Paper deep dive
TRNet: Topography-Guided Frequency Rectification and Structure-Aware Decoding for Multimodal Paddy Rice Segmentation
Kaiwen Xiao, Chunlong Fu, Liping Zheng, Yanfeng Su
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/9/2026, 3:03:13 AM
Summary
The paper introduces TRNet, a multimodal segmentation network for mapping paddy rice in mountainous regions using very-high-resolution RGB imagery and coarse topographic data (DEM and slope). TRNet employs a Topographic Energy-Spectral Rectification (TESR) module to modulate visual features based on terrain conditions, suppressing steep-slope clutter and enhancing low-slope rice cues. It also uses a Topography-guided Paddy Structure Decoder (TPSD) to refine boundaries and interiors using coarse topography as context. Experiments on Area A and Area B in Hongya County, China, demonstrate that TRNet outperforms Dual-Encoder U-Net, achieving higher IoU scores and reducing false positives on steep terrain.
Entities (9)
Relation Signals (9)
TRNet → usesmodule → Topographic Energy-Spectral Rectification
confidence 95% · Its two components are Topographic Energy-Spectral Rectification (TESR) and the Topography-guided Paddy Structure Decoder (TPSD).
TRNet → usesmodule → Topography-guided Paddy Structure Decoder
confidence 95% · Its two components are Topographic Energy-Spectral Rectification (TESR) and the Topography-guided Paddy Structure Decoder (TPSD).
TRNet → evaluatedon → Area B
confidence 90% · Experiments used ... held-out Area B
TRNet → evaluatedon → Area A
confidence 90% · Experiments used an Area A internal test set
TRNet → outperforms → Dual-Encoder U-Net
confidence 90% · TRNet achieved rice intersection-over-union (IoU) values ... exceeding the original Dual-Encoder U-Net
TRNet → processesinput → TanDEM-X
confidence 90% · We present TRNet for ... a 5-m TanDEM-X digital elevation model (DEM)
TRNet → processesinput → GaoJing-1
confidence 90% · We present TRNet for 0.5-m GaoJing-1 red--green--blue (RGB) imagery
Topographic Energy-Spectral Rectification → suppresses → steep-slope clutter
confidence 85% · suppress steep-slope clutter and conditionally enhance compatible low-slope rice cues
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Mapping paddy rice from very-high-resolution imagery in mountainous and hilly regions is difficult because terrain alters optical appearance and increases confusion with visually similar vegetation. We present TRNet for 0.5-m GaoJing-1 red--green--blue (RGB) imagery, a 5-m TanDEM-X digital elevation model (DEM), and derived slope. Separate visual and terrain encoders preserve modality-specific features. At an early encoder stage, Topographic Energy-Spectral Rectification applies terrain-conditioned low-frequency modulation and asymmetric high-frequency regulation to suppress steep-slope clutter and conditionally enhance compatible low-slope rice cues. The Topography-guided Paddy Structure Decoder combines semantic, rice--background boundary, and interior cues, using coarse terrain as context. Experiments used an Area A internal test set and held-out Area B, which had steeper terrain and lower rice prevalence. TRNet achieved rice intersection-over-union (IoU) values of 85.10\% and 80.68\%, exceeding the original Dual-Encoder U-Net by 9.15 and 18.83 percentage points, respectively. Ablation and slope-stratified results linked these gains to frequency rectification, structure learning, and fewer steep-terrain false positives. The results support coarse topography as a contextual prior for very-high-resolution paddy rice mapping.
Tags
Links
- Source: https://arxiv.org/abs/2608.04154v1
- Canonical: https://arxiv.org/abs/2608.04154v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
73,841 characters extracted from source content.
Expand or collapse full text
TRNet: Topography-Guided Frequency Rectification and Structure-Aware Decoding for Multimodal Paddy Rice Segmentation Kaiwen Xiao, Chunlong Fu, Liping Zheng, and Yanfeng Su Kaiwen Xiao, Chunlong Fu, Liping Zheng, and Yanfeng Su are with the School of Computer Science, Sichuan University Jinjiang College, Meishan 620860, China. Corresponding author: Kaiwen Xiao. Abstract Mapping paddy rice from very-high-resolution imagery in mountainous and hilly regions is difficult because terrain alters optical appearance and increases confusion with visually similar vegetation. We present TRNet for 0.5-m GaoJing-1 red–green–blue (RGB) imagery, a 5-m TanDEM-X digital elevation model (DEM), and derived slope. Separate visual and terrain encoders preserve modality-specific features. At an early encoder stage, Topographic Energy-Spectral Rectification applies terrain-conditioned low-frequency modulation and asymmetric high-frequency regulation to suppress steep-slope clutter and conditionally enhance compatible low-slope rice cues. The Topography-guided Paddy Structure Decoder combines semantic, rice–background boundary, and interior cues, using coarse terrain as context. Experiments used an Area A internal test set and held-out Area B, which had steeper terrain and lower rice prevalence. TRNet achieved rice intersection-over-union (IoU) values of 85.10% and 80.68%, exceeding the original Dual-Encoder U-Net by 9.15 and 18.83 percentage points, respectively. Ablation and slope-stratified results linked these gains to frequency rectification, structure learning, and fewer steep-terrain false positives. The results support coarse topography as a contextual prior for very-high-resolution paddy rice mapping. I Introduction Accurate paddy rice maps support agricultural management, irrigation planning, crop statistics, food-security assessment, and regional monitoring [38, 42, 36, 34]. Remote sensing substantially advances paddy rice mapping by enabling efficient and increasingly accurate observation over large areas [9, 8, 20, 21]. With the continued development of remote sensing technology, very-high-resolution (VHR) imagery further extends rice mapping toward parcel-level analysis by resolving field margins, parcel geometry, and within-field texture that are obscured in coarser regional-scale observations [41]. The resulting fine-grained spatial information can be effectively learned by encoder–decoder segmentation networks [24, 5]. However, mapping paddy rice from VHR imagery remains challenging in mountainous and hilly regions, where small and fragmented rice fields are embedded in complex terrain and heterogeneous land-cover backgrounds [31]. One difficulty arises from topographic variation, which alters optical illumination and land-cover appearance, while mountain shadows introduce strong radiometric discontinuities [3, 26]. Under VHR observation, these effects appear as salient local contrasts and textures that can obscure or imitate vegetation and field structure [4]. Segmentation models may consequently respond strongly to steep non-rice terrain, making terrain-induced visual detail difficult to distinguish from genuine rice structure [43]. A separate difficulty concerns fragmented terraces and irregular field margins, which divide rice cover into small regions whose complex contours and narrow interiors are difficult to recover [15, 30]. Although multi-scale networks and edge-aware extraction improve cropland and field-boundary mapping, semantic predictions can still contain incomplete boundaries and inconsistent region interiors when these structures are not explicitly supervised [28]. To mitigate the confusion between terrain-induced visual detail and genuine rice structure, relevant approaches can be broadly grouped into two categories: multisource fusion and frequency-domain modeling. Multisource fusion introduces complementary observations and controls how they interact with optical features. The Gather-to-Guide Network aggregates multimodal features to guide RGB representations [44], whereas CMX rectifies modality-specific features before Transformer-based cross-modal fusion [40]. For optical–topographic data, DMTFNet separately encodes Sentinel-2 and terrain features and combines them through multiscale Transformer decoding and adaptive fusion [7]; FCA-DeepLab instead integrates optical and DEM branches using spatial–channel attention [27]. These methods preserve complementary modality information and regulate cross-modal interaction. However, when DEM and slope products are substantially coarser than VHR imagery, fusion cannot recover field-scale spatial detail absent from the topographic sources, leaving unresolved how coarse terrain cues can distinguish terrain-induced visual responses without being treated as fine spatial evidence. Frequency-domain modeling addresses spatial-scale variation within feature representations. SFFNet combines wavelet-derived frequency features with spatial features [37], while hierarchical wavelet feature enhancement decomposes features at multiple scales, strengthens high-frequency subbands, and feeds them back to the corresponding network stages [12]. Dual-domain processing further aligns wavelet features with spatial constraints [33], and learned frequency-domain fusion extends such interaction to heterogeneous remote-sensing modalities [1]. These methods make scale-specific feature regulation explicit. Nevertheless, frequency decomposition or high-frequency enhancement alone cannot determine whether fine detail represents genuine field structure or terrain-induced texture, leaving the frequency components without an explicit topographic criterion for response regulation. The second challenge, incomplete boundaries and inconsistent interiors in fragmented rice regions, has mainly been approached through explicit spatial structure supervision, which imposes geometric constraints beyond region-overlap objectives. E2EVAP combines semantic–contour interaction with a topological loss to vectorize smallholder parcel boundaries [22]. BsiNet jointly predicts parcel masks, boundaries, and distance maps [16]. BSNet fuses boundary and semantic representations for farmland parcel mapping [32]. A semantic edge-aware multitask network jointly learns region and edge representations while aggregating local and global edge semantics [11]. Boundary-specific losses further penalize contour displacement or progressively move predicted boundaries toward their labels [10, 29]. Collectively, these methods improve geometric delineation through complementary structural predictions or direct boundary constraints. However, structural supervision alone provides no explicit criterion for distinguishing genuine rice structure from visually ambiguous boundary and interior responses, especially when terrain-induced appearance variations obscure or imitate field patterns. Topographic context offers complementary evidence of rice–terrain compatibility, but how to couple it with boundary and interior cues for joint refinement remains underexplored. To address these challenges, we propose an asymmetric framework guided by topography, called the Topography-Regulated Network (TRNet). Its two components are Topographic Energy-Spectral Rectification (TESR) and the Topography-guided Paddy Structure Decoder (TPSD). TESR supplies a topographic criterion for frequency-specific visual regulation, whereas TPSD couples boundary and interior cues with coarse topographic context. Its contributions are: • An asymmetric terrain-as-context design integrates VHR imagery with coarser topographic data. Separate visual and terrain encoders preserve modality-specific representations: RGB provides the primary semantic and spatial evidence, while DEM and slope regulate visual interpretation. • TESR uses terrain-conditioned low-frequency modulation and asymmetric high-frequency regulation to suppress steep-slope clutter and conditionally enhance compatible low-slope rice cues. TPSD couples boundary and interior-depth cues with coarse topographic context, allowing visual and topographic residuals to jointly refine rice–background boundaries and region interiors. Both modules are trained from binary rice masks. • Experiments on the Area A internal test set and geographically held-out Area B yield Rice IoU values of 85.10% and 80.68%, outperforming Dual-Encoder U-Net by 9.15 and 18.83 percentage points, respectively. Component, frequency-path, decoder, and slope-stratified analyses further support the complementary roles of TESR and TPSD, including lower background false-positive and rice false-negative rates on steep terrain. Figure 1: Study areas in Hongya County, China: (a) county location and elevation, (b) Areas A and B, and representative GaoJing-1 imagery of (c) Area A and (d) Area B. I Study areas and datasets I-A Study areas The study was conducted in mountainous and hilly Hongya County, Meishan City, Sichuan Province, China. Two geographically separate subareas were selected (Fig. 1). Area A supported training, validation, and internal testing, whereas Area B was reserved for held-out cross-area testing. Area A contains extensive agricultural land and frequent rice fields, with a mean slope of 8.28∘8.28 . Area B is steeper and rice-sparser, with a mean slope of 25.90∘25.90 . Differences in terrain, land cover, and rice prevalence make Area B a combined cross-area shift rather than an isolated topographic test. I-B Data sources and preprocessing The optical data comprised orthorectified, pan-sharpened 0.50.5-m GaoJing-1 (SuperView-1) RGB images acquired in July 2023. Terrain information came from the commercial 55-m TanDEM-X dataset acquired in July 2023 at its original spatial resolution. The TanDEM-X processing context followed Rossi et al. [25], and slope in degrees was calculated from the DEM in ArcGIS Pro. All layers used WGS 84/UTM Zone 48N (EPSG:32648). DEM and slope were bilinearly resampled to 0.50.5 m and co-registered with RGB for grid alignment without adding terrain detail beyond the source DEM resolution. The five-channel RGB–DEM–slope input was clipped for outliers and standardized with fixed training-set channel statistics. The same statistics recover slope in degrees for TESR, while the original degree-valued slope is retained for TPSD and the terrain-focused loss (Section I). I-C Dataset construction and split Thirty annotators from an agricultural research institute interpreted the 0.50.5-m RGB imagery without terrain data, and the resulting annotations and co-registered layers were divided into 1,762 non-overlapping 512×512512× 512-pixel patches. Each patch contained a five-channel GeoTIFF and PNG mask. The 1,462 Area A patches were split into 1,169 training, 146 validation, and 147 internal-test patches, with rice comprising 16.51%, 15.61%, and 17.25% of valid pixels, respectively. This fixed split was used in all experiments. Non-overlap prevents pixel duplication, although the random Area A split measures within-area rather than spatially independent transfer. The 300 geographically separate Area B test patches were excluded from model development and contained 4.80% rice. Rice pixels concentrate on flatter terrain: 68.6% occur below 2∘2 , and only 3.2% occur at slopes of at least 15∘15 (Fig. 2). Correspondingly, rice prevalence declines from 26.2% to 3.7% across these intervals. This strong but non-exclusive slope prior motivates terrain-aware modulation without treating slope as a hard decision rule. Figure 2: Empirical relationship between slope and paddy occurrence in the Area A training split. (a) Class-conditional slope distributions, normalized separately for rice and background. (b) Fraction of rice pixels within each slope interval. Values are percentages over valid pixels. I Methods I-A Overall architecture of TRNet Figure 3: TRNet architecture. RGB and DEM–slope streams feed five aligned encoder stages with 64, 128, 256, 512, and 1024 channels; degree-valued slope supports TESR, TPSD, and the terrain-aware loss. Terrain-to-RGB fusion operates at E0E_0 and E2E_2–E4E_4. At E1E_1, TESR uses T1T_1 and degree-valued slope to produce E1′E_1 (shown as RE1RE_1) for E2E_2 and the decoder skip. The single RGB decoder combines the fused E4E_4 bottleneck with skips from E3E_3, E2E_2, E1′E_1 , and E0E_0. TPSD refines the prediction using boundary, interior, and topographic cues. Training combines boundary WBCE + Dice, eroded-interior Smooth L1, slope-aware asymmetric cross-entropy, and foreground Dice losses. TRNet is inspired by the dual-encoder paradigm of Lu et al. [17]. Both architectures use modality-specific encoders and asymmetric cross-modal interaction within a U-Net framework, but they assign different roles to the two modalities. In the original network, optical features enter the DEM companion encoder and the resulting companion features feed an attention-gated U-Net decoder. In TRNet, terrain features instead regulate the RGB pathway, allowing topography to provide contextual guidance while RGB remains the primary source of semantic and spatial evidence (Fig. 3). TRNet implements this role assignment across five resolution-aligned encoder stages. At stages other than E1E_1, terrain and visual features are concatenated and projected into the RGB pathway. At E1E_1, TESR replaces this interaction and sends its single rectified output to the next RGB stage and the corresponding decoder skip. Only the RGB pathway has a decoder, while the degree-valued slope map is retained outside the backbone for TESR, TPSD, and the terrain-aware training objectives. For controlled comparisons, we define TRNet-Base as the same asymmetric backbone and U-Net-style decoder without either proposed module. TRNet-Base restores the standard concatenation–projection interaction at E1E_1 in place of TESR and maps the final decoder feature directly to semantic logits with a conventional 1×11× 1 classifier in place of TPSD. It is therefore the ablation reference for TESR and TPSD, whereas the original Dual-Encoder U-Net is retained only as a prior-art comparison. For an input sample of spatial size H×WH× W, let =[rgb,dem,enc]∈ℝ5×H×W, =[I_rgb,D_dem,S_enc] ^5× H× W, (1) =[dem,enc]∈ℝ2×H×W. =[D_dem,S_enc] ^2× H× W. denote the five-channel input and its terrain subset. Here, rgb∈ℝ3×H×WI_rgb ^3× H× W, whereas demD_dem and encS_enc are the standardized DEM and slope channels. The degree-valued slope map ∘∈ℝ1×H×WS ^1× H× W is retained separately for TESR, TPSD, and the terrain-focused loss. Let lR_l and lT_l denote the RGB and terrain encoder outputs at stage l, respectively, and let lF_l denote the feature propagated along the RGB pathway after cross-modal interaction. At the full-resolution stage, 0=Ev0(rgb),0=Et0().R_0=E_v^0 (I_rgb ), _0=E_t^0 (T ). (2) PlP_l is a 1×11× 1 Conv–BatchNorm–ReLU projection that maps the concatenated 2Cl2C_l channels back to ClC_l channels. 0=P0([0∥0]).F_0=P_0 ( [R_0 \|T_0 ] ). (3) At the half-resolution stage, the two encoders first compute 1=Ev1(0),1=Et1(0).R_1=E_v^1 (F_0 ), _1=E_t^1 (T_0 ). (4) We identify 1R_1 with the unrectified visual feature 1E_1 in Fig. 3. TESR replaces the concatenation–projection operation at this stage and uses ∘S as an auxiliary physical anchor: 1=1′=TESR(1,1;∘).F_1=E_1 =TESR (R_1,T_1;S ). (5) The output 1F_1 is used directly, without concatenation with 1R_1 or 1T_1. For the remaining stages, the terrain encoder continues independently, whereas each RGB stage receives the preceding cross-modal feature: l=Evl(l−1),l=Etl(l−1),l∈2,3,4.R_l=E_v^l (F_l-1 ), _l=E_t^l (T_l-1 ), l∈\2,3,4\. (6) These modality-specific outputs are concatenated and projected to update the feature propagated along the RGB pathway: l=Pl([l∥l]),l∈2,3,4.F_l=P_l ( [R_l \|T_l ] ), l∈\2,3,4\. (7) The aligned encoder outputs and propagated features follow l,l,l∈ℝCl×H/2l×W/2l, _l,T_l,F_l ^C_l× H/2^l× W/2^l, (8) Cl=2l+6,l∈0,…,4. C_l=2^l+6, l∈\0,…,4\. The features 0F_0–3F_3 are retained as decoder skips, and 4F_4 is the fused bottleneck. The four-stage visual decoder starts from 4F_4. We set D0=4F_D_0=F_4 and compute Dl=ϕl([Upl(Dl−1)∥l]),l∈1,…,4.F_D_l= _l ( [Up_l (F_D_l-1 ) \|K_l ] ), l∈\1,…,4\. (9) where (1,2,3,4)=(3,2,1,0). (K_1,K_2,K_3,K_4 )= (F_3,F_2,F_1,F_0 ). (10) UplUp_l denotes twofold bilinear interpolation followed by channel reduction, and ϕl _l fuses the upsampled feature with the same-resolution skip. The decoder outputs have 512, 256, 128, and 64 channels from D1F_D_1 to D4F_D_4, respectively. TPSD takes the full-resolution feature D4F_D_4 and the degree-valued slope map ∘S and returns (seg,b,int)=TPSD(D4,∘). (Z_seg,Z_b,Z_int )=TPSD (F_D_4,S ). (11) All outputs are H×WH× W. seg∈ℝ2×H×WZ_seg ^2× H× W contains background and rice logits, whereas bZ_b and intZ_int are one-channel auxiliary logits used during training. Inference exposes only the two-class segmentation output. I-B Topographic Energy-Spectral Rectification Figure 4: Topographic Energy-Spectral Rectification (TESR) at the E1/T1E_1/T_1 interaction point. Both feature inputs have shape 128×h×w128× h× w. Fixed Haar DWT decomposes E1E_1 into LL, LHLH, HLHL, and HH, each of shape 128×h/2×w/2128× h/2× w/2. A stride-2 terrain projection maps T1T_1 to TfT_f. The low-frequency path predicts γ and β for LL′=(1+tanhγ)⊙LL+βL =(1+ γ) L+β. In the high-frequency path, high- and low-slope priors PhP_h and PlP_l are combined with visual compatibility CrC_r and a high-slope-attenuated learned residual to form the terrain-guided energy E. The monotonically decreasing gate G=σ(−E/TE)G=σ(-E/T_E) is smoothed by size-preserving 3×33× 3 average pooling to obtain G¯ G, from which the centered gain Ahf=1+α(2G¯−1)A_hf=1+α(2 G-1) is shared by LHLH, HLHL, and HH. Fixed Haar iDWT reconstructs the sole output E1′E_1 , which proceeds directly to the next RGB encoder stage and the corresponding decoder skip. TESR is motivated by two asymmetric terrain-related risks: texture-induced false responses on steep slopes and weak rice responses on flatter terrain. Accordingly, it suppresses terrain-inconsistent high-frequency responses at high slopes while permitting confidence-conditioned enhancement at low slopes; Fig. 4 summarizes this asymmetric rectification process. Existing remote-sensing segmentation methods use wavelet decomposition to combine spatial and frequency-domain representations or to reintroduce enhanced high-frequency subbands [37, 12], while cross-modal rectification methods use one modality to calibrate features from another [40]. TESR couples these two lines of work to an explicit physical anchor: terrain-conditioned FiLM modulates low-frequency visual responses, whereas degree-valued slope drives an asymmetric terrain-guided gate that suppresses high-slope detail and permits only confidence-conditioned enhancement at low slopes. The inverse transform then reconstructs a single rectified feature for the visual pathway rather than treating the terrain stream as a second fine-resolution representation. TESR takes the aligned shallow features =1≡1X=E_1 _1 and =1U=T_1, where ,∈ℝ128×h×wX,U ^128× h× w, h=H/2h=H/2, and w=W/2w=W/2. The retained slope map ∘S provides an explicit physical anchor. A fixed Haar transform decomposes the visual feature into one low-frequency and three directional high-frequency subbands [18]. Terrain-conditioned FiLM rectifies the low-frequency subband, while terrain-guided energy controls the high-frequency subbands. TESR reconstructs the single rectified visual output 1′≡1E_1 _1. Formally, TESR first applies a fixed, channel-wise Haar discrete wavelet transform: (LL,LH,HL,HH)=DWTHaar(). (X_L,X_LH,X_HL,X_H )=DWT_Haar (X ). (12) Each subband has shape 128×h/2×w/2128× h/2× w/2. LLX_L carries coarse visual semantics, whereas LHX_LH, HLX_HL, and HHX_H encode directional feature variations rather than image-space boundaries. We define the terrain projection as f=ϕt().T_f= _t (U ). (13) ϕt _t is a stride-2 3×33× 3 Conv–BatchNorm–ReLU block. Its output fT_f has shape 64×h/2×w/264× h/2× w/2. fT_f predicts spatially varying FiLM scale γ and bias β. FiLM applies a conditioning-dependent feature-wise affine transformation [23]: [;]=ϕlf(f). [ γ; β ]= _lf (T_f ). (14) ϕlf _lf uses 3×33× 3 Conv–ReLU and 1×11× 1 Conv, then splits its output into two 128-channel maps aligned with LLX_L. These parameters rectify the low-frequency subband as ~LL=(1+tanh())⊙LL+. X_L= (1+ ( γ ) ) _L+ β. (15) The componentwise multiplicative factor lies in (0,2)(0,2) and therefore provides bounded suppression or amplification, while β shifts the feature baseline. The high-frequency path first resizes the degree-valued slope map to the wavelet grid: w=Resizebilinearh/2×w/2(∘).S_w=Resize_bilinear^h/2× w/2 (S ). (16) High- and low-slope soft priors are defined by h _h =σ(w−θhτh), =σ ( S_w- _h _h ), θh _h =10∘, =0 , (17) l _l =σ(θl−wτl), =σ ( _l-S_w _l ), θl _l =5∘. =5 . Here, σ(x)=(1+exp(−x))−1σ(x)=(1+ (-x))^-1 denotes the sigmoid function; θh _h and θl _l are the fixed high- and low-slope thresholds, respectively; and τh,τl>0 _h, _l>0 control their transition widths. Because low slope alone does not imply rice presence, a visual compatibility gate is predicted from the low-frequency feature: r=σ(ϕc(LL)).C_r=σ ( _c (X_L ) ). (18) Let whw_h and wlw_l be independent unconstrained learnable scalars for the high-slope suppression and low-slope enhancement branches, respectively. Their positive strengths are ah=softplus(wh),al=softplus(wl).a_h=softplus (w_h ), a_l=softplus (w_l ). (19) Here, softplus(x)=log(1+exp(x))softplus(x)= (1+ (x)), ensuring that ah,al>0a_h,a_l>0 throughout optimization. The physically interpretable signed energy is phys=ahh−all⊙r.E_phys=a_hP_h-a_lP_l _r. (20) The positive high-slope term favors high-frequency suppression, whereas the negative low-slope term permits enhancement only where the visual feature is compatible with rice. A learned terrain energy retains data-driven flexibility outside the high-slope regime: raw=ϕhf(f),res=b(1−h)⊙tanh(raw),E_raw= _hf (T_f ), _res=b (1-P_h ) (E_raw ), (21) where b>0b>0 bounds the residual amplitude. Multiplication by 1−h1-P_h prevents the learned residual from overriding the high-slope anchor. The final energy is =phys+res.E=E_phys+E_res. (22) It is first converted into a monotonically decreasing gate: =σ(−TE),TE=0.35.G=σ (- ET_E ), T_E=0.35. (23) Here, TET_E is the energy-temperature parameter that controls the sensitivity of the mapping from E to the gate values G. A size-preserving local average pool suppresses isolated fluctuations in the gate map: ¯=AvgPool3×3s=1,p=1(). G=AvgPool^s=1,p=1_3× 3 (G ). (24) The smoothed gate is then mapped to a centered high-frequency gain: hf=1+α(2¯−1).A_hf=1+α (2 G-1 ). (25) Here, α>0α>0 controls the modulation range. Positive, negative, and zero energy yield <0.5G<0.5, >0.5G>0.5, and =0.5G=0.5, respectively, before local smoothing. Consequently, hf<1A_hf<1, hf>1A_hf>1, and hf=1A_hf=1 denote suppression, enhancement, and neutrality. The gain is shared across all three directional subbands: ~LH X_LH =hf⊙LH, =A_hf _LH, (26) ~HL X_HL =hf⊙HL, =A_hf _HL, ~HH X_H =hf⊙HH. =A_hf _H. TESR reconstructs the rectified feature with the inverse Haar transform: 1′=iDWTHaar(~LL,~LH,~HL,~HH).E_1 =iDWT_Haar ( X_L, X_LH, X_HL, X_H ). (27) The reconstructed 1′E_1 retains the 128×h×w128× h× w input shape. I-C Topography-guided Paddy Structure Decoder Figure 5: Topography-guided Paddy Structure Decoder (TPSD). The full-resolution decoder feature is first transformed into the base feature F. (A) The visual branch derives contextual and high-detail contrasts from F to predict interior-depth and boundary maps, which are fed back through visual refinement to form Δv _v. (B) The topographic branch combines coarse slope cS_c, normalized Sobel magnitude cG_c, and cross-scale consistency cR_c to construct coarse context cC_c. The context and consistency paths are upsampled separately and combined with the boundary-free propagation mask to form Δt _t. (C) Learnable tanh -scaled visual and topographic residuals are added to F before dropout and the final 1×11× 1 two-class classifier. In the diagram, PiP_i and PbP_b denote the auxiliary interior-depth and boundary probabilities supervised during training. TPSD is motivated by the need to refine rice–background boundaries without fragmenting internally coherent rice regions. It therefore combines boundary and interior cues with coarse topographic context, as summarized in Fig. 5. Existing agricultural-field decoders improve spatial geometry through joint mask, boundary, and distance-map prediction or through boundary–semantic feature fusion [16, 11, 32]. TPSD builds on structure-aware decoding but couples a decoder-derived structural residual with an explicit topographic residual: boundary and interior-depth predictions refine the shared decoder feature, while cross-scale consistency and predicted boundaries regulate the propagation of coarse slope context. The two residuals therefore jointly refine semantic structure without treating the coarse terrain map as fine boundary evidence. TPSD takes D4∈ℝ64×H×WF_D_4 ^64× H× W and degree-valued slope ∘S . Here, paddy structure refers to rice–background boundaries and an erosion-derived measure of interior depth. It does not denote cadastral or instance boundaries, which are absent from the binary mask. A 3×33× 3 Conv–SyncBN–ReLU block produces =ϕfcn(D4),F= _fcn (F_D_4 ), (28) where ∈ℝ64×H×WF ^64× H× W. TPSD contrasts fine local variation with broader context using stride-one, size-preserving pooling: smooth5=AvgPool5×5s=1().F_smooth^5=AvgPool^s=1_5× 5 (F ). (29) The broad-neighborhood context contrast and fine-detail contrast are then computed as ctx _ctx =smooth5−, =F_smooth^5-F, (30) high _high =−AvgPool3×3s=1(). =F-AvgPool^s=1_3× 3\! (F ). ctxF_ctx and highF_high capture broad-context contrast and fine deviations. Together with F, they form a shared 64-channel feature: s=ϕs([,ctx,high]).F_s= _s ( [F,F_ctx,F_high ] ). (31) ϕs _s is a 3×33× 3 Conv–BN–ReLU block followed by a 1×11× 1 convolution. This shared representation is used by both structural branches to predict the rice–background boundary and rice-region interior-depth maps: b=ϕb(s),b=σ(b).Z_b= _b (F_s ), _b=σ (Z_b ). (32) ϕb _b and ϕint _int are independent heads; each uses 3×33× 3 Conv–BN–ReLU to reduce 64 channels to 32 and a 1×11× 1 convolution to produce one logit channel. int=ϕint(s),int=σ(int).Z_int= _int (F_s ), _int=σ (Z_int ). (33) Their probabilities are fed back into visual refinement: Δv=ϕv([,s,b,int]). _v= _v ( [F,F_s,P_b,P_int ] ). (34) ϕv _v maps the concatenated inputs to a 64-channel residual. Thus, both structural branches affect the segmentation logits directly. The topographic path forms a detached, normalized slope reference that controls scaling rather than class suitability: n=stopgrad(clip(∘30∘,0,1)).S_n=stopgrad (clip ( S 30 ,0,1 ) ). (35) Adaptive averaging produces the coarse slope map: c=AAPH/8×W/8(n).S_c=AAP_H/8× W/8 (S_n ). (36) Resizing cS_c permits a full-to-coarse consistency comparison: ^n=ResizebilinearH×W(c). S_n=Resize_bilinear^H× W (S_c ). (37) The reconstruction discrepancy and its coarse-scale form are s=|n−^n|,s,c=AAPH/8×W/8(s). _s= |S_n- S_n |, _s,c=AAP_H/8× W/8 ( _s ). (38) The cross-scale slope-consistency weight is defined as c=stopgrad(exp(−τs,c)),τ=6.R_c=stopgrad ( (-τ _s,c ) ), τ=6. (39) cR_c is a deterministic slope-consistency weight, not a DEM-accuracy probability; larger values indicate closer reconstruction. Per-sample normalized Sobel magnitude gives the coarse transition map: c=MinMax(‖∇Sobelc‖2).G_c=MinMax ( \| _SobelS_c \|_2 ). (40) MinMaxMinMax is numerically stabilized and maps constant inputs to zero. The three coarse maps form a 64-channel context feature: c=ϕc([c,c,c]).C_c= _c ( [S_c,G_c,R_c ] ). (41) ϕc _c uses two 3×33× 3 Conv–BN–ReLU blocks with 32 hidden channels and a 1×11× 1 convolution. The context and consistency maps are resized to the full decoder resolution: =ResizebilinearH×W(c), =Resize_bilinear^H× W (C_c ), (42) topo _topo =ResizebilinearH×W(c). =Resize_bilinear^H× W (R_c ). The transition information in cG_c is already encoded in cC_c and therefore reaches the full-resolution residual through C; cG_c is not upsampled as a separate residual factor. The boundary-aware propagation mask is p=1−stopgrad(b).M_p=1-stopgrad (P_b ). (43) The topographic residual is defined by Δt=⊙topo⊙p. _t=C _topo _p. (44) This suppresses topographic propagation at inconsistent slopes and predicted boundaries. Detaching bP_b prevents segmentation gradients from shifting boundaries toward coarse context. The visual and topographic residuals are injected into the base feature through two learnable scalar parameters, αs _s and αt _t, respectively: r=+tanh(αs)Δv+tanh(αt)Δt.F_r=F+tanh\! ( _s ) _v+tanh\! ( _t ) _t. (45) The classifier is seg=Conv1×1(Dropout(r)).Z_seg=Conv_1× 1 (Dropout (F_r ) ). (46) I-D Training objectives Let ∈0,1,255H×WY∈\0,1,255\^H× W encode background, rice, and ignored pixels. The valid-pixel and rice masks are defined as valid=[≠255],rice=[=1].M_valid= [Y≠ 255 ], _rice= [Y=1 ]. (47) Let Ωv _v be the set of valid pixels, and let yi∈0,1y_i∈\0,1\ be the rice label at pixel i∈Ωvi∈ _v. The overall training objective is ℒ=ℒTCE+ℒDice+0.5ℒboundary+0.2ℒinterior. =L_TCE\!+L_Dice\!+0.5L_boundary\!+0.2L_interior. (48) For degree-valued slope si∘s_i and logits i∈ℝ2z_i ^2, high- and low-slope memberships with slope-sensitivity parameter k>0k>0 are hi=σ(k(si∘−θh)),θh=10∘.h_i=σ\! (k(s_i - _h) ), _h=10 . (49) The complementary membership assigns larger weights to pixels increasingly below the low-slope threshold: li=σ(k(θl−si∘)),θl=5∘.l_i=σ\! (k( _l-s_i ) ), _l=5 . (50) Both lie in (0,1)(0,1) and provide overlapping soft pixel weights without affecting inference. The final configuration uses k=0.75k=0.75 for both membership functions. The pixel weight is wi=1+wfp(1−yi)hi+wfnyili.w_i=1+w_fp(1-y_i)h_i+w_fny_il_i. (51) Here, wfpw_fp and wfnw_fn control the emphasis on high-slope background and low-slope rice, respectively. The final configuration sets wfp=1.0w_fp=1.0 and wfn=0.25w_fn=0.25. The resulting slope-aware asymmetric cross-entropy loss is ℒTCE=1|Ωv|∑i∈ΩvwiCE(i,yi).L_TCE= 1| _v| _i∈ _vw_iCE (z_i,y_i ). (52) This weighting emphasizes high-slope background and low-slope rice independently of prediction errors. The Dice loss directly optimizes foreground overlap [19]. Let pi=Softmax(i)1,p_i=Softmax (z_i )_1, (53) denote rice probability. The final configuration disables low-slope positive weighting in the Dice term (wlow,pos=0w_low,pos=0), so the model uses the standard foreground Dice loss: ℒDice=1−2∑i∈Ωvpiyi+ϵ∑i∈Ωvpi+∑i∈Ωvyi+ϵ.L_Dice=1- 2 _i∈ _vp_iy_i+ε _i∈ _vp_i+ _i∈ _vy_i+ε. (54) Multi-task agricultural field delineation has used boundary supervision alongside semantic prediction [16]. Here, the structural boundary target is generated from the binary rice mask using dilation and erosion: gt=(Dilater=2(rice)−Eroder=2(rice))⊙valid.B^gt= (Dilate_r=2 (M_rice )-Erode_r=2 (M_rice ) ) _valid. (55) This gives a finite-width semantic boundary without introducing instance boundaries. Eight erosions define the interior target: rice(0)=rice,rice(k)=Erode3×3(rice(k−1)),k∈1,…,8. gatheredM_rice^(0)=M_rice,\\ M_rice^(k)=Erode_3× 3 (M_rice^(k-1) ), k∈\1,…,8\. gathered (56) Averaging these erosion-survival masks gives a truncated morphological depth proxy rather than a Euclidean distance or instance annotation: gt=18∑k=18rice(k).I^gt= 18 _k=1^8M_rice^(k). (57) For N+N_+ boundary and N valid pixels, stabilized counts are N~+ N_+ =max(N+,1), = (N_+,1 ), (58) N~ N =max(N,1), = (N,1 ), N~− N_- =max(N~−N~+,1). = ( N- N_+,1 ). The class-balanced boundary weight is ωb,i=N~−N~Bigt+N~+N~(1−Bigt). _b,i= N_- NB_i^gt+ N_+ N (1-B_i^gt ). (59) The normalized weighted boundary loss is ℒWBCE=∑i∈Ωvωb,iℓBCE(Bigt,Pb,i)max(∑i∈Ωvωb,i,1).L_WBCE= _i∈ _v _b,i _BCE\! (B_i^gt,P_b,i ) \! ( _i∈ _v _b,i,1 ). (60) Here, ℓBCE(y,p) _BCE(y,p) denotes the standard binary cross-entropy; the denominator normalizes the boundary weights and remains finite for degenerate samples. The boundary Dice loss is ℒbDice=1−2∑i∈ΩvPb,iBigt+ϵ∑i∈ΩvPb,i+∑i∈ΩvBigt+ϵ.L_bDice=1- 2 _i∈ _vP_b,iB_i^gt+ε _i∈ _vP_b,i+ _i∈ _vB_i^gt+ε. (61) The boundary objective is ℒboundary=0.5ℒWBCE+ℒbDice.L_boundary=0.5L_WBCE+L_bDice. (62) For interior probability Pint,iP_int,i, the pixel weight is qi=Mvalid,i(0.25+0.75yi).q_i=M_valid,i (0.25+0.75y_i ). (63) For di=Pint,i−Iigtd_i=P_int,i-I_i^gt, the Smooth L1 term is ℓint,i=12di2,|di|<1,|di|−12,|di|≥1. _int,i= cases 12d_i^2,&|d_i|<1,\\ |d_i|- 12,&|d_i|≥ 1. cases (64) The interior objective is ℒinterior=∑iqiℓint,imax(∑iqi,1).L_interior= _iq_i _int,i ( _iq_i,1 ). (65) The structural losses supervise predictions that also refine the final segmentation features. I-E Evaluation metrics Paddy rice is the positive class; TPTP, FPFP, and FNFN are the corresponding confusion counts. Precision is the fraction of predicted rice pixels that are correct: Precision=TPTP+FP,Precision= TPTP+FP, (66) Recall is the fraction of reference rice pixels that are recovered: Recall=TPTP+FN,Recall= TPTP+FN, (67) The F1F_1 score is the harmonic mean of precision and recall: F1=2×Precision×RecallPrecision+Recall,F_1= 2×Precision×RecallPrecision+Recall, (68) Rice IoU measures predicted–reference region overlap: IoU=TPTP+FP+FN,IoU= TPTP+FP+FN, (69) Mean IoU averages the rice and background IoU values: mIoU=IoUrice+IoUbackground2.mIoU= IoU_rice+IoU_background2. (70) For the complementary boundary metric, let ℬpB_p and ℬgB_g be the one-pixel rice–background contours extracted from the predicted and reference masks. With Euclidean distance d(⋅,ℬ)d(·,B) to the nearest contour pixel and tolerance δ=2δ=2 pixels, boundary precision and recall are Pb P_b =∑∈ℬp[d(,ℬg)≤δ]|ℬp|, = _x _pI\! [d(x,B_g)≤δ ] |B_p |, (71) Rb R_b =∑∈ℬg[d(,ℬp)≤δ]|ℬg|. = _x _gI\! [d(x,B_p)≤δ ] |B_g |. The corresponding boundary score is F1boundary=2PbRbPb+Rb.F_1^boundary= 2P_bR_bP_b+R_b. (72) For the terrain diagnostic, define the valid high-slope background set as Ω0,≥15∘=i∈Ωv:yi=0,si∘≥15∘. _0,≥ 15 = \i∈ _v:y_i=0,\;s_i ≥ 15 \. (73) The high-slope false-positive rate is FPR≥15∘=∑i∈Ω0,≥15∘[y^i=1]|Ω0,≥15∘|,FPR_≥ 15 = _i∈ _0,≥ 15 I\! [ y_i=1 ] | _0,≥ 15 |, (74) where y^i y_i is the predicted class. This is the fraction of valid background pixels at slopes of at least 15∘15 misclassified as rice. Confusion counts exclude label-255 pixels. Rice IoU is primary because it measures foreground overlap under class imbalance; the other metrics characterize commission, omission, their balance, and class-average overlap. IV Experiments and Analysis IV-A Experimental settings All models use the splits in Section I-C. Training and validation-based checkpoint selection use Area A; evaluation uses its internal test set and held-out Area B. RGB-only and multimodal models use three and five channels, respectively. All models are trained from scratch under a common protocol and evaluated without post-processing. Dual-Encoder U-Net, TRNet-Base, and TRNet are each trained five times with independent random seeds; their Rice IoU values in both areas are reported as mean ± standard deviation. Training uses 512×512512× 512 patches with random scaling, cropping, and horizontal flipping. Stochastic gradient descent runs for 30,000 iterations with batch size 2, initial learning rate 0.003, momentum 0.9, and weight decay 0.0005. A polynomial schedule follows 1,000 warm-up iterations. Whole-image inference uses an NVIDIA GeForce RTX 3090. Decoder and objective ablations also report rice–background Boundary F1 with 2-pixel tolerance and high-slope FPR over background pixels with slope ≥15∘≥ 15 . Settings are fixed within each ablation group. Rice IoU uses the main semantic evaluator; the other metrics are complementary diagnostics from the same checkpoint. IV-B Comparison with baseline methods The CNN baselines are U-Net [24], DeepLabV3+ [2], and OCRNet [39]. The Transformer baselines are SegFormer [35] and GloTS [14]. The state-space-based baselines are SegMAN [6], which combines sliding local attention with dynamic state-space modeling, and Samba [45]. We also evaluate the remote-sensing model FADNet [13]. All baseline methods are reproduced from their official code releases under the unified experimental protocol described above. RGB-only and early-concatenation RGB–DEM–slope variants isolate the effect of added terrain channels. The original Dual-Encoder U-Net serves as a prior-art comparison. TRNet-Base retains our redesigned asymmetric backbone and decoder but replaces TESR and TPSD with the standard interaction and semantic head defined in Section I-A; it serves as the controlled architectural baseline for all module ablations. TABLE I: Area A internal-test comparison. RGB+T denotes early RGB–DEM–slope concatenation; Separate denotes modality-specific encoder streams before cross-modal interaction. Rice IoU for Dual-Encoder U-Net, TRNet-Base, and TRNet is reported as mean ± standard deviation over five runs; the remaining values are percentages from a single run. Method Input/fusion Rice IoU Rice F1 Rice Precision Rice Recall mIoU DeepLabV3+ RGB+T 62.35 76.81 81.52 72.62 76.86 DeepLabV3+ RGB 62.87 77.20 81.20 73.59 77.14 FADNet RGB 63.01 77.31 84.15 71.49 77.38 OCRNet RGB 63.65 77.79 83.44 72.86 77.72 SegMAN RGB+T 65.05 78.82 83.13 74.94 78.54 GloTS RGB+T 66.38 79.79 84.40 75.66 79.39 FADNet RGB+T 66.52 79.89 79.84 79.95 79.24 GloTS RGB 66.60 79.95 82.31 77.73 79.42 OCRNet RGB+T 67.05 80.28 77.45 83.32 79.40 SegFormer RGB+T 67.65 80.70 80.79 80.62 79.96 SegMAN RGB 68.07 81.01 82.48 79.58 80.31 SegFormer RGB 69.73 82.16 87.44 77.49 81.51 Samba RGB 73.75 84.89 81.42 88.68 83.66 U-Net RGB+T 74.33 85.27 85.96 84.59 84.21 Samba RGB+T 75.29 85.90 84.93 86.89 84.75 U-Net RGB 75.48 86.03 84.73 87.36 84.86 Dual-Encoder U-Net Separate 75.95 ± 0.28 86.33 86.17 86.48 85.20 TRNet-Base Separate 75.99 ± 0.11 86.36 82.50 90.60 85.07 TRNet Separate 85.10 ± 0.26 91.95 89.59 94.43 90.85 On Area A, TRNet exceeds the original Dual-Encoder U-Net, TRNet-Base, and the strongest RGB-only baseline by 9.15, 9.11, and 9.62 Rice-IoU points, respectively (Table I). Relative to the original network, Rice precision and recall increase by 3.42 and 7.95 points, respectively. Relative to TRNet-Base, they increase by 7.09 and 3.83 points, indicating that TRNet improves both false-positive control and rice-region coverage. Figure 6 provides the corresponding qualitative comparison on representative Area A scenes. Early terrain concatenation improves only three of eight architectures and changes mean Rice IoU by +0.18+0.18 points. TRNet’s gain therefore reflects explicit terrain–visual interaction rather than additional channels alone. Figure 6: Qualitative comparison on four representative Area A test scenes. Columns show the RGB image with ground-truth overlay, followed by error overlays for TRNet, TRNet-Base, RGB U-Net, the original Dual-Encoder U-Net, and Samba; colors denote true positives, false negatives, and false positives. TABLE I: Held-out Area B comparison. RGB+T denotes early concatenation; Separate denotes modality-specific encoder streams before cross-modal interaction. The table includes all multimodal models and the strongest Area A RGB-only model. Rice IoU for Dual-Encoder U-Net, TRNet-Base, and TRNet is reported as mean ± standard deviation over five runs; the remaining values are percentages from a single run. Method Input/fusion Rice IoU Rice F1 Rice Precision Rice Recall mIoU DeepLabV3+ RGB+T 51.26 67.78 66.39 69.23 74.00 SegFormer RGB+T 51.90 68.34 64.24 72.99 74.27 OCRNet RGB+T 53.05 69.33 63.94 75.70 74.86 GloTS RGB+T 53.24 69.49 61.66 79.58 74.88 SegMAN RGB+T 54.30 70.38 64.21 77.85 75.52 FADNet RGB+T 55.45 71.34 70.12 72.61 76.28 Samba RGB+T 56.51 72.21 64.92 81.36 76.69 U-Net RGB+T 57.08 72.68 62.70 86.44 76.92 U-Net RGB 57.75 73.22 61.11 91.32 77.20 Dual-Encoder U-Net Separate 61.85 ± 0.19 76.43 66.77 89.34 79.54 TRNet-Base Separate 64.38 ± 0.14 78.33 70.40 88.28 80.97 TRNet Separate 80.68 ± 0.24 89.31 86.06 92.81 89.78 On Area B, TRNet exceeds the original Dual-Encoder U-Net and TRNet-Base by 18.83 and 16.30 Rice-IoU points, respectively (Table I). Relative to the original network, Rice precision and recall rise by 19.29 and 3.47 points. Relative to TRNet-Base, they rise by 15.66 and 4.53 points, showing that the complete model substantially improves false-positive control while further increasing rice coverage. Area B differs in terrain, land cover, and rice prevalence but shares Area A’s imagery source, preprocessing, and annotation protocol; it therefore measures cross-area transfer rather than independent external validation. Overall, TRNet consistently outperforms the single-stream methods, the original Dual-Encoder U-Net, and TRNet-Base in Areas A and B. TRNet-Base exceeds the original network by 0.04 and 2.53 Rice-IoU points in Areas A and B, respectively, which remains substantially smaller than the gains of the complete model. Redesigning the backbone alone therefore does not explain TRNet’s improvement. Instead, TRNet improves false-positive control while maintaining or increasing rice coverage, with fewer fragmented rice strips and isolated false positives. IV-C Ablation studies IV-C1 TESR encoder placement and frequency paths TABLE I: Area A TESR ablation across encoder stages and frequency paths with TPSD enabled in all variants. LF-only retains low-frequency FiLM; HF-only retains high-frequency gating. Variant Rice IoU Rice F1 Rice Precision Rice Recall mIoU Without TESR 76.79 86.87 79.82 95.30 85.42 TESR at E0 80.19 89.01 86.30 91.89 87.77 TESR at E1 85.10 91.95 89.59 94.43 90.85 TESR at E2 77.51 87.33 80.94 94.81 85.92 TESR at E3 77.42 87.27 80.58 95.18 85.84 TESR at E4 77.16 87.10 83.76 90.73 85.83 Frequency-path variants LF-only 76.54 86.71 80.50 93.95 85.30 HF-only 77.34 87.22 81.16 94.27 85.83 Table I evaluates TESR placement across encoder stages E0E0–E4E4 and compares the complete module with its individual frequency paths on Area A. TPSD remains enabled in every row, so the table isolates changes to TESR. LF-only and HF-only retain only the low- and high-frequency paths, respectively. At E1, TESR improves Rice IoU by 8.31 points over no TESR; precision rises from 79.82% to 89.59% while recall changes from 95.30% to 94.43%, indicating fewer false positives. E1 reaches 85.10% Rice IoU and 90.85% mIoU, beating the next-best E0 placement by 4.91 and 3.08 points, whereas E2–E4 yield 77.16–77.51% Rice IoU. LF-only and HF-only reach 76.54% and 77.34% Rice IoU. Relative to them, E1 improves precision by 9.09 and 8.43 points and recall by 0.48 and 0.16 points, supporting complementary frequency paths. Figure 7: TESR effects in three high-slope scenes. Columns show the RGB image with ground-truth overlay, slope, energy gate, error overlays for TRNet-Base and TRNet-Base + TESR, and ground truth. White circles link suppressed gate responses to removed baseline false positives. Figure 7 compares TRNet-Base with its TESR-equipped variant in three high-slope, rice-sparse scenes. The white-circled regions show that TESR attenuates terrain-inconsistent high-slope responses that become isolated false positives in the variant without TESR, while sparse rice regions remain intact. The visualization therefore supports selective steep-terrain suppression rather than blanket removal. IV-C2 TPSD and training objectives Table IV compares full and reduced TPSD variants with TESR and the final objective fixed. Objective variants retain the complete architecture and cumulatively replace CE with TCE, add boundary and interior supervision, and apply the balanced weights in Eq. (48). TABLE IV: Area A TPSD and objective ablations. Boundary F1 uses 2-pixel tolerance; high-slope FPR uses background pixels at ≥15∘≥ 15 . Group Variant Rice IoU ↑ Boundary F1 ↑ High-slope FPR ↓ TPSD TESR w/o TPSD 81.48 49.90 3.11 Boundary-only TPSD 82.36 55.80 2.77 Visual-structure TPSD 84.55 60.15 2.31 Full TPSD 85.10 64.96 1.35 Objective CE + Dice 76.58 50.72 2.95 TCE + Dice 78.23 51.67 1.83 + Boundary supervision 80.22 57.46 1.87 + Interior supervision 84.57 60.13 1.58 Final balanced objective 85.10 64.96 1.35 Full TPSD exceeds the TESR-equipped decoder by 3.62 Rice-IoU and 15.06 Boundary-F1 points while reducing high-slope FPR from 3.11% to 1.35%. Against visual-structure TPSD, the stronger reduced variant, it adds 0.55 and 4.81 points and lowers FPR by 0.96 points. This supports complementary visual structure and topographic context. Figure 8: Qualitative comparison of TRNet-Base and TRNet-Base + TPSD. Columns show the RGB image with ground-truth overlay, slope, ground truth, the two predictions, and paired detail crops. Red circles highlight boundary and interior differences; white and black indicate rice and background. Figure 8 compares TRNet-Base with its TPSD-equipped variant. The red-circled crops show that TPSD restores more continuous rice–background contours and closes internal breaks, yielding regions closer to the reference. These comparisons concern semantic structure rather than instance or cadastral delineation. Replacing CE with TCE raises Rice IoU and Boundary F1 from 76.58% and 50.72% to 78.23% and 51.67%; Dice remains standard because wlow,pos=0w_low,pos=0. Boundary and interior supervision raise them to 84.57% and 60.13%. High-slope FPR declines overall despite an increase from 1.83% to 1.87% after boundary supervision. The final balanced objective reaches 85.10% Rice IoU, 64.96% Boundary F1, and 1.35% FPR. Over the strongest intermediate variant, gains are 0.53 and 4.83 points, with 0.23-point lower FPR. IV-C3 Overall component ablation Table V summarizes the overall component ablation. Adding TESR to TRNet-Base improves Rice IoU by 5.49 points, whereas adding TPSD alone yields a 0.80-point gain. Enabling both modules improves TRNet-Base by 9.11 points and exceeds the single-module variants by 3.62 and 8.31 points, respectively, supporting their complementary use. TABLE V: Overall component ablation on Area A. A check mark denotes an enabled module; Δ is the Rice IoU change from TRNet-Base. Values are percentage points. Variant TESR TPSD Rice IoU Δ TRNet-Base – – 75.99 – TRNet-Base + TESR ✓ – 81.48 +5.49 TRNet-Base + TPSD – ✓ 76.79 +0.80 Full TRNet ✓ ✓ 85.10 +9.11 IV-D Terrain-stratified and cross-area analysis We pool valid pixels from 147 Area A and 300 Area B test patches into [0,2)[0,2), [2,5)[2,5), [5,10)[5,10), [10,15)[10,15), and [15,∞)[15,∞) degree intervals. Aggregated confusion counts yield Rice IoU, background FPR, and rice FNR for RGB U-Net, RGB+T U-Net, TRNet-Base, the original Dual-Encoder U-Net, and TRNet (Fig. 9). Figure 9: Terrain-stratified performance in Area A (a–c) and Area B (d–f). Curves show Rice IoU, background FPR, and rice FNR for five models across five slope intervals using pooled valid pixels; label-255 pixels are excluded. Steep-terrain FPR evaluates TESR suppression, while low-slope Rice IoU and FNR test whether confidence-conditioned enhancement preserves rice coverage. In the lowest-slope interval, the FNR curves of TRNet are the lowest in both areas, consistent with the confidence-conditioned low-slope path of TESR. TRNet has the highest Rice IoU in every interval. Its margin over the strongest comparator grows from 7.23 to 11.76 points from the lowest to highest slopes in Area A and from 8.54 to 22.25 points in Area B. At ≥15∘≥ 15 , its background FPR is 1.35% and 0.58%, below the baseline ranges of 2.02–2.84% in Area A and 1.51–2.49% in Area B. Rice FNR is 9.45% and 8.80%, also below all baselines in the two areas. The joint reductions in false positives and false negatives support improved steep-terrain discrimination rather than a trade-off between suppression and rice coverage. The same ranking holds at 10∘10 and 15∘15 cutoffs. TRNet FPR is 1.15% and 1.35% in Area A and 0.64% and 0.58% in Area B, below the best baseline ranges of 1.83–2.02% and 1.51–1.61%, respectively. Rice FNR remains within 9.45–10.03% and 8.80–9.22%, and is lower than the corresponding baselines at both cutoffs. The joint reduction is therefore not tied to a single slope threshold. Lower steep-terrain FPR with preserved coverage supports TESR’s asymmetric operation in both areas. IV-E Integrated analysis In the evaluated setting, the results support using topography to regulate visual features rather than simply treating DEM and slope as additional input channels. Early concatenation of DEM and slope changed the mean Rice IoU of eight architectures by only +0.18+0.18 points, whereas TRNet produced consistent gains over the RGB-only baselines, the original Dual-Encoder U-Net, and TRNet-Base. Within the TESR ablation, the advantage was concentrated at E1E_1; applying the same operation at deeper stages yielded substantially smaller gains. This pattern suggests that terrain-conditioned frequency rectification is most useful while the visual representation still retains field-scale spatial detail. The interpretation remains architectural rather than causal, but it agrees with the intended division of roles: RGB supplies the fine geometry, while terrain modifies responses that are less compatible with the local topographic context. The structure and terrain diagnostics further clarify the source of the improvement. Relative to the TESR-equipped decoder, full TPSD increased Boundary F1 by 15.06 points and reduced high-slope FPR from 3.11% to 1.35%. The complete model also retained low rice omission rates in the steepest interval while suppressing background responses in both test areas. These results do not imply that the upsampled 55-m DEM delineates 0.50.5-m field boundaries. Rather, they support a complementary mechanism in which predicted visual structure limits the propagation of coarse topographic context. This distinction is important in fragmented mountain agriculture, where elevation data can indicate implausible rice locations but cannot replace the spatial detail of VHR imagery. Area B provides a demanding held-out cross-area test, but it is not independent external validation. It shares the sensor, acquisition period, preprocessing, and annotation protocol of Area A, while terrain, land cover, and rice prevalence change together. The current evidence is also limited to one county, one growing season, and binary semantic masks. Consequently, the results do not establish multi-season or cross-sensor transfer, cadastral parcel delineation, or sensitivity to DEM source and resolution. Future work should prioritize evaluation in geographically independent regions across acquisition dates and sensors, systematically examine sensitivity to the source, resolution, and noise of the DEM, and quantify variation across repeated training runs and computational cost. Instance-aware assessment would additionally require field-level annotations that are not available in the present dataset. V Conclusion TRNet shows that coarse topographic data can improve VHR paddy rice segmentation when they are used to regulate visual interpretation rather than to supply fine boundary evidence. The model achieved Rice IoU values of 85.10% on the Area A internal test set and 80.68% on held-out Area B, exceeding the original Dual-Encoder U-Net by 9.15 and 18.83 percentage points, respectively. These improvements were accompanied by better boundary delineation and fewer false positives on steep slopes, supporting the complementary roles of terrain-conditioned frequency rectification and structure-aware decoding. More broadly, the results indicate that addressing the scale mismatch between modalities can be more effective than directly fusing upsampled terrain with optical features in fragmented mountainous landscapes. Because both test areas share the same sensor, acquisition period, and annotation protocol, however, the current evidence does not establish operational generalization. Future work should prioritize evaluation in independent regions across seasons and sensors and systematically examine sensitivity to the source, resolution, and noise of the DEM. References [1] G. Chen, F. Sun, W. Jing, W. Zou, D. Di, Y. Song, and L. Fan (2025) Learning frequency-domain fusion for multimodal remote sensing semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing 63, p. 1–16. External Links: Document Cited by: §I. [2] L. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam (2018) Encoder–decoder with atrous separable convolution for semantic image segmentation. In Computer Vision – ECCV 2018, p. 833–851. External Links: Document Cited by: §IV-B. [3] R. Chen, G. Yin, W. Zhao, K. Yan, S. Wu, D. Hao, and G. Liu (2023) Topographic correction of optical remote sensing images in mountainous areas: a systematic review. IEEE Geoscience and Remote Sensing Magazine 11 (4), p. 125–145. External Links: Document Cited by: §I. [4] S. Cui, W. Chen, W. Gu, L. Yang, and X. Shi (2023) SiamC Transformer: siamese coupling Swin Transformer multi-scale semantic segmentation network for vegetation extraction under shadow conditions. Computers and Electronics in Agriculture 213, p. 108245. External Links: Document Cited by: §I. [5] F. I. Diakogiannis, F. Waldner, P. Caccetta, and C. Wu (2020) ResUNet-a: a deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing 162, p. 94–114. External Links: Document Cited by: §I. [6] Y. Fu, M. Lou, and Y. Yu (2025) SegMAN: omni-scale context modeling with state space models and local attention for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 19077–19087. Cited by: §IV-B. [7] X. Gao and P. Lu (2026) Dual-encoder multiscale transformer fusion network for landslide detection integrating Sentinel-2 spectral and topographic clues. Landslides. External Links: Document Cited by: §I. [8] Y. Gao, Y. Pan, X. Zhu, L. Li, S. Ren, C. Zhao, and X. Zheng (2023) FARM: a fully automated rice mapping framework combining Sentinel-1 SAR and Sentinel-2 multi-temporal imagery. Computers and Electronics in Agriculture 213, p. 108262. External Links: Document Cited by: §I. [9] S. Jeong, S. Kang, K. Jang, H. Lee, S. Hong, and D. Ko (2012) Development of variable threshold models for detection of irrigated paddy rice fields and irrigation timing in heterogeneous land cover. Agricultural Water Management 115, p. 83–91. External Links: Document Cited by: §I. [10] H. Kervadec, J. Bouchtiba, C. Desrosiers, E. Granger, J. Dolz, and I. Ben Ayed (2021) Boundary loss for highly unbalanced segmentation. Medical Image Analysis 67, p. 101851. External Links: Document Cited by: §I. [11] M. Li, J. Long, A. Stein, and X. Wang (2023) Using a semantic edge-aware multi-task neural network to delineate agricultural parcels from remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing 200, p. 24–40. External Links: Document Cited by: §I, §I-C. [12] Y. Li, Z. Liu, J. Yang, and H. Zhang (2023) Wavelet transform feature enhancement for semantic segmentation of remote sensing images. Remote Sensing 15 (24), p. 5644. External Links: Document Cited by: §I, §I-B. [13] J. Liu, X. Xu, M. Bilal, and J. Jiang (2024) FADNet: greenhouse identification with fusion attention mechanism and deformable convolution. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 17, p. 7170–7178. External Links: Document Cited by: §IV-B. [14] Y. Liu, Y. Zhang, Y. Wang, and S. Mei (2023) Rethinking transformers for semantic segmentation of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing 61, p. 1–15. Cited by: §IV-B. [15] Z. Liu, G. Chen, B. Tang, Q. Wen, R. Tan, and Y. Huang (2024) Regional scale terrace mapping in fragmented mountainous areas using multi-source remote sensing data and sample purification strategy. Science of the Total Environment 925, p. 171366. External Links: Document Cited by: §I. [16] J. Long, M. Li, X. Wang, and A. Stein (2022) Delineation of agricultural fields using multi-task BsiNet from high-resolution satellite images. International Journal of Applied Earth Observation and Geoinformation 112, p. 102871. External Links: Document Cited by: §I, §I-C, §I-D. [17] W. Lu, Y. Hu, Z. Zhang, and W. Cao (2023) A dual-encoder U-Net for landslide detection using Sentinel-2 and DEM data. Landslides 20, p. 1975–1987. External Links: Document Cited by: §I-A. [18] S. G. Mallat (1989) A theory for multiresolution signal decomposition: the wavelet representation. IEEE Transactions on Pattern Analysis and Machine Intelligence 11 (7), p. 674–693. External Links: Document Cited by: §I-B. [19] F. Milletari, N. Navab, and S. Ahmadi (2016) V-Net: fully convolutional neural networks for volumetric medical image segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), p. 565–571. External Links: Document Cited by: §I-D. [20] R. Ni, J. Tian, X. Li, D. Yin, J. Li, H. Gong, J. Zhang, L. Zhu, and D. Wu (2021) An enhanced pixel-based phenological feature for accurate paddy rice mapping with Sentinel-2 imagery in Google Earth Engine. ISPRS Journal of Photogrammetry and Remote Sensing 178, p. 282–296. External Links: Document Cited by: §I. [21] X. Pan, J. Xu, X. Li, and J. Zhao (2025) Highly transferable paddy field identification model based on SAR index and transformer. Computers and Electronics in Agriculture 237, p. 110790. External Links: Document Cited by: §I. [22] Y. Pan, X. Wang, L. Zhang, and Y. Zhong (2023) E2EVAP: end-to-end vectorization of smallholder agricultural parcel boundaries from high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing 203, p. 246–264. External Links: Document Cited by: §I. [23] E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville (2018) FiLM: visual reasoning with a general conditioning layer. Proceedings of the AAAI Conference on Artificial Intelligence 32 (1), p. 3942–3951. External Links: Document Cited by: §I-B. [24] O. Ronneberger, P. Fischer, and T. Brox (2015) U-Net: convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, p. 234–241. External Links: Document Cited by: §I, §IV-B. [25] C. Rossi, F. Rodriguez Gonzalez, T. Fritz, N. Yague-Martinez, and M. Eineder (2012) TanDEM-X calibrated raw DEM generation. ISPRS Journal of Photogrammetry and Remote Sensing 73, p. 12–20. External Links: Document Cited by: §I-B. [26] Z. Song, J. Lu, Y. Yue, X. Tong, and K. Wang (2026) A GAN-based method for mountain shadow restoration in optical remote sensing image. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 19, p. 21245–21254. External Links: Document Cited by: §I. [27] W. Tuo, J. Zeng, F. Wu, W. Wu, X. Wang, and X. Zhao (2026) Landslide detection using multimodal data fusion and an improved DeepLabV3+ model. Scientific Reports 16 (1), p. 1383. External Links: Document Cited by: §I. [28] F. Waldner and F. I. Diakogiannis (2020) Deep learning on edge: extracting field boundaries from satellite images with a convolutional neural network. Remote Sensing of Environment 245, p. 111741. External Links: Document Cited by: §I. [29] C. Wang, Y. Zhang, M. Cui, P. Ren, Y. Yang, X. Xie, X. Hua, H. Bao, and W. Xu (2022) Active boundary loss for semantic segmentation. Proceedings of the AAAI Conference on Artificial Intelligence 36 (2), p. 2397–2405. External Links: Document Cited by: §I. [30] L. Wang, H. Ma, Y. Gao, S. Chen, S. Yang, P. Lu, L. Fan, and Y. Wang (2024) Small- and medium-sized rice fields identification in hilly areas using all available Sentinel-1/2 images. Plant Methods 20 (1), p. 25. External Links: Document Cited by: §I. [31] L. Wang, H. Ma, J. Li, Y. Gao, L. Fan, Z. Yang, Y. Yang, and C. Wang (2022) An automated extraction of small- and middle-sized rice fields under complex terrain based on SAR time series: a case study of chongqing. Computers and Electronics in Agriculture 200, p. 107232. External Links: Document Cited by: §I. [32] S. Wang, Y. Zhou, X. Yang, L. Feng, T. Wu, and J. Luo (2023) BSNet: boundary-semantic-fusion network for farmland parcel mapping in high-resolution satellite images. Computers and Electronics in Agriculture 206, p. 107683. External Links: Document Cited by: §I, §I-C. [33] G. Wei, J. Xu, W. Yan, Q. Chong, H. Xing, and M. Ni (2024) Dual-domain fusion network based on wavelet frequency decomposition and fuzzy spatial constraint for remote sensing image segmentation. Remote Sensing 16 (19), p. 3594. External Links: Document Cited by: §I. [34] P. Wei, D. Chai, R. Huang, D. Peng, T. Lin, J. Sha, W. Sun, and J. Huang (2022) Rice mapping based on Sentinel-1 images using the coupling of prior knowledge and deep semantic segmentation network: a case study in northeast china from 2019 to 2021. International Journal of Applied Earth Observation and Geoinformation 112, p. 102948. External Links: Document Cited by: §I. [35] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo (2021) SegFormer: simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems 34, p. 12077–12090. Cited by: §IV-B. [36] J. Xu, H. Wang, L. Qiu, H. Wang, and Y. Mu (2025) Mapping county-level rice planting areas by joint use of high-resolution optical and time series SAR imagery. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18, p. 10547–10561. External Links: Document Cited by: §I. [37] Y. Yang, G. Yuan, and J. Li (2024) SFFNet: a wavelet-based spatial and frequency domain fusion network for remote sensing segmentation. IEEE Transactions on Geoscience and Remote Sensing 62, p. 1–17. External Links: Document Cited by: §I, §I-B. [38] Z. Yu, L. Di, S. Shrestha, C. Zhang, L. Guo, F. Qamar, and T. J. Mayer (2023) RiceMapEngine: a Google Earth Engine-based web application for fast paddy rice mapping. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 16, p. 7264–7275. External Links: Document Cited by: §I. [39] Y. Yuan, X. Chen, and J. Wang (2020) Object-contextual representations for semantic segmentation. In European conference on computer vision, p. 173–190. Cited by: §IV-B. [40] J. Zhang, H. Liu, K. Yang, X. Hu, R. Liu, and R. Stiefelhagen (2023) CMX: cross-modal fusion for RGB-X semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems 24 (12), p. 14679–14694. External Links: Document Cited by: §I, §I-B. [41] P. Zhang, S. Hu, W. Li, and C. Zhang (2020) Parcel-level mapping of crops in a smallholder agricultural area: a case of central china using single-temporal VHSR imagery. Computers and Electronics in Agriculture 175, p. 105581. External Links: Document Cited by: §I. [42] S. Zhang, P. Li, Y. Xie, W. Shao, and X. Tian (2025) Classification of paddy rice planting area through feature selection method using Sentinel-1/2 time series images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18, p. 8747–8762. External Links: Document Cited by: §I. [43] Y. Zhao, H. Cheng, S. Peng, S. Liu, Y. Xie, and B. Liu (2026) High-resolution (0.9 m) terrace mapping in low-latitude hilly regions using deep learning and area–slope denoising: a case study from guangdong province, china. International Journal of Applied Earth Observation and Geoinformation 146, p. 105126. External Links: Document Cited by: §I. [44] X. Zheng, X. Wu, L. Huan, W. He, and H. Zhang (2022) A gather-to-guide network for remote sensing semantic segmentation of RGB and auxiliary image. IEEE Transactions on Geoscience and Remote Sensing 60, p. 1–15. External Links: Document Cited by: §I. [45] Q. Zhu, Y. Cai, Y. Fang, Y. Yang, C. Chen, L. Fan, and A. Nguyen (2024) Samba: semantic segmentation of remotely sensed images with state space model. Heliyon 10 (19). Cited by: §IV-B.