BDANet: Multiscale Convolutional Neural Network with Cross-directional Attention for Building Damage Assessment from Satellite Images

• Automated satellite-image information extraction is motivated by time-critical disaster response, since expert analysis can be time-consuming over large areas.Source: §I · p. 1
• Post-disaster images alone may lack precise contours of intact buildings, while damage assessment must classify damage levels and include undamaged buildings.Unlike change detection, building damage assessment also requires identifying unchanged buildings.Source: §I · p. 1
• About 24.2% of minor-damage samples were misclassified as no damage by the ResNet-50 baseline on xBD, illustrating the visual similarity of difficult classes.Source: §I · p. 2
• The two-stage framework first segments buildings, then assesses damage using a two-branch multi-scale CNN that processes pre- and post-disaster images separately.Building segmentation guides the locations assessed in Stage 2, which reuses network weights from Stage 1.Source: §I · p. 2
• The proposed MFF module aggregates features across scales, while CDA uses channel-wise and spatial-wise correlations to fuse pre- and post-disaster features.CutMix is applied selectively to difficult classes; experiments report better classification performance than applying CutMix to all classes.Source: §I · p. 2
• BDANet achieves state-of-the-art performance on the large-scale xBD building-damage-assessment benchmark.Source: §I · p. 2

A. Building Damage Assessment

• Post-disaster-only assessment methods include SVM and object-based machine-learning approaches,, but may lack precise contours of intact buildings; paired imagery has also been used to extract textural and structural features.Source: A. Building Damage Assessment · p. 3
• Some prior work simplified assessment to binary damage/no-damage labels, while later approaches considered multiple levels, including minor and major damage,.Source: A. Building Damage Assessment · p. 3

B. U-Net

• U-Net combines downsampling and upsampling to aggregate features at multiple scales, with skip connections integrating low- and high-level features.Source: B. U-Net · p. 3
• Although U-Net variants have been used for remote-sensing segmentation and related tasks,,, the paper says the architecture cannot be directly used for building damage assessment.Source: B. U-Net · p. 3

C. Attention-based Models

• Attention methods have been used in remote-sensing classification and segmentation,, but the cited non-local attention approach to paired-image damage assessment has high computational cost on high-resolution features [13],.Source: C. Attention-based Models · p. 3

D. Data Augmentation Strategies

• CutMix cuts and pastes image patches, unlike Mixup's linear combination of image samples; BDANet uses CutMix to address difficult classes.Source: D. Data Augmentation Strategies · p. 4

III. PROPOSED METHOD

• BDANet uses pre-disaster imagery in Stage 1 to produce building masks, then feeds pre- and post-disaster images into separate Stage 2 branches with MFF and CDA modules.Source: §III · p. 4

A. Two-stage Framework

• A two-stage framework separates building localization from damage assessment, addressing the need to classify damage levels and detect undamaged buildings.The paper notes that change-detection networks focus on image differences and may ignore unchanged objects such as undamaged buildings.Source: A. Two-stage Framework · p. 4
• Stage 1 uses pre-disaster imagery in a U-Net to produce a binary building-distribution map for guiding the assessment stage.The paper motivates this stage by noting that damaged post-disaster buildings make simultaneous segmentation and damage assessment difficult.Source: A. Two-stage Framework · pp. 4–5
• Stage 2 feeds pre- and post-disaster images into two branches and initializes their shared convolutional kernels with weights from Stage 1.The resulting segmentation predictions identify building locations, while the damage output is guided by those locations.Source: A. Two-stage Framework · p. 5

B. Multi-scale Feature Fusion Module

• Three image resolutions—1×, 0.5×, and 0.25×—feed separate streams whose features are embedded at encoder layers for multi-scale representation.The 1× and 0.5× stream features are concatenated and dimension-reduced; the 0.25× stream is concatenated with a later ResNet block.Source: B. Multi-scale Feature Fusion Module · p. 5

C. Cross-directional Attention Module

• CDA cross-aggregates channel and spatial information from pre- and post-disaster features to recalibrate both branches.Channel feature fusion feeds into spatial feature fusion; the module uses convolution and matrix operations.Source: C. Cross-directional Attention Module · pp. 5–6

D. Data Augmentation with CutMix

• The authors use CutMix to increase difficult-class sample sizes, aiming to improve feature representations and encourage attention to those classes.The motivating challenge is visual similarity among damage classes.Source: D. Data Augmentation with CutMix · p. 6
• CutMix takes patches from difficult-class samples—minor and major damage in xBD—and pastes them into randomly selected sample images.The augmented image pairs and labels are generated using a binary mask.Source: D. Data Augmentation with CutMix · p. 6

A. Dataset and Evaluation

• xBD covers 19 disasters and more than 800,000 building annotations, with paired 1024 × 1024 RGB images at 0.8 meter per pixel.Damage labels comprise no damage, minor damage, major damage, and destroyed; the distribution is highly skewed toward no damage.Source: A. Dataset and Evaluation · pp. 6–7
• Building segmentation is evaluated with F1b, while damage assessment uses F1d, the harmonic mean of class-wise damage-classification F1 scores.F1b uses true-positive, false-positive, and false-negative pixels; F1s is a weighted average of F1b and F1d.Source: A. Dataset and Evaluation · p. 7

B. Implementation Details

• Training uses 512 × 512 crops, ResNet-50 for all compared networks, cross-entropy loss, and Adam; the segmentation and damage stages run for 120 and 25 epochs.Testing predictions use the original 1024 × 1024 image size.Source: B. Implementation Details · p. 7

C. Results Analysis

• The vanilla framework exceeds RescueNet and Weber’s method on F1s, F1b, and F1d, with over 2% improvement in F1b and minor-damage F1 rising from 0.493 to 0.616.The authors attribute the segmentation advantage in part to the decoder and skip connections.Source: C. Results Analysis · p. 7
• 0.782 vs 0.757 in F1d: the full framework improves on the vanilla network, with gains across damage levels and higher F1s than RescueNet and Weber’s method.The comparison considers building segmentation and damage assessment together; Table V reports xBD results.Source: Table V · p. 8, C. Results Analysis · p. 8
• Stage 1 segmentation results are reported to match ground-truth maps well and guide building locations for Stage 2 damage assessment.Figure 5 orders pre-disaster images, segmentation ground truth, and Stage 1 results by row.Source: C. Results Analysis · p. 8, Figure 5 · p. 8
• In visual comparisons, BDANet has fewer damage-assessment errors overall; minor and major damage remain difficult classes for the compared methods.Figure 6 places ground truth beside FCN, SegNet, DeepLab, and BDANet predictions.Source: C. Results Analysis · p. 8

D. Ablation Study

• 1.3% in F1s and 2.7% in minor-damage accuracy (0.578 vs 0.605): CutMix improves overall performance, especially when focused on minor and major damage.Applying CutMix to all classes yields barely any improvement, whereas class-specific augmentation produces a significant gain.Source: D. Ablation Study · pp. 8, 11, Table VII · p. 9
• 0.5% in F1s: using all three MFF streams improves overall damage-assessment performance; the 0.5× and 0.25× streams individually add about 0.4% and 0.2%.The 1× stream captures local detail, while lower-resolution streams capture more global information.Source: D. Ablation Study · p. 8, Table VIII · p. 9, Figure 7 · p. 10
• About 0.6% improvement over the vanilla network: using all three CDA modules performs better than placing attention only at dconv1 or dconv3.The authors report a better gain at larger-scale dconv3 than at small-scale dconv1; Table IX concerns these positions.Source: Table IX · p. 9, D. Ablation Study · p. 10
• BDANet’s CDA outperforms the evaluated SE attention variants; SE channel attention alone adds only 0.1% in F1s.The comparison evaluates channel attention, spatial attention, and their combination.Source: D. Ablation Study · p. 10, Table X · p. 10

E. Computational Cost and Training Efficiency

• 107.6 vs 92.9 GFLOPs: adding MFF and CDA raises computational cost, while the full framework adds less than 2M parameters and about 60G FLOPs over vanilla.CutMix is applied during preprocessing and adds no network computational resources; added runtime is 0.016 seconds per image.Source: E. Computational Cost and Training… · p. 11
• At least 5 epochs are saved when Stage 1 weights initialize Stage 2; validation accuracy starts higher and loss lower, with faster convergence.The comparison is reported for Stage 2 fine-tuning against training without Stage 1 initialization.Source: E. Computational Cost and Training… · p. 11

F. Evaluation on Building Change Detection Dataset

• LEVIR-CD evaluation used 509 image pairs for training and 128 for testing, with Stage 2 alone trained for change detection.The dataset contains 637 high-resolution images from Texas cities, captured from 2002 to 2018.Source: F. Evaluation on Building Change… · p. 11
• BDANet achieved the best precision, recall, and F1 scores among the compared change-detection methods on LEVIR-CD.The authors also report better visual performance than other methods.Source: F. Evaluation on Building Change… · pp. 11–12

V. CONCLUSION

• The authors report state-of-the-art building damage assessment performance on xBD and improvements from CutMix on difficult damage levels.They describe MFF as learning features at multiple scales and CDA as exploring correlations between pre- and post-disaster features.Source: §V · p. 12