Paper deep dive
Embedded Bi-Temporal Building Damage Assessment for On-Board Data Reduction
Thomas Goudemant, Benjamin Francesconi, Marjorie Bellizzi, Adrien Dorise
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Rapid assessment of building damage after natural disasters is essential to support emergency response. Earth Observation satellites can acquire relevant imagery shortly after an event, but exploitation is limited by uplink and downlink capacity and by ground-processing latency. We address this with a bi-temporal building damage assessment pipeline built on a siamese detector derived from YOLOX, designed to compress information at both ends of the ground/space link. On the ground, pre-disaster reference images are encoded into a compact latent space -- compressed by up to a factor of 64 -- and uplinked to the satellite. On board, this reference is compared with a fresh post-disaster acquisition so that the downlink carries only actionable object-level products, bounding boxes and damage classes, instead of full scenes. This cuts the data exchanged in both directions, while on xBD the strongly compressed reference still preserves most of the detection performance. Because on-board acquisitions suffer from residual pre/post co-registration errors, we introduce a latent-space shift estimation and correction module that regresses the global offset from the coarse feature level and realigns the post-disaster features before fusion. It substantially improves robustness to de-registration -- especially under large shifts, where fusion-only variants collapse -- while also raising nominal accuracy and remaining compatible with the strongest compression. We finally port the pipeline to two embedded targets, a Xilinx Versal VCK190 and an NVIDIA Jetson AGX Orin, and report hardware performance (latency, throughput, power efficiency). The core detector and its compression port cleanly to both, but the operators needed for long-range robustness survive only on the Jetson GPU, whereas the Versal DPU does not.
Tags
Links
- Source: https://arxiv.org/abs/2609.37013v1
- Canonical: https://arxiv.org/abs/2609.37013v1
Trouble viewing inline? Open PDF directly →
Full Text
40,553 characters extracted from source content.
Expand or collapse full text
Embedded Bi-Temporal Building Damage Assessment for On-Board Data Reduction ∗ Thomas Goudemant 1 , Benjamin Francesconi 1 , Marjorie Bellizzi 1 , Adrien Dorise 1,2 1 IRT Saint Exupéry, France 2 CNES, France thomas.goudemant@irt-saintexupery.com Preprint — accepted at OBPDC 2026, Barcelona, October 2026 Abstract Rapid assessment of building damage after natural disasters or conflict events is essential to support emer- gency response and prioritise field interventions. Earth Observation satellites can acquire relevant imagery shortly after an event, but exploitation is limited by up- link and downlink capacity and by ground-processing latency. We address this with a bi-temporal build- ing damage assessment pipeline built on a siamese detector derived from YOLOX and designed to com- press information at both ends of the ground/space link. On the ground, pre-disaster reference images are encoded into a compact latent space (compressed by up to a factor of 64) and uplinked to the satellite. On board, this reference is compared with a fresh post-disaster acquisition so that the downlink carries only actionable object-level products, bounding boxes and damage classes, instead of full scenes. This cuts the data exchanged in both directions and shortens the time to an actionable result, while on xBD the strongly compressed reference still preserves most of the detection performance. Because on-board acquisitions suffer from residual pre/post co-registration errors, we introduce a latent- space shift estimation and correction module that re- gresses the global offset from the coarse feature level and realigns the post-disaster features before fusion. It substantially improves robustness to de-registration, especially under large shifts where fusion-only variants collapse, while also raising nominal accuracy and re- maining compatible with the strongest compression. We finally port the pipeline to two embedded tar- gets, a Xilinx Versal VCK190 and an NVIDIA Jet- son AGX Orin, and report deployment impact and hardware performance (latency, throughput, power efficiency). We find that the core detector and its com- pression port cleanly to both, but that the operators needed for long-range robustness survive only on the Jetson GPU, which supports the full robust pipeline, whereas the Versal DPU does not. ∗ Accepted at OBPDC 2026, Barcelona, October 2026. This is the authors’ preprint version.© 2026 the authors, licensed under C BY 4.0. 1 Introduction Timely assessment of building damage after natural disasters or armed conflicts is critical to support emer- gency response, allocate limited resources and pri- oritise field interventions [12]. High-resolution (HR) Earth Observation (EO) satellites provide large-scale, repeatable coverage of impacted regions, and deep learning has significantly improved automated dam- age assessment from satellite imagery [12,21]. Most high-performing methods formulate the task as a bi- temporal change detection problem, jointly exploiting pre- and post-disaster imagery to disambiguate newly destroyed structures from pre-existing ones [18,24]. Such pipelines, however, remain largely ground-centric: they require full-image downlink and offline processing before actionable products are generated [19], which delays decision-making and consumes communication resources over areas of limited operational interest. Advances in intelligent satellite systems and em- bedded AI accelerators enable moving part of the processing closer to the sensor [23], and on-board ob- ject detection has been demonstrated for several EO tasks [15,19]. In this context, on-board AI can be used as a data-reduction and prioritisation stage: rather than transmitting complete scenes, the satellite ex- tracts and downlinks only actionable information. In our previous work [9], we introduced a ground/on- board architecture for bi-temporal building damage assessment that decouples pre- and post-event process- ing: pre-disaster scenes are encoded on the ground into compact latent representations, uplinked and stored on board, then compared with newly acquired post- disaster images to produce object-level products. That study established the benefit of siamese processing, latent-space compression and shift augmentation, and used classical cross-attention as a reference mechanism for robustness to residual co-registration errors. It did not, however, resolve two open issues: (i) robustness to large pre/post de-registration degraded well before the model’s effective receptive field was reached, and (i) the mechanisms giving robustness were not shown to be compatible with an actual embedded target. This paper extends that line of work along both axes. First, on the algorithmic side, we introduce a latent-space shift estimation and correction module 1 arXiv:2609.37013v1 [cs.CV] 29 Sep 2026 that explicitly regresses the global pre/post offset from the coarse feature level and realigns the post-disaster features before fusion. It substantially improves ro- bustness to de-registration while also raising nominal accuracy, and it remains compatible with aggressive latent compression up to a factor of 64; detailed figures are reported in Section 4. Second, on the deployment side, we port the on-board pipeline to two representa- tive embedded accelerators, a Xilinx Versal VCK190 and an NVIDIA Jetson AGX Orin, and report both the impact of deployment on algorithmic performance and the hardware performance (latency, throughput, power efficiency). We show that the core detection pipeline ports cleanly, but that the operators required for long- range robustness (softmax cross-attention and grid- sampling-based alignment) are either non-portable or non-robust on the Versal DPU, whereas the Jetson GPU supports the full robust pipeline. Contributions. (i) A latent-space shift estimation and correction module for robust bi-temporal damage assessment under strong de-registration, compatible with latent compression and designed for embedded (GPU) inference. (i) An updated robustness anal- ysis contrasting shift augmentation, cross-attention, portable (DPU-compatible) attention, and the pro- posed alignment module. (i) An embedded deploy- ment study on Versal VCK190 and Jetson AGX Orin, reporting algorithmic impact and hardware perfor- mance, and characterising which robustness mecha- nisms survive each target. 2 Related work Building damage assessment from EO imagery is com- monly cast as a bi-temporal change detection prob- lem, since post-only observation confuses damaged buildings with background structures [1,18]; fully con- volutional siamese networks compare pre-/post-event images in a shared representation space [3]. Most benchmarks pose the task as dense pixel-level segmen- tation [12], but pixel-wise labelling can yield seman- tically inconsistent predictions at the building scale when damage affects only part of a structure [24]. Object-level formulations instead take the building as the unit of analysis, localising each structure and assigning it a single damage state; this produces more coherent, compact outputs [13] that are well suited to on-board data reduction, where only actionable object-level products need be downlinked. For ro- bustness to residual co-registration, non-local mech- anisms are attractive because a feature can retrieve context from a displaced location: self-attention does this [16,20], but its spatial softmax is poorly supported by fixed-function edge accelerators. Lightweight con- volutional attention such as CBAM [22] is hardware- friendly, but its spatial component is a multiplicative gate with a fixed convolutional receptive field: it re- weights features locally and cannot fetch information from a displaced location, so it remains strictly local. Ground Segment Compressed pre-disaster latent Pre-disaster EO images Uplink Disaster Ground Encoder (encode + compress) On-board Segment (satellite) Building BBoxes On-board Detection & Coregistration Module Damage Level Cropped images Downlink Post-disaster de-registrated images Low-rate Products Generation x64 compression rate (57 kB vs 3.15 MB image) Figure 1: System concept. The ground segment en- codes pre-disaster references into a compressed latent (×64) uplinked to the satellite; the on-board segment detects and co-registers against the fresh post-disaster image and downlinks only low-rate products (bound- ing boxes, damage level, optional crops over damaged regions) instead of full scenes. On the deployment side, embedded AI in orbit has been demonstrated for object and change detection on spaceborne hardware [10,11,19,23], and system-level demonstrators have explored reactive ground–space architectures that shorten the decision–action loop [6]; yet these efforts rarely combine bi-temporal reasoning, object-level modelling and the constraints of embed- ded deployment [4]. We refer the reader to [9] for the full survey and adopt its object-level, single-stage detection formulation; this paper focuses on the two new contributions: an explicit latent-space alignment module and an embedded portability study. 3 Method 3.1 Ground/on-board system concept We consider an operational EO scenario in which rapid building damage assessment is required to support post-disaster response. Following a disaster alert or the identification of an area of interest, the ground segment selects the relevant pre-disaster reference im- age(s), encodes them into compact multi-scale latent representations, and uplinks only the latents over the targeted area. The on-board segment processes the newly acquired post-disaster image and compares it with the stored pre-disaster latents to produce com- pact object-level products: building bounding boxes and damage classes, with the option to selectively downlink image crops over damaged regions instead of full scenes. This separation (Fig. 1) reduces the uplink volume (compact latents instead of full-resolution pre-disaster images) and the downlink volume (damage products and selected crops instead of full post-disaster scenes). It also introduces the constraints that drive our de- sign: uplink bandwidth and on-board memory forbid transmitting complete pre-disaster scenes, and lim- ited on-board geolocation accuracy induces residual pre/post misregistration. The detailed network is shown in Fig. 2. 2 3.2 Base detection architecture The base detector and the two baseline robustness mechanisms below (shift augmentation and attention) are those of [9], where they are detailed; we only sum- marise them here. The backbone is YOLOX-S [8], whose CSPDarknet produces three feature maps (D3, D4, D5) at strides 8, 16 and 32. These levels differ in nature: D3, at stride 8 and the finest spatial resolution, stays close to the input image and retains fine geomet- ric detail, which makes it the level on which spatial registration can still be performed; D5, at stride 32, is spatially coarse but semantically rich, encoding object- level representations. YOLOX’s Path Aggregation FPN (PAFPN) [14] then fuses these maps with both a top-down and a bottom-up path, propagating D5’s semantic context into the finer, better-localised D3/D4 levels and, in turn, their spatial detail back up. We start from a classic YOLOX and simply feed it the pre- and post-disaster images stacked along the channel axis (a 6-channel input); we refer to this base- line as early fusion, since the two images are mixed at the very input. We then evolve it into a siamese architecture: a single shared encoder processes the pre- and the post-disaster image independently, and the two feature sets are concatenated, together with their element-wise difference, before the FPN. This de- composition lets the network reason separately about each image (e.g. distinguishing building edges from ground) while learning a common representation of both, which improves detection over early fusion. Cru- cially, sharing one encoder maps directly onto the ground/on-board split: the same encoder is used on the ground to encode the pre-disaster reference and on board to encode the newly acquired post-disaster image, so pre-disaster features are computed on the ground and reused on board. Compressed pre-disaster latents. The raw multi- scale pre-disaster features exceed the input image size (≈3.7 MB vs. 3.15 MB at 1024 2 ), so a lightweight 1×1 convolution reduces the channel count by a factorr before uplink, with a symmetric projection restoring dimensionality on board. Tab. 1 reports the resulting int8 footprint:r=64 shrinks the latent to 57 kB, two orders of magnitude below the image, whiler=128 leaves a single latent channel at the finest level (D3), which we show is insufficient for the alignment module (Sec. 4.3). 3.3Robustness to pre/post de- registration Residual spatial misregistration between the pre- and post-disaster images is unavoidable on board, due to acquisition-geometry differences and limited on- board registration. Under such shifts, strictly local feature comparison fails to retrieve the relevant pre- disaster information at the correct location, which mostly degrades damage classification (localisation is more stable, as boxes are predicted in the pre-disaster Table 1: Size of the pre-disaster latent representation under channel compression, for a 1024×1024×3 input, reported as an equivalent int8 footprint. Spatial resolutions: 128 2 (D3), 64 2 (D4), 32 2 (D5). ConfigurationD3 D4 D5 Size Input image (1024 2 ×3) – 3.15 MB Latent (no comp.)128 256 512 3.67 MB Latent (comp. r=8)16 32 64 0.46 MB Latent (comp. r=64)24857 kB Latent (comp. r=128)12429 kB reference frame) [9]. We consider three families of mechanisms, of increasing effectiveness. (a) Shift augmentation and (b) attention. Shift augmentation applies, during training, a random dis- placement (offsets in [0,150] px) to the post-disaster image only, with ground truth kept in the pre-disaster frame; it is used in all variants unless stated. Cross- attention lets each pre-disaster location attend over the whole post-disaster map and retrieve possibly shifted features, using the query–key–value formulation of [20] on the coarse levels (D4–D5); itsHW × HWsoft- max, however, is not supported by the Versal DPU (Sec. 3.4). We therefore additionally consider portable attention gates built only from DPU-compatible opera- tors, the most representative being a spatial-difference gate adapted from the spatial branch of CBAM [22]: a depthwise 7×7 convolution over the signed pre−post difference, followed by a pointwise convolution and a sigmoid gate. Being purely convolutional, it has a strictly local receptive field. (c) Latent-space shift estimation and correc- tion (proposed). We estimate the global pre/post offset explicitly and use it to realign the post-disaster features before fusion (Fig. 2). An offset head pre- dicts one global displacement (δ x ,δ y ) per image by cross-correlating the pre and post features. It works on D3, the finest and least-compressed level, whose geometric detail is best for matching; a stride-4 convo- lution first subsamples it to 32×32 to keep the search cheap. On a grid of (2R+1) 2 candidate offsets, we cosine-correlate the pre and shifted post embeddings, and a soft-argmax over the grid gives a sub-pixel off- set. We useR=13, covering shifts of up to 416 px. A coarse grid is sufficient in practice: the correction only needs to bring the residual shift back within the perceptive field, where detection stays robust (≈20 px, Fig. 3), and the soft-argmax already interpolates be- low one cell. The offset then shifts the post-disaster D3/D4/D5 features toward the pre features by bilinear resampling, each level scaled by its stride. The whole module is differentiable, so it is trained end-to-end by the detection loss plus an auxiliary loss that regresses the known training shift. The module stays exportable on embedded accelera- tors: the offset search is a single vectorised correlation 3 and the resampling reduces to one grid-sampling oper- ation. Grid-sampling is a native TensorRT operator, so the alignment is portable on the Jetson GPU but not on the Versal DPU (Sec. 3.4), which it therefore targets. 3.4 Embedded targets and portability We deploy the on-board branch on two representative accelerators. The Xilinx Versal VCK190, whose adaptive-compute architecture [7] is a candidate for next-generation edge computing in space [17], car- ries aDPUCVDX8G_ISA3_C32B6DPU (batch 6) pro- grammed through Vitis AI with int8 quantisation and a fixed-function operator set. The NVIDIA Jetson AGX Orin [2] is a general-purpose edge GPU pro- grammed through TensorRT, capped here to a 30 W power mode. The DPU imposes a hard portability constraint: the on-board model must compile to a single DPU subgraph, otherwise the runtime aborts. The siamese detector and its latent-compressed variants meet this constraint and port cleanly, as does the portable at- tention gate (the spatial-difference gate). Classical cross-attention and the alignment module, by contrast, are not portable to the DPU. In short, the DPU sup- ports the core bi-temporal detector, compression and the portable attention gate, but not the operators pro- viding long-range robustness through cross-attention or explicit alignment, which the Jetson GPU does support. 4 Experiments and results 4.1 Dataset and protocol We conduct all experiments on the xBD dataset [12] (xView2 Challenge), built from very-high-resolution optical imagery of the Maxar/DigitalGlobe Open Data program. The images are orthorectified RGB prod- ucts, provided as coarsely registered pre-/post-event pairs of size 1024×1024 at a ground sampling distance below 0.8 m/px. Buildings are annotated with four damage levels (no damage, minor, major, destroyed). We reformulate the segmentation annotations as ob- ject detection by converting polygons to axis-aligned boxes, and use the standard split (excluding Tier 3): 4,665 pairs, 60/20/20% train/val/test. All variants use a YOLOX-S backbone and identical optimisation settings. Detection is evaluated with precision, recall, F1 and mAP@0.5. Robustness is measured by apply- ing synthetic shifts at test time to the post-disaster image only, with ground truth kept in the pre-disaster frame and all metrics computed on the common spatial support induced by the maximum displacement, so that every shift is scored on an identical image and annotation subset. 4.2 Nominal detection performance Tab. 2 reports performance without test-time shift. Siamese processing improves over early fusion, con- firming the benefit of separating pre- and post-disaster feature extraction before comparison. Cross-attention and the portable spatial-difference gate leave nominal accuracy essentially unchanged. The proposed shift estimation and correction module gives the best nomi- nal mAP@0.5 (57.8%), and combining it with latent compressionr=64 costs only≈1 p of mAP while shrinking the uplinked latent by two orders of mag- nitude (Tab. 1). This confirms that the pre-disaster information required for the task can be represented with a much smaller volume than the original image. Table 2: Nominal detection performance on xBD (no test-time shift), %. All variants use shift augmentation unless stated. ConfigurationP R F1 mAP@0.5 Early Fusion (baseline)58.3 55.0 56.355.1 Siamese (no shift-aug)59.0 58.0 58.256.4 Siamese58.5 57.4 57.856.6 Siamese + Cross-Attention61.1 56.0 58.456.5 Siamese + Portable Attn (Sp.-Diff) 58.9 57.3 57.956.7 Siamese + Shift Est. & Corr.61.5 57.3 59.257.8 + Latent Compression ×6461.3 55.8 58.356.7 4.3 Robustness to de-registration Fig. 3 reports mAP@0.5 as a function of the ap- plied post-disaster shift. Without shift augmentation (the early-fusion and siamese-no-aug baselines) perfor- mance collapses within a few tens of pixels, i.e. typical building extents. Shift augmentation alone slows the degradation but its mAP@0.5 still falls to 26.6% at a 400-pixel shift (from 56.6% nominal). Cross-attention adds genuine long-range robustness (43.0% at 400 px) but does not port to the DPU (it runs on the Jetson GPU only), while the portable spatial-difference gate behaves like the plain siamese model (29.2%): its local receptive field cannot retrieve distant features, so it does not deliver the robustness of true attention. This is a negative result we make explicit, since it is exactly this gap that motivates an explicit alignment. The proposed shift estimation and correction module is the clear best variant: it keeps mAP essentially flat across the whole shift range, holding 49.4% at 400 px (from 57.8% nominal), against 26.6% for shift augmen- tation and 43.0% for cross-attention, while also improv- ing nominal accuracy. Notably, mAP stays flat well beyond the 150 px training range of the shift augmenta- tion (vertical line in Fig. 3): the model extrapolates to shifts larger than any seen during training, because the offset head estimates the displacement geometrically rather than memorising trained offsets. Adding latent compression×64 preserves this behaviour almost ex- actly (47.0% at 400 px), so aggressive data reduction and strong de-registration tolerance are compatible. 4 Pre-disaster image (RGB) Onboard YOLOX-S siamese model + Latent Offset Alignment D4 post + aligned D5 post + aligned Latent Of fset Alignenment D3 post + aligned D3 post D4 post D5 post Feat Decompression CSPDarknet Backbone D4 pre D5 pre D3 pre D5c D3c D4c FPN Neck + Detection Head Offset Estimation Pre/Post Offset Post-disaster image (RGB) D3 fusion D4 fusion D5 fusion CSPDarknet Backbone Ground Encoder Feat Compression D4 pre D3 pre D5 pre Outputs BBoxes Damage Class D3 pre (recon.) stride 8, 128² learned stride-4 conv decimate128² → 32² cos. cross-correlation ±13 px in x, y: 27²=729 candidate shifts (±416 px) soft- argmax global shift (δx, δy) Detailed Offset Estimation D3 post stride 8, 128² Figure 2: Proposed ground/on-board architecture (details in Sec. 3.3). Compressed pre-disaster latents are uplinked and expanded on board; the offset-estimation module regresses a global shift (δ x ,δ y ) from the D3 latent, which the latent offset alignment applies to realign the post D3/D4/D5 features before multi-scale fusion. 15102050100200400 Post-disaster image shift (pixels) 20 30 40 50 60 mAP (%) training limit (150px) Robustness to Misregistration mAP@0.5 vs. Shift Early Fusion Siamese (no shift-aug) Siamese Siamese + Cross-Attn Siamese + Portable Attn Siamese + Latent Offset Align Siamese + Latent Offset Align + Compression ×64 Siamese + Latent Offset Align + Compression ×128 Figure 3: Robustness to pre/post misregistration on xBD: mAP@0.5 vs. applied post-disaster shift (log-x), averaged over shift directions. The vertical line marks the 150 px training range of the shift augmentation; the proposed module (green) stays flat beyond it, whereas shift-augmentation and attention baselines degrade. Offset estimation quality and ther=128 limit. Fig. 4 reports the module’s offset estimation error. It is low and grows smoothly within the useful range, and compression×64 tracks the uncompressed esti- mator almost exactly, confirming that the offset can be recovered directly from the compact latent. Push- ing to×128, however, leaves a single latent channel at D3 (Tab. 1): the correlation signal becomes too weak and the estimation error roughly doubles. The alignment can then no longer lock onto the shift, and this propagates to the end task: the post features are mis-warped, so damage classification degrades and the overall detection mAP@0.5 drops well below the×64 configuration under shift, undoing the robustness gain. r=128 removes the information the alignment module needs, sor=64 is the aggressive-compression limit at which the robust pipeline still works. 050100150200250300350400 Applied shift (pixels) 0 50 100 150 200 250 300 Offset estimation error (px, mean ± std) training limit (150px) Shift Estimation Error CoarseD3 (sans compression) CoarseD3 + Comp64 CoarseD3 + Comp128 Figure 4: Offset estimation error (mean±std) vs. applied shift. Compression×64 matches the uncom- pressed estimator;×128 (a single D3 latent channel) roughly doubles the error and breaks alignment. The rise beyond∼150 px reflects extrapolation outside the training range. 4.4Embedded deployment: algorith- mic impact Tab. 3 reports the effect of porting on mAP@0.5. Quan- tisation is near-neutral for all compressed/portable models on both targets: on the Jetson, the plain siamese detector loses only 0.31 p in int8; on the Ver- sal, the×128-compressed detector and the portable spatial-difference proxy stay within±1.3 p of their float32 references. The exception is the full alignment model (with compression×64) on the Jetson, which drops 3.24 p already in FP16. We attribute this to a porting bug in the alignment path, not to numerical precision. Two arguments sup- port this. First, FP16 is otherwise near-lossless on this target (all other models lose<0.4 p), so a 3 p drop appearing in FP16, before any int8 rounding, cannot be a quantisation effect. Second, the drop is localised to the one component that distinguishes this model: 5 the grid-sampling warp of the alignment module. The operator does export and run (the engine builds and produces detections), but its numerical output diverges from the reference PyTorch model. The likely cause is a mismatch in the resampling convention between the two implementations (corner alignment, the normali- sation of sampling coordinates, or the border-padding mode), any of which shifts the warped features by a fraction of a cell and thus corrupts the pre/post comparison that damage classification relies on. This is an implementation-level porting issue, still open at the time of writing, and must be resolved before this model is trusted in production on the Jetson; it does not affect the other reported models, none of which use grid sampling. Table 3: Impact of quantisation/deployment on port- ing, per model (mAP@0.5, %). Rows compare each model’s float32 reference against its on-target quan- tised version, not models across platforms. PlatformModelfloat32 Quant.∆ Jetson (TensorRT) Siamese57.90 57.59 (int8) −0.31 Jetson (TensorRT) S.E.C. + C×6457.63 54.39 (FP16) −3.24 ∗ VCK190 (Vitis AI) Siamese + C×128 52.66 53.98 (int8) +1.32 VCK190 (Vitis AI) Portable Sp.-Diff 58.85 58.03 (int8) −0.81 ∗ open porting issue on the grid-sampling alignment path, not a quantisation effect. 4.5Embedded deployment: hardware performance We measure end-to-end latency (broken down into preprocess / inference / postprocess), throughput and power on both targets. On the Jetson we sweep five axes: model, precision (FP16/int8), NMS mode, batch size and image size (Fig. 5); NMS runs either as a CPU post-processing step (external) or fused into the engine as a GPU TensorRT plugin (integrated, faster); on the VCK190 we can only sweep model and image size (Fig. 6): the other axes are fixed in the DPU image (changing the batch size requires re- building it, too costly to sweep), and NMS runs on the CPU, unsupported by the DPU. Unless stated, the reference point is 1024×1024, integrated NMS, Jetson at batch 1/FP16/30 W mode and VCK190 at batch 6/int8. Power is read from each board’s on- board monitors: the Jetson’s three GPU/CPU/SYS rails (INA3221) and the VCK190’s 17 INA226 rails. Throughput is reported in Mpixels/s, which is compa- rable across image sizes and batch settings. Per-model behaviour. On the Jetson, the plain siamese detector runs at 25.6 ms / 41.0 Mpx/s / 288 GFLOP/s/W; adding cross-attention is almost free (27.8 ms), whereas the alignment module is≈40% slower (38.5–38.8 ms) because of the multi-level grid- sampling warp, not because of attention. Latent compression×64 brings no latency gain over the un- compressed alignment model (38.8 vs. 38.5 ms): it shrinks the tensor exchanged between ground and edge, which matters for the uplink, but does not change the on-board compute that dominates latency. On the VCK190, compression×128 is instead the fastest model (67.9 ms / 92.7 Mpx/s), 8% faster than the plain detector, because DPU compute is bound by the feature volume and compression lightens it directly; the portable spatial-difference proxy is the slowest (81.9 ms), the cost of its extra spatial gate. Precision, NMS and image/batch sweeps. The Jetson sweep (Fig. 5) isolates several system knobs. Precision: switching FP16 to int8 cuts latency by≈13% and lifts efficiency markedly (up to 465 GFLOP/s/W on the plain detector) at near- constant power. NMS mode: integrated NMS is con- sistently 3–5 ms cheaper in postprocess than external NMS, since it avoids the GPU-to-CPU round-trip; we thus use it for all hardware measurements, external NMS being needed only for uncapped-detection mAP evaluation. Batch size: throughput rises from batch 1 to≈3 (+13%) and then plateaus, as the GPU already saturates at 1024 2 ; power stays nearly flat. Image size: throughput in Mpixels/s grows with resolution (9.4→32.4 Mpx/s from 256 to 2048) as the fixed per- pixel cost is better amortised, while frame rate falls and power rises with load. On the VCK190, 512 px is the most efficient operating point (3.6 vs. 2.6 Mpx/s/W at 1024/2048). Cross-platform summary. Tab. 4 collects the best portable model per target plus the common noatt reference. On the plain detector (the only fair cross- platform comparison), the VCK190 reaches≈2×the throughput of the Jetson (85 vs. 41 Mpx/s) but draws ≈2.4×the power (29.8 vs. 12.5 W), so energy effi- ciency favours the Jetson (288 vs. 172 GFLOP/s/W), consistent with a general-purpose GPU tuned for FP16/int8 versus an int8-only DPU on a develop- ment board. Each platform’s best model behaves differently. On the VCK190 compression is essentially free: the compressed best model matches the plain detector in efficiency (171 vs. 172 GFLOP/s/W) and is even marginally faster. On the Jetson, enabling the full robust pipeline (S.E.C. alignment) is what costs: efficiency drops from 288 to 181 GFLOP/s/W and latency rises 25.6→38.8 ms (≈40%), bringing the Jetson down to roughly the Versal’s efficiency, in ex- change for the long-range robustness the DPU cannot provide. Table 4: Hardware performance at 1024 2 . Only noatt (Siamese) is compared across platforms; per-platform best models use different compression and are not cross- compared. Jetson: FP16, batch 1, 30 W; VCK190: int8, batch 6. Platform ModelGFLOPs Lat. Thr. Power Eff. (/img) (ms) (Mpx/s) (W) (GF/s/W) Jetson Orin Siamese69.825.6 41.012.5288 VCK190Siamese69.874.1 85.029.8172 Jetson Orin S.E.C. + C×64 (best)70.438.8 27.112.0181 VCK190Siamese + C×128 (best) 69.967.9 92.732.3171 6 Figure 5: Jetson AGX Orin sweep. Top: latency (stacked bars: preprocess / inference / postprocess, left axis) and throughput (line, right axis). Bottom: power (mean bar with a min–max whisker, left axis) and energy efficiency (line, right axis). Five axes: model, precision, NMS mode, batch size, image size. Figure 6: VCK190 sweep (int8, batch 6), two axes (model, image size). From left to right: latency/throughput by model, latency/throughput by image size, power/efficiency by model, power/efficiency by image size. The DPU batch size is fixed and the NMS/precision knobs of the Jetson path do not apply. 5 Discussion Data-reduction use case. The pipeline reduces data on both links. For a representative 100 km 2 urban disaster area (≈Paris proper) at 0.8 m/px, uplinking full-resolution pre-disaster references would require ≈450 MB, whereas×64-compressed latents bring the uplink below 10 MB (two orders of magnitude less). On the downlink, on-board inference transmits compact detection products (boxes, damage levels, optional crops over damaged regions) instead of full scenes. At the measured throughput (∼27–85 Mpx/s), inference over such an area takes a few seconds: compute is not the bottleneck, the downlink is, which is exactly what compressing the products on board reduces. Robustness vs. portability. Our results clearly separate the two targets. The alignment module is the best answer to de-registration (flat mAP across the operational shift range and the best nominal ac- curacy) but relies on grid sampling, a GPU operator absent from the Versal DPU. Classical cross-attention is next best on robustness yet equally non-portable (spatial softmax). The DPU-compatible alternatives (a portable spatial-difference gate, and compression) keep the model deployable but do not recover long- range robustness (the gate is only as robust as the plain siamese model). Thus, on a fixed-function DPU, on-board correction of large de-registration remains open, and alternatives, such as TriCCOT [5], are be- ing explored. In contrast, an embedded GPU (Jetson) supports the full robust pipeline today. Limitations. This work is a step toward raw on- board data, not a validation on it. We operate on L1C- like inputs (ortho-rectified, calibrated, only coarsely co-registered RGB pairs, as in xBD [12]) and move closer to raw acquisitions by removing the need for ac- curate co-registration: the network tolerates a residual pre/post pixel shift rather than assuming aligned pairs. Part of the real variability is already exercised by xBD and thus implicitly evaluated: off-nadir differences sur- face as a fixed per-image pixel drift [12], exactly the global shift our alignment targets, and illumination 7 and seasonal changes span nineteen events over several years, though a fine-grained evaluation of each factor is left for future work. Other gaps remain: on-board sensor calibration, scale and rotation differences (only translation is handled today), cross-sensor pairs, and validation on true raw acquisitions. 6 Conclusion We presented an on-board data-reduction architecture for bi-temporal building damage assessment that com- presses pre-disaster information on the ground and extracts compact damage products on board. Building on our previous work [9], it addresses three challenges of this distributed ground/on-board use case. First, it cuts the uplink data volume by about two orders of magnitude (up to×64) with no notable loss in de- tection, making the pre-disaster reference practical to transmit. Second, it moves closer to raw on-board inputs: a latent-space shift estimation and correc- tion module regresses the pre/post offset and realigns the post-disaster features before fusion, removing the need for accurate co-registration and outperforming shift augmentation and cross-attention under strong de-registration, with robustness to fully raw acquisi- tions remaining future work. Third, it is deployed on embedded hardware (to our knowledge, the first on- board, dual ground/on-board detector of its kind) on a Versal VCK190 and a Jetson AGX Orin. The core detector and its compression port cleanly, while the op- erators giving long-range robustness (spatial-softmax attention, grid-sampling alignment) are non-portable on the fixed-function DPU, whereas the Jetson GPU runs the full pipeline. Several directions follow: an asymmetric, larger ground-side encoder (it runs on the ground) could yield richer references at no on-board cost; multi-temporal series and multimodal inputs (e.g. SAR with optical) would add robustness to illumina- tion and weather. References [1] G. Andresini, A. Appice, D. Ienco, and D. Malerba. Seneca: Change detection in optical imagery using siamese net- works with active-transfer learning. Expert Systems with Applications, 214:119123, 2023. [2]M. Barnell, C. Raymond, M. Wilson, D. Isereau, and E. Cote. Ultra low-power deep learning applications at the edge with Jetson Orin AGX hardware. In IEEE High Performance Extreme Computing Conf. (HPEC), pages 1–4, 2022. [3]R. Caye Daudt, B. Le Saux, and A. Boulch. Fully convo- lutional siamese networks for change detection. In 2018 25th IEEE International Conference on Image Processing (ICIP), pages 4063–4067, 2018. [4]A. Dorise, M. Bellizzi, A. Girard, B. Francesconi, and S. May. Explaining raw data complexity to improve satellite on-board processing. In 2025 European Data Handling & Data Processing Conference (EDHPC), pages 1–8, Oct 2025. [5]A. Dorise, M. Bellizzi, J. Cohen, and S. May. Triccot: Tri-part convolutional conformal transformer for onboard space object detection, 2026. URLhttps://arxiv.org/ abs/2609.08659. [6] B. Francesconi, L. Palluel, T. Goudemant, H. Meleiro, B. Marchand, O. Thiery, and F. Creil. From event to action: A reactive loop demonstrator for earth observation based on modular ai-driven components. In Actes de la 7ème Conference on Artificial Intelligence for Defense (CAID 2025), 2025. [7]B. Gaide, D. Gaitonde, C. Ravishankar, and T. Bauer. Xil- inx adaptive compute acceleration platform: Versal archi- tecture. In ACM/SIGDA Int. Symp. Field-Programmable Gate Arrays (FPGA), pages 84–93, 2019. [8] Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. [9]T. Goudemant and B. Francesconi. Optimizing latent repre- sentations for robust building damage assessment onboard earth observation satellites. In Proc. IEEE/CVF CVPR Workshops, pages 10232–10240, 2026. [10]T. Goudemant, B. Francesconi, H. Farhat, L. Daniel, O. Thiery, E. Kervennic, A. Girard, and S. Mzoughi. Dé- tection de navires embarquable à bord de satellites. In Actes de la 4ème Conference on Artificial Intelligence for Defense (CAID 2022), 2022. [11]T. Goudemant, B. Francesconi, M. Aubrun, E. Kervennic, I. Grenet, Y. Bobichon, and M. Bellizzi. Onboard anomaly detection for marine environmental protection. IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., 17:7918–7931, 2024. [12] R. Gupta, R. Hosfelt, S. Sajeev, N. Patel, B. Goodman, J. Doshi, E. Heim, H. Choset, and M. Gaston. xbd: A dataset for assessing building damage from satellite imagery, 2019. [13]H. Jiang, M. Peng, Y. Zhong, H. Xie, Z. Hao, J. Lin, X. Ma, and X. Hu. A survey on deep learning-based change detec- tion from high-resolution remote sensing images. Remote Sensing, 14(7), 2022. [14]T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie. Feature pyramid networks for object detection. In IEEE Conf. Computer Vision and Pattern Recognition (CVPR), pages 2117–2125, 2017. [15]N. Melega, B. A. Carnicero, N. Longepe, A. Paskeviciute, V. Marchese, O. Aragon, I. Babkina, A. Marin, J. Nalepa, L. Buckley, et al. Implementation of theφsat-2 on board image processing chain. In Sensors, systems, and next- generation satellites XXVII, volume 12729, pages 264–276, 2023. [16]A. Mohammadian and F. Ghaderi. Siamixformer: a fully- transformer siamese network with temporal fusion for accu- rate building detection and change detection in bi-temporal remote sensing images. International Journal of Remote Sensing, 44(12):3660–3678, June 2023. [17]N. Perryman, C. Wilson, and A. George. Evaluation of Xilinx Versal architecture for next-gen edge computing in space. In IEEE Aerospace Conference, pages 1–11, 2023. [18]J. Shao, L. Tang, M. Liu, G. Shao, L. Sun, and Q. Qiu. Bdd- net: A general protocol for mapping buildings damaged by a wide range of disasters based on satellite imagery. Remote Sensing, 12(10), 2020. [19]Y. Shen, D. Liu, J. Chen, Z. Wang, Z. Wang, and Q. Zhang. On-board multi-class geospatial object detection based on convolutional neural network for high resolution remote sensing images. Remote Sensing, 15(16), 2023. [20]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. CoRR, abs/1706.03762, 2017. [21] L. Wang, J. Wu, Y. Yang, R. Tang, and R. Ya. Deep learn- ing models for hazard-damaged building detection using remote sensing datasets: A comprehensive review. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:15301–15318, 2024. [22]S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon. CBAM: Convolutional block attention module. In European Conf. Computer Vision (ECCV), pages 3–19, 2018. [23]B. Zhang, Y. Wu, B. Zhao, J. Chanussot, D. Hong, J. Yao, and L. Gao. Progress and challenges in intelligent remote sensing satellite systems. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15: 1814–1822, 2022. [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, 265:112636, 2021. 8