ w of 0.737, whereas GPT-4o achieved 0.686 and 0.738, respectively. GPT-4o performed better for overall severity classification, reaching an agreement of 0.739 and a κ w of 0.688, compared with 0.711 and 0.630 for Qwen3.6-27B. InternVL3-8B was substantially less consistent, with an indicator-level agreement of 0.405 and a κ w of 0.331, SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE13 TABLE V AGREEMENT AND QUADRATIC WEIGHTED KAPPA (κ w ) EVALUATION BETWEEN GEMINI AND TWO EXPERT ANNOTATIONS FOR DISASTER DAMAGE SEVERITY ASSESSMENT WITH 20% EXPERT VERIFICATION. IndicatorLevel VLM vs Expert AVLM vs Expert BExpert A vs Expert B Agr.κ w Agr.κ w Agr.κ w Flooding and water impact Low 0.9660.9540.9320.9060.9320.908 Medium0.9400.9630.9400.9480.9490.953 High0.9400.9640.9320.9570.9740.984 Overall0.9490.9680.9340.9530.9520.964 Building and structural damage Low0.9830.9160.9830.9270.9910.846 Medium0.9490.9480.9570.9310.9150.912 High0.9320.9480.9660.9740.9150.936 Overall0.9540.9620.9690.9670.9400.943 Road and access disruption Low0.9910.9880.9910.9880.9830.976 Medium0.9320.9380.9490.9290.9060.914 High0.9490.9210.9830.9740.9490.920 Overall0.9570.9700.9740.9750.9460.962 Vegetation and natural damage Low0.9660.8610.9910.9620.9740.898 Medium0.9320.9360.9230.8810.9150.898 High 0.9490.9440.9320.9240.9490.942 Overall 0.9490.9500.9490.9330.9460.939 Infrastructure and facility damage Low0.9910.9550.9910.9550.9830.913 Medium0.9150.9060.9150.9020.9150.907 High 0.9150.8970.9060.8860.9230.908 Overall 0.9400.9440.9370.9410.9400.946 Debris and scene disruption Low0.9660.8950.9490.8550.9490.855 Medium0.9230.9180.9060.8460.9150.861 High0.9570.9410.9570.9410.9320.903 Overall0.9490.9570.9370.9350.9320.931 Overall indicator-level agreement0.9500.9620.9500.9540.9430.951 Overall severity-level agreement0.9150.9240.9490.9540.9540.960 Severity underestimation rate 0.066—0.040—0.014— Severity overestimation rate0.020—0.011—0.031— and an overall severity agreement of 0.338 with a κ w of 0.226. The gap suggests that reliable damage assessment requires vi- sual object recognition and the integration of multiple damage cues into an ordered scene-level judgment. At the category-overall level, GPT-4o achieved the strongest results for flooding and water impact, building and structural damage, vegetation and natural damage, and infrastructure and facility damage. Its overall agreement and κ w reached 0.777 and 0.848 for flooding, 0.790 and 0.817 for building damage, and 0.776 and 0.780 for vegetation damage. However, no model consistently performed best across all severity levels. For example, Qwen3.6-27B performed best for medium-level flooding, whereas GPT-4o performed best for low- and high- level flooding. Qwen3.6-27B also achieved the best results for high-level building damage, although GPT-4o performed better for this category overall. Results for road disruption and in- frastructure damage were similarly mixed between agreement and κ w . Qwen3.6-27B showed a clearer advantage for debris and scene disruption across all severity levels, obtaining an overall agreement of 0.663 and a κ w of 0.657, compared with 0.528 and 0.527 for GPT-4o. Infrastructure and facility damage remained difficult for both leading models, with overall agree- ment close to 0.60, possibly because utility failures, facility functionality, and partially occluded infrastructure are difficult to assess from ground-level disaster photographs. The error distributions reveal different tendencies among the evaluated models. Qwen3.6-27B had a relatively low underestimation rate of 0.066 but an overestimation rate of 0.223, indicating a tendency to assign higher severity levels when errors occurred. GPT-4o showed a more balanced profile, with underestimation and overestimation rates of 0.155 and 0.106, respectively. InternVL3-8B produced an overestima- tion rate of 0.620 despite an underestimation rate of only 0.041, indicating frequent assignment of higher-than-reference severity levels and helping to explain its weak agreement and kappa scores. GPT-4o was the most reliable model for overall severity classification, while Qwen3.6-27B provided competitive indicator-level consistency and performed partic- ularly well for debris-dominated scenes. The remaining errors mainly involved ambiguous cases with overlapping damage indicators, unclear functional impacts, or severity distinctions requiring evidence beyond the image. The examples in Fig. 7 show how the three VLMs interpret the same damage evidence. In the flooded scene, all models identify severe water impact and restricted road access but differ in their assessments of buildings, vegetation, and in- frastructure. In the structural-collapse scene, they consistently recognize collapsed walls and heavy debris while interpreting road accessibility and facility damage differently. The main damage type is generally recognized, whereas secondary indi- SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE14 TABLE VI AGREEMENT AND QUADRATIC WEIGHTED KAPPA EVALUATION OF VLMS FOR DISASTER DAMAGE SEVERITY ASSESSMENT USING GEMINI AS A HUMAN ANNOTATION SUBSTITUTE. BOLD THE BEST-PERFORMING RESULTS IN EACH CATEGORY. IndicatorLevelQwen Agr.Qwenκ w GPT Agr.GPTκ w InternVL Agr.InternVLκ w Flooding and water impact Low0.6610.6010.7170.6670.3940.306 Medium 0.7890.8510.7720.8440.5830.462 High 0.8020.8420.8770.9150.6270.427 Overall0.7640.8310.7770.8480.5510.465 Building and structural damage Low0.8290.5810.9270.5910.4230.175 Medium0.7350.7570.7640.8100.2390.161 High0.7550.7490.6820.7370.5540.122 Overall 0.7600.7880.7900.8170.3410.262 Road and access disruption Low 0.5880.4000.7120.6200.2990.219 Medium 0.6910.5840.5750.6060.5390.227 High0.7670.4990.7420.5890.7170.050 Overall0.6830.6690.6400.6930.5220.382 Vegetation and natural damage Low0.7480.5970.8800.7000.4590.235 Medium0.6620.6920.7510.7790.3010.278 High 0.7380.6600.7110.7030.4750.296 Overall 0.6950.7180.7760.7800.3700.352 Infrastructure and facility damage Low0.7270.3770.7940.3330.2890.024 Medium0.5400.4460.5610.5700.1610.058 High0.6560.3830.4810.4230.5630.054 Overall0.6030.5820.6050.6030.2670.169 Debris and scene disruption Low0.5830.4490.5560.3860.3330.160 Medium0.6330.5320.5120.4060.2890.075 High0.8450.6400.5410.1600.7200.124 Overall0.6630.6570.5280.5270.3830.266 Overall indicator-level agreement0.6950.7370.6860.7380.4050.331 Overall severity-level agreement0.7110.6300.7390.6880.3380.226 Severity underestimation rate0.066—0.155—0.041— Severity overestimation rate0.223—0.106—0.620— cators and functional impacts produce more disagreement. V. DISCUSSION Cross-view geolocalization provides the spatial anchor for interpreting crowdsourced disaster imagery. Unlike conven- tional queries, disaster VGI may depict flooded roads, col- lapsed buildings, debris-covered surroundings, or disturbed vegetation that no longer resembles the available SVI and RSI references. Stronger foundation-model representations im- prove the extraction of persistent geographic features, while multi-view fusion reduces dependence on any single source. VGI retains local objects and visible damage, SVI provides street-level context and surrounding structures, and RSI rep- resents broader layouts that may remain recognizable despite local disturbance. Retrieval-based methods are better suited to fine-grained geolocalization when the reference gallery has adequate coverage, whereas probabilistic models represent plausible geographic regions and uncertainty when the avail- able cues are incomplete [51, 52]. Disaster geolocalization can therefore benefit from combining precise reference matching with uncertainty-aware estimation rather than relying on a single geolocalization mechanism. Retrieval scores cannot explain why a candidate geolocation is plausible or reveal whether high similarity reflects genuine geographic correspondence. This limitation becomes important when different places contain similar buildings, road patterns, vegetation, or urban forms. Although recent remote-sensing VLMs have demonstrated strong capabilities in grounded conversation and region-level image interpretation [35, 38], their potential for reasoning over cross-view geolocaliza- tion results remains insufficiently studied. GRDisaster applies VLM-based spatial reasoning to the retrieved VGI–SVI–RSI correspondences and organizes the supporting evidence into a dedicated set of cross-view reasoning indicators. Building ge- ometry, road organization, intersections, facility arrangements, environmental features, and global-scene characteristics are evaluated separately rather than compressed into one similarity score. Structural and relational cues generally provide stronger evidence because their relative configurations can remain identifiable after substantial appearance changes, whereas veg- etation, terrain, waterbodies, and regional backgrounds are more sensitive to season, scale, viewing direction, and disas- ter disturbance. The indicator structure makes disagreements traceable to specific cues and helps distinguish geographically meaningful correspondence from broad semantic resemblance. The second set of reasoning indicators addresses disaster damage severity through flooding, structural damage, road access, vegetation disturbance, infrastructure condition, and debris. Existing disaster assessment studies have largely con- centrated on change detection and building-level damage clas- SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE15 F l o o d AccuratePartialInaccurate Qwen3.6-27BGPT-4o InternVL3-8B H u r r i c a n e a b W i l d f i r e E a r t h q u a k e c d Qwen3.6-27BGPT-4oInternVL3-8B Flooding and water impactBuilding and structural damageRoad and access disruption Vegetation and natural damageInfrastructure and facility damageDebris and scene disruption Fig. 7. Examples of VLM-based disaster damage reasoning across six dimensions. Three VLMs are compared in flooding and water impact, building and structural damage, road and access disruption, vegetation and natural damage, infrastructure and facility damage, and debris and scene disruption. The circles in each group follow this order, with different colors indicating accuracy. sification from pre- and post-event satellite imagery [23]. GRDisaster instead extends VLM reasoning from determin- ing whether images represent the same place to interpreting multiple forms of damage visible at that place. The models generally recognize dominant damage patterns, while greater disagreement appears in secondary effects and functional impacts. Standing water may be visible without establishing whether a road remains passable, and exterior damage may not reveal whether a building or facility remains operational. Such ambiguities explain why infrastructure functionality, accessibility, and adjacent severity levels are harder to assess than directly observable flooding or collapse. Differences in overestimation and underestimation also indicate that reliable severity assessment depends on calibration as well as visual recognition. Applying VLM-based reasoning to both geolo- calization validation and damage assessment distinguishes GRDisaster from studies that address these tasks indepen- dently. Expert verification remains necessary when indicators conflict or judgments depend on evidence beyond the image. The two indicator systems should be regarded as an initial and extensible organization of relevant evidence rather than a complete or definitive taxonomy. Future studies may refine, prioritize, or expand them for different disasters, geographic contexts, and operational needs. The two indicator systems connect geographic credibility with disaster impact while preserving the different roles of the underlying methods. Roads, buildings, waterbodies, veg- etation, and facilities may support a geolocation match and also reveal how the event has altered access, structure, and environmental conditions. Deterministic retrieval supplies pre- cise reference-based anchoring, probabilistic geolocalization describes spatial uncertainty, and VLM reasoning examines the evidence behind correspondence and severity judgments. Crowdsourced imagery can thus be interpreted as a georef- erenced record of local conditions rather than as an isolated visual observation. The outputs should nevertheless be treated as structured decision support rather than autonomous conclu- sions, especially when reference imagery is outdated, gallery coverage is uneven, important objects are occluded, or the scene differs from the regions and disaster types represented in PhotoMappers. Single-image reasoning also has limited ability to determine hidden structural damage, infrastructure functionality, and the temporal progression of an event. Considering model capability, availability, inference cost, and representation across different developers, the evaluation includes Qwen3.6-27B [59], GPT-4o [60], and InternVL3- 8B [61]. The findings provide evidence for the effectiveness of GRDisaster and VLM-based spatial reasoning, but they are limited to the evaluated models and versions rather than VLMs as a whole, particularly because commercial and open- source models evolve rapidly and may respond differently to prompt design. Future evaluation should cover a broader range of model families, parameter scales, and deployment settings, with explicit comparisons of reasoning accuracy, computational cost, latency, and prompt sensitivity. Pre- and post-event image sequences, road networks, elevation, weather records, and infrastructure databases could help separate per- sistent geographic structures from temporary disaster effects. SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE16 Additional regions, event categories, image sources, reference densities, and temporal gaps are also needed to assess ge- ographic transfer and disaster-type generalization. Calibrated indicator-level confidence, explicit recognition of insufficient visual evidence, lightweight models for field deployment, and selective expert review could further improve the reliability and practical value of VLM-assisted disaster mapping. VI. CONCLUSIONS This study introduced GRDisaster, a multi-task geospatial reasoning framework for converting crowdsourced disaster im- agery into georeferenced and interpretable information. Built on the newly curated PhotoMappers dataset of 8,780 human- validated VGI–SVI–RSI cross-view triplets (26,340 images), it combines deterministic and probabilistic cross-view geolocal- ization with explainable VLM-based reasoning for match ver- ification and damage severity classification. The results show that foundation-model features, complementary cross-view in- formation, and optional re-ranking improve geographic match- ing. Two structured indicator systems support interpretable spatial validation and damage assessment, while expert ver- ification confirms the reliability of the generated annotations. Tests with Qwen3.6-27B, GPT-4o, and InternVL3-8B demon- strate meaningful zero-shot reasoning, although performance varies across models and evidence types. GRDisaster provides a scalable and expert-verifiable framework for GeoAI-assisted disaster mapping, but its generalizability remains limited by reference coverage, limited visual evidence from individual images, and the range of evaluated disasters and models. Future work should evaluate broader regions, disaster types, and VLMs and incorporate temporal and auxiliary geospatial information to improve generalization and reliability. REFERENCES [1] H. Li, F. Deuser, W. Yin, X. Luo, P. Walther, G. Mai, W. Huang, and M. Werner, “Cross-view geolocalization and disaster map- ping with street-view and vhr satellite imagery: A case study of hurricane ian,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 220, p. 841–854, 2025. [2] S. Zhao, I. Prapas, Z. Xiong, I. Karasante, I. Papoutsis, G. Camps-Valls, and X. X. Zhu, “Causal graph neural networks for robust wildfire forecasting across geographic shifts,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 236, p. 654–667, 2026. [3] K. R. Shahi, M. M. Kuglitsch, J. B. Bove, M. Ronco, P. Ghamisi, Y. Sun, G. Duca, M. V. Gargiulo, A. Berlin, J.-P. J ̈ ap ̈ ol ̈ a et al., “Governing generative ai in disaster risk management,” 2026. [4] X. X. Zhu, Y. Wang, M. Kochupillai, M. Werner, M. H ̈ aberle, E. J. Hoffmann, H. Taubenb ̈ ock, D. Tuia, A. Levering, N. Jacobs et al., “Geoinformation harvesting from social media data: A community remote sensing approach,” IEEE Geoscience and Remote Sensing Magazine, vol. 10, no. 4, p. 150–180, 2023. [5] W. Yin, Y. Xue, Z. Liu, H. Li, and M. Werner, “Llm-enhanced disaster geolocalization using implicit geoinformation from multimodal data: A case study of hurricane harvey,” Interna- tional Journal of Applied Earth Observation and Geoinforma- tion, vol. 137, p. 104423, 2025. [6] W. Yin, F. Deuser, Z. Liu, J. Wei, X. Luo, M. Werner, H. Li, and Y. Xue, “Triple-objective cross-view geolocalization of disaster- related vgi: the case of hurricane ian,” International Journal of Geographical Information Science, vol. 40, no. 1, p. 194–216, 2026. [7] F. Deuser, K. Habel, and N. Oswald, “Sample4geo: Hard neg- ative sampling for cross-view geo-localisation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, p. 16 847–16 856. [8] E. Saralioglu and O. Gungor, “Crowdsourcing in remote sens- ing: A review of applications and future directions,” IEEE Geoscience and Remote Sensing Magazine, vol. 8, no. 4, p. 89–110, 2020. [9] H. Huang, Y. Yu, T. Chen, N. Yokoya, J. Li, and A. Plaza, “Sar and social-media-based change detection with dual-threshold fusion for flood inundation mapping,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, p. 15 278–15 290, 2025. [10] W. Zheng, H. Huang, J. Li, and Z. Liu, “Weakly supervised multimodal fusion of remote sensing and social media data for urban flood mapping,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2026. [11] X. Xing, B. Yu, C. Kang, B. Huang, J. Gong, and Y. Liu, “The synergy between remote sensing and social sensing in urban studies: Review and perspectives,” IEEE Geoscience and Remote Sensing Magazine, vol. 12, no. 1, p. 108–137, 2024. [12] J. Ye, H. Lin, L. Ou, D. Chen, Z. Wang, Q. Zhu, C. He, and W. Li, “Where am i? cross-view geo-localization with natural language descriptions,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, p. 5890– 5900. [13] H. Li, C. Xu, W. Yang, L. Mi, H. Yu, H. Zhang, and G.- S. Xia, “Unsupervised multiview uav image geolocalization via iterative rendering,” IEEE Transactions on Geoscience and Remote Sensing, vol. 63, p. 1–15, 2025. [14] Z. Xiong, F. Zhang, Y. Wang, Y. Shi, and X. X. Zhu, “Earthnets: Empowering artificial intelligence for earth observation,” IEEE Geoscience and Remote Sensing Magazine, 2024. [15] X. Weng, C. Pang, and G.-S. Xia, “Vision-language modeling meets remote sensing: Models, datasets, and perspectives,” IEEE Geoscience and Remote Sensing Magazine, vol. 13, no. 3, p. 276–323, 2025. [16] X. X. Zhu, Z. Xiong, Y. Wang, A. J. Stewart, K. Heidler, Y. Wang, Z. Yuan, T. Dujardin, Q. Xu, and Y. Shi, “On the foundations of earth foundation models,” Communications Earth & Environment, 2026. [17] P. Ghamisi, W. Yu, A. Marinoni, C. M. Gevaert, C. Persello, S. Selvakumaran, M. Girotto, B. P. Horton, P. Rufin, P. Hostert et al., “Responsible artificial intelligence for earth observation: Achievable and realistic paths to serve the collective good,” IEEE Geoscience and Remote Sensing Magazine, 2025. [18] Y. Li, W. Li, B. Dang, Y. Wang, W. Chen, L. Wang, B. Yang, and Y. Zhang, “Full-scope vectorization of geographical elements from large-size remote sensing imagery,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026. [19] W. Yin, F. Deuser, J. Jiang, Z. Liu, X. Luo, Z. Zheng, M. Werner, H. Li, and Y. Xue, “Sagingeo: A space-aerial-ground integrated framework for vgi geolocalization in multi-disaster scenarios,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2026. [20] C. Wen, Y. Lin, X. Qu, N. Li, Y. Liao, X. Li, and H. Lin, “Remote sensing retrieval-augmented generation: Bridging re- mote sensing imagery and comprehensive knowledge with a multimodal dataset and retrieval-augmented generation model,” IEEE Geoscience and Remote Sensing Magazine, 2026. [21] Z. Liu, D. Zhao, B. Yuan, and Z. Jiang, “Rescueadi: Adap- tive disaster interpretation in remote sensing images with autonomous agents,” IEEE Transactions on Geoscience and Remote Sensing, vol. 63, p. 1–14, 2025. [22] K. Janowicz, S. Gao, G. McKenzie, Y. Hu, and B. Bhaduri, “Geoai: spatially explicit artificial intelligence techniques for geographic knowledge discovery and beyond,” International SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE17 Journal of Geographical Information Science, vol. 34, no. 4, p. 625–636, 2020. [23] R. Gupta, B. Goodman, N. Patel, R. Hosfelt, S. Sajeev, E. Heim, J. Doshi, K. Lucas, H. Choset, and M. Gaston, “Creating xbd: A dataset for assessing building damage from satellite imagery,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2019, p. 10–17. [24] Z. Zheng, Y. Zhong, J. Wang, A. Ma, and L. Zhang, “Build- ing damage assessment for rapid disaster response with a deep object-based semantic change detection framework: From natural disasters to man-made disasters,” Remote Sensing of Environment, vol. 265, p. 112636, 2021. [25] Y. Xu, P. Ghamisi, and Q. Weng, “Towards realistic remote sensing dataset distillation with discriminative prototype-guided diffusion,” arXiv preprint arXiv:2601.15829, 2026. [26] Y. Li, Y. Wang, L. Yu, B. Dang, G. Xu, Z. Zhong, Y. Wu, X. Guo, K. Wu, Z. Li et al., “Learning to reason over multi- granularity knowledge graph for zero-shot urban land-use map- ping,” Remote Sensing of Environment, vol. 330, p. 114961, 2025. [27] W. Yin, Z. Liu, N. Mou, W. Li, D. Hong, and H. Li, “Disastertd: Disaster toponym disambiguation using multimodal llms and cross-view geolocalization,” IEEE Transactions on Geoscience and Remote Sensing, 2026. [28] J. Ye, J. He, X. Zhang, Y. Lin, H. Lin, C. He, and W. Li, “Satellite image synthesis from street view with fine-grained spatial textual guidance: A novel framework,” IEEE Geoscience and Remote Sensing Magazine, vol. 13, no. 3, p. 395–414, 2025. [29] B. Tang, R. Lu, X. Yang, Y. Li, Y. Li, D. Zhang, and S. Chen, “R2ploc: A region-to-point uav visual geo-localization frame- work leveraging hierarchical semantic representation,” IEEE Transactions on Geoscience and Remote Sensing, 2025. [30] H. Lin, D. Hong, S. Ge, C. Luo, K. Jiang, H. Jin, and C. Wen, “Rs-moe: A vision–language model with mixture of experts for remote sensing image captioning and visual question answer- ing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 63, p. 1–18, 2025. [31] W. Li, Y. Li, Y. Zhou, D. Song, J. Zhang, Z. Wei, and A.-A. Liu, “Knowledge-guided prompt learning for tropical cyclone intensity estimation,” IEEE Transactions on Geoscience and Remote Sensing, vol. 63, p. 1–13, 2025. [32] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning. PmLR, 2021, p. 8748–8763. [33] J.-B. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Has- son, K. Lenc, A. Mensch, K. Millican, M. Reynolds et al., “Flamingo: a visual language model for few-shot learning,” Advances in neural information processing systems, vol. 35, p. 23 716–23 736, 2022. [34] D. Hong, C. Li, X. Li, G. Camps-Valls, and J. Chanus- sot, “Foundation models in remote sensing: Evolving from unimodality to multimodality,” IEEE Geoscience and Remote Sensing Magazine, 2026. [35] F. Liu, D. Chen, Z. Guan, X. Zhou, J. Zhu, Q. Ye, L. Fu, and J. Zhou, “Remoteclip: A vision language foundation model for remote sensing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, p. 1–16, 2024. [36] Z. Zhang, T. Zhao, Y. Guo, and J. Yin, “Rs5m and georsclip: A large-scale vision-language dataset and a large vision-language model for remote sensing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, p. 1–23, 2024. [37] Z. Xiong, Y. Wang, W. Yu, A. J. Stewart, J. Zhao, N. Lehmann, T. Dujardin, Z. Yuan, P. Ghamisi, and X. X. Zhu, “Dofa- clip: Multimodal vision-language foundation models for earth observation,” arXiv preprint arXiv:2503.06312, 2025. [38] K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. Khan, “Geochat: Grounded large vision-language model for remote sensing,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, p. 27 831–27 840. [39] W. Zhang, M. Cai, T. Zhang, Y. Zhuang, and X. Mao, “Earthgpt: A universal multimodal large language model for multisensor image comprehension in remote sensing domain,” IEEE Trans- actions on Geoscience and Remote Sensing, vol. 62, p. 1–20, 2024. [40] Y. Shu, B. Ren, Z. Xiong, X. X. Zhu, B. Demir, N. Sebe, and P. Rota, “Terrascope: Pixel-grounded visual reasoning for earth observation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, p. 16 712– 16 722. [41] S. Cheng, P. Chen, L. Zhu, and Y. Yan, “Beyond the geotag: assessing implicit geoprivacy risks in visual user-generated content,” International Journal of Geographical Information Science, p. 1–26, 2026. [42] P. S. Earle, D. C. Bowden, and M. Guy, “Twitter earthquake detection: earthquake monitoring in a social world,” Annals of geophysics, vol. 54, no. 6, 2011. [43] S. Zhao, Z. Xiong, J. Zhao, and X. X. Zhu, “Exebench: Benchmarking foundation models on extreme earth events,” arXiv preprint arXiv:2505.08529, 2025. [44] Y. Zhu, H. Yang, Y. Lu, and Q. Huang, “Simple, effective and general: A new backbone for cross-view image geo- localization,” arXiv preprint arXiv:2302.01572, 2023. [45] S. Zhu, M. Shah, and C. Chen, “Transgeo: Transformer is all you need for cross-view image geo-localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, p. 1162–1171. [46] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, 2022, p. 11 976–11 986. [47] M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al., “Dinov2: Learning robust visual features without super- vision,” Transactions on Machine Learning Research Journal, 2024. [48] O. Sim ́ eoni, H. V. Vo, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V. Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa et al., “Dinov3,” arXiv preprint arXiv:2508.10104, 2025. [49] R. Sinkhorn and P. Knopp, “Concerning nonnegative matrices and doubly stochastic matrices,” Pacific Journal of Mathematics, vol. 21, no. 2, p. 343–348, 1967. [50] M. Cuturi, “Sinkhorn distances: Lightspeed computation of optimal transport,” Advances in neural information processing systems, vol. 26, 2013. [51] N. Dufour, V. Kalogeiton, D. Picard, and L. Landrieu, “Around the world in 80 timesteps: A generative approach to global vi- sual geolocation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, p. 23 016– 23 026. [52] H. Li, F. Deuser, W. Yin, S. Knoblauch, W. Zhao, F. Biljecki, Y. Xue, and W. Huang, “Towards generative location awareness for disaster response: A probabilistic cross-view geolocalization approach,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 237, p. 130–145, 2026. [53] R. T. Chen and Y. Lipman, “Flow matching on general geome- tries,” in International Conference on Learning Representations, vol. 2024, 2024, p. 47 922–47 945. [54] J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou et al., “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural infor- mation processing systems, vol. 35, p. 24 824–24 837, 2022. [55] T. Doshi and Gemini Team, “Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber,” Google Blog, July 2026, accessed: 2026-08-11. [Online]. Available: https://blog. SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE18 google/innovation-and-ai/models-and-research/gemini-models/ gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/ [56] Arena, “Vision Arena: English leaderboard,” https://arena.ai/ leaderboard/vision/english, 2026, accessed: 2026-08-11. [57] X. Wu, L. Xiao, Y. Sun, J. Zhang, T. Ma, and L. He, “A survey of human-in-the-loop for machine learning,” Future Generation Computer Systems, vol. 135, p. 364–381, 2022. [58] L. Lin, D. Shi, A. Han, F. Chen, Q. Chen, J. Li, Z. Li, J. Zhang, Z. Sun, and J. Gao, “Act as human: Multimodal large language model data annotation with critical thinking,” Advances in Neural Information Processing Systems, vol. 38, p. 24 107–24 138, 2026. [59] Qwen Team, “Qwen3.6-27B,” Hugging Face Model Card, 2026, accessed: 2026-08-11. [Online]. Available: https://huggingface. co/Qwen/Qwen3.6-27B [60] OpenAI,“GPT-4osystemcard,”arXivpreprint arXiv:2410.21276, 2024. [61] J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, H. Tian, Y. Duan, W. Su, J. Shao et al., “InternVL3: Exploring advanced training and test-time recipes for open-source multimodal mod- els,” arXiv preprint arXiv:2504.10479, 2025. [62] J. Cohen, “Weighted kappa: Nominal scale agreement provi- sion for scaled disagreement or partial credit.” Psychological bulletin, vol. 70, no. 4, p. 213, 1968. [63] S. Hou, Z. Liu, L. Xie, G. Chen, H. Jiao, S. Wu, X. Guan, and H. Wu, “Postgiser: The first end-to-end fine-tuned large lan- guage model for postgis geosql query generation,” Information Processing & Management, vol. 64, no. 1, p. 105042, 2027.

Instant research discovery