CRD-Net: Frequency-Adaptive Feature Injection and Change Decoupling for Building Damage Assessment

Conference: ECCV2026
Paper: Official ECCV page
PDF: Full paper
Authors: Yao Zheng, Yuanxin Ye, Tan Shu, Liwei Cai
Area: Remote Sensing
Keywords: Building damage assessment, frequency-adaptive fusion, change representation decoupling, prototype contrastive learning, DINOv3

TL;DR

CRD-Net injects frequency-separated CNN details into DINOv3, separates change-related and invariant information in bitemporal features, and uses prototype-guided hard-sample contrastive learning to recognize damage, achieving overall F1 scores of 81.26% on xBD and 81.88% on EBD without leading every subtask.

Background & Motivation

Pre- and post-disaster satellite images must answer both where buildings are and how severely they are damaged. Localization relies on stable contours and structures, whereas damage classification may depend on small differences such as roof cracks or deformed edges. Visual foundation models such as DINOv3 offer strong global semantic representations, but large-stride patch embedding can discard local detail early. Upsampling low-resolution semantic features cannot recover damage cues that have already disappeared. ChangeViT addresses this through spatial detail injection, although direct fusion can also introduce redundant background information.

Preserving detail in the encoder does not make every temporal difference a damage signal. Illumination changes, background textures, and shared intact-building structures remain entangled with genuine damage, making minor damage particularly easy to overlook. ChangeMamba can establish long-range spatiotemporal relationships, but aggregation alone does not guarantee a clean change representation. Even after separating background, minor damage and intact buildings may remain close in feature space. The paper therefore addresses three successive questions: whether details enter the representation, whether irrelevant changes contaminate it, and whether visually similar categories can be separated.

Core Idea: selectively restore local structure by frequency, decouple spatiotemporal features into change and invariant subspaces, and apply different contrastive constraints to intact buildings and background so that weak damage is not reduced to a small difference within a dominant background.

Method

Overall Architecture

The input is a pair of pre- and post-disaster remote sensing images, and the outputs are a building localization mask and damage-level predictions. The encoder combines LoRA-adapted DINOv3 with an auxiliary ResNet that supplies local spatial features; Spatio-Frequency Adaptive Enhancement fuses the two representations. The localization branch progressively upsamples pre-disaster features only, preventing post-disaster appearance changes from disrupting building localization.

The damage branch builds on the ChangeMamba decoder and proceeds through Mamba Spatiotemporal Fusion, Orthogonal Subspace Decoupling, and Prototype-Guided Hard-Sample Contrast. The latter two are the added mechanisms in the CRD² decoder for cleaning representations and improving class separation. Decoupling modifies the change features passed to the prediction head, whereas contrast primarily supplies training constraints rather than a classification module that requires extra labels at inference time.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Pre- and post-disaster images<br/>DINOv3 + ResNet"] --> B["Spatio-Frequency Adaptive Enhancement"]
    B -->|Pre-disaster features| C["Building localization mask"]
    B -->|Bitemporal features| D["Mamba Spatiotemporal Fusion"]
    D --> E["Orthogonal Subspace Decoupling"]
    E --> F["Damage-level predictions"]
    E -.->|Training constraint| G["Prototype-Guided Hard-Sample Contrast"]

Key Designs

1. Spatio-Frequency Adaptive Enhancement: inject details by frequency band instead of directly adding two feature streams

DINOv3 features have a native spatial resolution of 1/16 of the input. Lightweight neck modules after layers 5, 11, 17, and 23 use standard and depthwise separable convolutions to construct a four-level pyramid at 1/4, 1/8, 1/16, and 1/32 resolution, with 96, 192, 384, and 768 channels. This supplies the multiscale inputs needed by the decoder. The independent ResNet branch, rather than upsampling itself, supplies details lost during initial embedding.

Spatio-Frequency Adaptive Feature Enhancement (SFAFE) first applies a two-dimensional FFT to same-scale CNN and DINOv3 features. Gaussian band-pass filters with learnable center frequencies, bandwidths, and gains separate low-, mid-, and high-frequency components. After inverse transformation to the spatial domain, these subbands use different fusion rules. Low and high frequencies use cross-attention with ViT queries and CNN keys and values; mid frequencies use MLP-generated gates to combine the streams. Low-frequency semantics and high-frequency structure can thus select content-dependent information from CNN features, while mid-frequency fusion adjusts their relative contributions. The module does not uniformly amplify every high-frequency response.

The fused subbands return to the frequency domain for aggregation and are then inverse-transformed into frequency-fused features. A parallel spatial branch concatenates the original streams and applies a 1×1 convolution to supplement channel semantics and mitigate numerical errors from frequency operations. CNN features, spatially fused features, and frequency-fused features jointly predict a spatial gate. The frequency and spatial outputs receive complementary weights, are concatenated, and pass through a 3×3 convolution. This second selection operates over spatial locations: a texture-rich frequency band should not cause background textures and building edges to receive equal emphasis in the decoder.

2. Mamba Spatiotemporal Fusion: establish bitemporal context before deciding which differences to retain

At each damage-decoder level, the method does not simply subtract post-disaster features from pre-disaster features. It organizes the bitemporal information through channel-wise concatenation, column-wise interleaving, and horizontal concatenation before state space modeling. These arrangements provide complementary spatiotemporal relationships, while level-wise skip connections fuse multiscale features so that small structural changes share a representation with broader building context. This follows ChangeMamba's decoding approach and supplies the input to subsequent decoupling; Mamba spatiotemporal decoding itself is not introduced by this paper.

The resulting features still contain disaster-induced changes, stable structures, and imaging disturbances. Rather than treating them as pure change, the model applies the subsequent decoupling and discrimination mechanisms at each decoder level. The localization branch instead progressively upsamples the pre-disaster pyramid. The two branches share enhanced encoder representations but optimize different objectives, and their contributions to the overall metric can involve trade-offs.

3. Orthogonal Subspace Decoupling: use invariant information to suppress spurious change rather than discard it

Orthogonal Subspace Decoupling (OSD) applies two independent projections to the fused features, producing change and invariant subspaces. The former is intended to preserve post-disaster structural changes, while the latter describes temporally stable scene content. To reduce redundancy, the module flattens and L2-normalizes the representations and penalizes their cross-correlation. The central constraint is:

\[ \mathcal{L}_{\mathrm{orth}}=\frac{1}{N}\left\|\bar Z_c\bar Z_i^{\top}\right\|_F^2. \]

Both matrices have shape \(D\times N\), where \(N\) is the number of spatial locations at the current scale and \(D\) is the projected feature dimension. The constraint encourages decorrelation, but orthogonality alone neither proves statistical independence nor guarantees that every change represents damage. The method also concatenates change and invariant features, applies a learned mapping and Sigmoid to obtain a change gate, and modulates the change features spatially. The invariant branch therefore conditions background suppression rather than supplying a fixed inverted hard mask. The intention is to reduce spurious changes from factors such as illumination while preserving locally discriminative responses.

4. Prototype-Guided Hard-Sample Contrast: distinguish resemblance to intact buildings from resemblance to background

A minor-damage region retains substantial intact-building structure even after decoupling, so requiring equal separation from every negative sample is inappropriate. Prototype-Guided Hard-sample Contrastive Learning (PHCL) maintains momentum-updated prototypes for damage, intact buildings, and background, and prioritizes hard pixels near ground-truth change boundaries. These are three coarse semantic prototypes, not one center for each of the four damage levels. Supervision of the damage prediction branch still handles the final severity categories.

To distinguish buildings from background for contrastive learning, the method uses the L2 energy of invariant-subspace features, together with the spatial mean and standard deviation of that energy map, to construct building confidence. This assumes that building textures typically produce stronger responses than flat background. For a damage anchor, intact-building negatives receive a higher base temperature to tolerate unavoidable structural similarity; background negatives receive a lower base temperature to strengthen separation. An additional temperature term decreases with normalized distance to the ground-truth change boundary. The mechanism therefore considers both negative-sample semantics and boundary uncertainty rather than mining hard examples solely by feature similarity.

The training constraint combines pixel-level InfoNCE with a prototype-center loss. The former regulates relationships between anchors and positive and negative samples, while the latter moves the mean anchor feature toward its ground-truth class prototype. Boundary and category supervision are used only during training; ground-truth change masks are not deployment inputs. Combining local contrast with momentum prototypes is intended to prevent scarce damage pixels from being dominated by the large number of background features in the current batch.

Loss & Training

Building localization combines BCE with Lovász-softmax weighted by 0.5. Damage classification combines multiclass cross-entropy with Lovász-softmax weighted by 0.75. The auxiliary objective adds orthogonality constraints across scales and PHCL, with a prototype-center loss weight of 1 and an orthogonality weight of 0.5. Localization, damage classification, and auxiliary objectives are jointly optimized. Although damage levels are ordered, this CE–Lovász combination should not be described as an explicit ordinal regression loss.

Training uses a single 48 GB NVIDIA A6000, SGD, an initial learning rate of \(1\times10^{-4}\), momentum 0.9, weight decay \(5\times10^{-4}\), batch size 8, and 100 epochs. Original xBD images of 1024×1024 are randomly cropped to 512×512, with rotation, random erasing, and Gaussian noise. The paper specifies LoRA adaptation of DINOv3 but does not fully disclose the LoRA rank, exact ResNet variant, or all contrastive temperature settings. Some cached equations also have missing symbols, so the account prioritizes mechanisms supported by the surrounding prose rather than inventing implementation details.

Key Experimental Results

Main Results

xBD uses four damage categories. EBD also has four severity levels, and the paper randomly splits it into training, validation, and test sets at 8:1:1 because no official split is provided. All values below are the paper's percentage-scale F1 scores, with higher being better. Differences between the datasets should not be interpreted directly as cross-domain generalization gains.

Localization F1 measures binary building localization. Classification F1 is the harmonic mean of the four per-category F1 scores, not their arithmetic macro-average. Overall F1 is a weighted sum of localization and classification. Equation (29) has missing symbols in the cache; checking the prose against the table values gives:

\[ F_{\mathrm{cls}}=\frac{4}{\sum_{c=1}^{4}1/F_c},\qquad F_{\mathrm{overall}}=0.3F_{\mathrm{loc}}+0.7F_{\mathrm{cls}}. \]

Dataset	Method	Localization F1 ↑	Classification F1 ↑	Overall F1 ↑	Minor-damage F1 ↑	Destroyed F1 ↑
xBD	ChangeMamba	85.91	76.68	79.45	59.10	86.44
xBD	FlowMamba	87.66	76.89	80.12	59.30	86.63
xBD	CRD-Net	87.10	78.76	81.26	62.68	86.70
EBD	DAHiTra	85.68	74.99	78.20	80.75	64.50
EBD	FlowMamba	91.20	75.98	80.55	85.58	60.79
EBD	CRD-Net	92.00	77.54	81.88	88.12	61.58

Relative to FlowMamba, CRD-Net improves overall F1 by 1.14 and 1.33 percentage points on xBD and EBD, respectively, and minor-damage F1 by 3.38 and 2.54 points. However, xBD localization F1 is 0.56 points lower, and EBD destroyed-class F1 trails DAHiTra by 2.92 points. The main advantage is improved damage discrimination rather than uniform leadership across all metrics.

Ablation Study

The following reproduces the selected configurations from Table 4. Every configuration marked with “+” adds modules independently to the DINOv3-LoRA baseline; the rows are not cumulative.

Config	Localization F1 ↑	Classification F1 ↑	Overall F1 ↑
DINOv3-Frozen	85.74	74.06	77.56
DINOv3-LoRA baseline	86.74	76.46	79.55
Baseline + SFAFE	87.46	77.98	80.83
Baseline + CRD²	86.22	78.33	80.70
Baseline + OSD only	87.33	77.08	80.16
Baseline + PHCL only	87.25	77.60	80.50
Baseline + SFAFE + CRD²	87.10	78.76	81.26

Adding SFAFE or CRD² individually improves overall F1 over the LoRA baseline by 1.28 or 1.15 percentage points. Relative to SFAFE alone, the full model increases classification F1 by 0.78 points while decreasing localization F1 by 0.36 points. The ablation supports complementary effects while showing that classification and localization do not always improve together.

Key Findings

• The most direct improvement is in xBD minor damage, but the independent gains from OSD and PHCL do not establish that orthogonality guarantees complete semantic separation. The paper provides no such theoretical guarantee.
• LoRA raises overall F1 from 77.56 for the frozen backbone to 79.55, indicating that task adaptation already contributes substantially. The entire improvement cannot be attributed to the new decoder.
• The complexity table reports 86.22 M parameters and 108.33 G FLOPs for CRD-Net, versus 68.52 M and 130.80 G for FlowMamba. Lower operation count with more parameters does not establish lower latency or memory consumption.

Highlights & Insights

• Frequency enhancement is not a claim that more high-frequency content is always better. Selecting fusion rules by band and then selecting locations through spatial gating addresses the coexistence of useful details and background textures.
• Invariant representations are not discarded information. They condition change gating and provide coarse building-versus-background cues for contrastive learning.
• A more permissive contrastive temperature for intact buildings acknowledges that damaged pixels still retain building structure. This is better aligned with fine-grained damage appearance than uniformly strengthening repulsion from every negative sample.

Limitations & Future Work

• The authors identify a remaining weakness against DAHiTra on the EBD destroyed category. Gains for minor damage should not be extrapolated to every disaster morphology.
• As a reading-based assessment, separating buildings from background through invariant-feature energy relies on a texture assumption. Dense vegetation or low-texture buildings may violate it, and the paper does not separately quantify the reliability of this proxy.
• EBD uses a random split without reported disaster-event-held-out evaluation. The results therefore do not establish generalization to unseen disaster areas, and variability across repeated runs is not reported.
• The complexity comparison lacks measured inference time, and some important hyperparameters are not fully disclosed. Reproduction should first verify these settings and severity-specific errors rather than only compare overall F1.

Related Work & Insights

• vs ChangeViT: Both supplement ViT with local detail. CRD-Net additionally separates frequency bands and introduces spatial gating rather than relying only on direct spatial feature injection.
• vs ChangeMamba: CRD-Net reuses its spatiotemporal decoder and adds OSD and PHCL to address change and category information that remains entangled after aggregation. The contribution is not simply replacing a component with Mamba.
• vs FlowMamba: The paper associates its strengths with optical-flow-related state space modeling and class-distance-aware constraints, whereas CRD-Net emphasizes feature injection, subspace decoupling, and semantically adaptive contrast. The xBD localization gap suggests that their strengths do not fully overlap.

Rating

• Novelty: 4/5. Frequency selection, change decoupling, and semantically adaptive contrast form a targeted combination, but rely on existing backbones and decoding architectures.
• Experimental Thoroughness: 3/5. Two datasets and module ablations are included, but event-held-out evaluation, variance, and measured latency are missing.
• Writing Quality: 3/5. The problems map clearly to the designs, although implementation details are incomplete and damaged equation extraction adds verification overhead.
• Value: 4/5. The gains for minor building damage are concrete, making the paper useful for readers studying fine-grained remote sensing change representations.