PaperlayerView exact revision on arXiv

Source-linked AI summary

Robust Building Damage Detection in Cross-Disaster Settings Using Domain Adaptation

Asmae Mouradi, Shruti Kshirsagar

arXiv:2603.14694v2cs.CVcs.AIcs.LG

TL;DR

Building damage models can lose accuracy and consistency when deployed in unseen regions because of domain shift. This paper adapts an xBD-pretrained two-stage ensemble to Ida-BD using supervised domain adaptation; unsharp-enhanced RGB attains a Macro-F1 of 0.5552.

  • Problem

    Building-damage models trained on multi-disaster benchmarks face domain shift in unseen regions and disaster scenarios, which degrades classification accuracy and consistency.

    Source: §I · p. 1

  • Method

    The paper transfers an xView2-based two-stage ensemble from xBD to Ida-BD using supervised domain adaptation and evaluates augmentation components.

    Source: §I · p. 2, A. Two-stage classification approach · p. 2

  • Results

    Unsharp-enhanced RGB input with supervised domain adaptation performs best, with a Macro-F1 of 0.5552.

    Source: §V · p. 6

  • Takeaways & Limitations

    The findings support domain-adaptive techniques for integrating automated damage assessment into disaster-response processes that depend on accurate outputs for human-machine teaming.

    Source: §V · p. 6

Abstract

from arXiv · show

Rapid structural damage assessment from remote sensing imagery is essential for timely disaster response. Within human-machine systems (HMS) for disaster management, automated damage detection provides decision-makers with actionable situational awareness. However, models trained on multi-disaster benchmarks often underperform in unseen geographic regions due to domain shift - a distributional mismatch between training and deployment data that undermines human trust in automated assessments. We explore a two-stage ensemble approach using supervised domain adaptation (SDA) for building damage classification across four severity classes. The pipeline adapts the xView2 first-place method to the Ida-BD dataset using SDA and systematically investigates the effect of individual augmentation components on classification performance. Comprehensive ablation experiments on the unseen Ida-BD test split demonstrate that SDA is indispensable: removing it causes damage detection to fail entirely. Our pipeline achieves the most robust performance using SDA with unsharp-enhanced RGB input, attaining a Macro-F1 of 0.5552. These results underscore the critical role of domain adaptation in building trustworthy automated damage assessment modules for HMS-integrated disaster response.

I. INTRODUCTION

Building damage models face domain shift in unseen regions and disaster scenarios, while the robustness of augmentation components remains underexplored. This paper proposes two-stage supervised adaptation from xBD to Ida-BD and evaluates augmentation and adaptation choices for disaster-response systems.

I. INTRODUCTION

  • Automated damage assessment is framed as decision support that can help analysts focus on severely affected regions and reduce manual screening effort.Source: §I · p. 1
  • Unseen regions and disaster scenarios create domain shift, degrading building-damage classification accuracy and consistency,,.The passage attributes distributional mismatch to differences in sensors, geography, atmospheric conditions, and disaster types.Source: §I · p. 1
  • Prior work found supervised fine-tuning on target-domain data improves generalization over pretrained-only and unsupervised adaptation baselines,.Source: §I · p. 1
  • The robustness of these algorithms and the effects of individual augmentation components on performance remain largely unexplored.Source: §I · pp. 1–2
  • The paper proposes transferring an xView2 first-place approach from xBD to Ida-BD using two-stage ensemble supervised domain adaptation [21],.Source: §I · p. 2
  • Its contributions include a segmentation-and-classification pipeline with fusion augmentation that surpasses results reported in, plus ablations intended to guide reliable HMS damage-assessment modules.Source: §I · p. 2

II. PROPOSED METHOD

The proposed method combines building localization and four-class damage classification in a two-stage pipeline. It fine-tunes xBD-pretrained models on annotated Ida-BD data to adapt to the target domain.

A. Two-stage classification approach

  • The pipeline first localizes buildings, then classifies damage from pre- and post-disaster imagery into No-Damage, Minor, Major, or Destroyed.Stage 1 generates a binary building mask that gates Stage 2 classification to building pixels.Source: A. Two-stage classification approach · p. 2
  • Twelve xBD-pretrained localization encoders are fine-tuned on augmented imagery, while damage-classification models are fine-tuned from xBD-pretrained weights onto Ida-BD.Source: A. Two-stage classification approach · p. 2
  • The method presents sequential building-mask gating as a way to decrease background false positives and improve damage estimates for human operators.Source: A. Two-stage classification approach · p. 2
  • Transfer learning initializes models with xBD weights before fine-tuning on annotated Ida-BD data; the comparison baseline applies xBD checkpoints directly without adaptation.Source: A. Two-stage classification approach · p. 2

B. Augmentation Methods

The augmentation methods add complementary visual cues to address appearance variation in post-disaster imagery. The section describes edge detection and CLAHE as ways to emphasize boundaries and subtle damage patterns.

B. Augmentation Methods

  • Post-disaster appearance variation motivates image-level augmentations used singly or in combination to add complementary cues for damage assessment.The passage identifies atmospheric conditions, sensor angles, and debris-induced texture changes as sources of variability.Source: B. Augmentation Methods · p. 2

1) Edge Detection:

  • Edge detection produces a single-channel map that emphasizes structural boundaries, including roof, wall, and debris-field edges.These boundaries may indicate damage and help the model detect transitions between intact and damaged regions.Source: 1) Edge Detection · p. 2

2) Contrast Enhancement (CLAHE):

  • CLAHE increases local contrast while limiting amplification on image tiles to avoid enhancing noise, highlighting subtle damage patterns such as discolored roofs and partially covered debris.The method is described as useful under variable lighting and for distinguishing No-Damage buildings from those with minimal surface damage.Source: 2) Contrast Enhancement (CLAHE) · p. 2

3) Unsharp Masking:

Unsharp masking amplifies fine structural details associated with severe damage. Fusion augmentation combines the original image with edge, contrast, and sharpening components, but their combination does not always improve performance.

3) Unsharp Masking:

  • Unsharp masking subtracts a Gaussian-blurred image and reinjects an amplified residual, accentuating cracks, roof fragmentation, debris, and deformed structures.Among the tested components, it is described as most effective for recognizing heavily damaged structures by amplifying high-frequency textural disruptions.Source: 3) Unsharp Masking · p. 2

4) Fusion Augmentation:

  • Fusion augmentation forms a weighted combination of the original image, edge map, CLAHE output, and unsharp-masked image, using fixed weights during training and evaluation.The components contribute boundary cues, locally enhanced contrast, and sharpened detail.Source: 4) Fusion Augmentation · p. 3
  • Combining all enhancement components does not always yield the best performance, as conflicting cues can degrade classification of minority damage classes.This motivates evaluation of individual components and pairwise combinations alongside the full fusion method, with each configuration tested using supervised domain adaptation.Source: 4) Fusion Augmentation · p. 3

III. EXPERIMENTAL SETUP

The experimental setup uses xBD as the source domain for initializing model weights and frames building localization and damage classification as the benchmark tasks.

III. EXPERIMENTAL SETUP

  • The experimental setup describes the datasets, benchmark system, and performance metric used in the experiments.Source: §III · p. 3

1) xBD Dataset (Source Domain):

  • xBD provides very-high-resolution pre- and post-disaster imagery across natural disasters, with building annotations spanning geographic regions.The dataset includes imagery organized into 1024 × 1024-pixel tiles.Source: 1) xBD Dataset (Source Domain) · p. 3
  • The xBD tasks are building localization and four-class damage classification, and the framework uses xBD only to initialize model weights.Its train, test, and validation sets cover identical disaster occurrences, while tier3 includes supplementary events.Source: 1) xBD Dataset (Source Domain) · p. 3

2) Ida-BD Dataset (Target Domain):

Ida-BD supplies the geographically and sensor-specific target data, while the xView2 first-place solution provides pretrained weights for the benchmark pipeline.

2) Ida-BD Dataset (Target Domain):

  • Ida-BD contains 87 pre- and post-disaster WorldView-2 image pairs from Louisiana before and after Hurricane Ida in 2021.The study uses an 80% training, 10% validation, and 10% held-out test split.Source: 2) Ida-BD Dataset (Target Domain) · p. 3
  • Ida-BD is the target domain for adaptation to differences in sensor features, regional building types, vegetation, and hurricane-specific damage signatures.The dataset’s discrete geography, disaster type, and sensor make it the study’s cross-disaster generalization target.Source: 2) Ida-BD Dataset (Target Domain) · p. 3

B. Benchmark System: xView2 Challenge

  • Table I reports F1 scores for the top-three xView2 Challenge solutions on the xBD test set.Source: Table I · p. 4
  • The xView2 first-place solution initializes the pipeline with an ensemble of four U-Net-based architectures and 12 models in total.The models use DPN92, ResNet34, SE-ResNeXt50, and SENet154 backbones, each trained with three random seeds.Source: B. Benchmark System: xView2 Challenge · p. 4

C. Performance Measures

The paper evaluates class-level F1 and uses Macro-F1 to summarize performance across the four damage classes; the results section examines cross-data building damage detection.

C. Performance Measures

  • For each class, F1 is the harmonic mean of precision and recall.Source: C. Performance Measures · p. 4
  • Macro-F1 averages F1 across the four damage classes into one summary metric for HMS-integrated response.The paper describes it as indicating whether predictions are balanced across the damage spectrum.Source: C. Performance Measures · p. 4

IV. RESULTS AND DISCUSSION

  • The results section presents findings for building damage detection in a cross-data setting.Source: §IV · p. 4

A. Effect of supervised domain adaptation

In the cross-data experiment, omitting supervised domain adaptation preserves localization performance but yields very low damage-class scores; the accompanying qualitative comparison describes cleaner localization with adaptation.

A. Effect of supervised domain adaptation

B. Ablation study: Effect of Individual Augmentation Components

With SDA, unsharp-enhanced RGB produced the strongest reported overall damage result, while other augmentation combinations traded performance across damage classes. The qualitative mask distinguishes all four severity levels.

B. Ablation study: Effect of Individual Augmentation Components

  • Unsharp masking + DA yields the highest Destroyed F1 (0.5156) and total damage Macro-F1 (0.5552), slightly beating the RGB-only baseline (0.5516).Table III reports the individual and combined augmentation ablation on the Ida-BD test set.Source: B. Ablation study: Effect of… · p. 4, Table III · p. 5
  • The RGB + Unsharp + SDA prediction mask distinguishes no damage, minor, major, and destroyed buildings with geographically consistent predictions.The mask uses green, yellow, orange, and red for those classes, respectively.Source: B. Ablation study: Effect of… · p. 5, Figure 4 · p. 5
  • Contrast + DA raises No-Damage F1 (0.7334) but lowers Destroyed F1 (0.0412); Edge detection + DA reaches Destroyed F1 (0.4362) and Macro-F1 (0.5359).The contrast configuration's weak Destroyed score is linked to masking small, scattered damage indications; edge cues help but do not match unsharp masking for severe damage.Source: B. Ablation study: Effect of… · p. 5
  • Contrast + Edges + DA leads localization F1 (0.8877) and No-Damage F1 (0.7338) but has Macro-F1 (0.4690); Unsharp + Contrast + DA leads Major-Damage F1 (0.5602).Their Destroyed F1 scores are 0.1957 and 0.3653, respectively.Source: B. Ablation study: Effect of… · p. 5
  • Full fusion augmentation + DA is ineffective, with conflicting feature cues appearing to confuse the classifier on minority classes.The passage attributes this to damage-specific signals being drowned out when all augmentation channels are stacked.Source: B. Ablation study: Effect of… · p. 5

C. Comparison with benchmark systems

On Ida-BD, the two-stage ensemble with supervised DA improves F1 scores across damage categories compared with benchmark systems. The conclusion identifies unsharp-enhanced RGB with supervised DA as the best configuration and points to lower-annotation adaptation as future work.

C. Comparison with benchmark systems

  • The two-stage ensemble with supervised DA improves F1 across damage categories, with Major-Damage rising from 0.192 to 0.556 (190%) and Destroyed from 0.117 to 0.516 (341%).The comparison uses benchmark systems on Ida-BD; the pipeline averages 12 models and uses optimized probability thresholding, building-pixel-restricted loss, and destroyed-aware crop sampling.Source: C. Comparison with benchmark systems · pp. 5–6

V. CONCLUSION

  • Unsharp-enhanced RGB with supervised DA performs best, with Macro-F1 of 0.5552; the study identifies domain adaptation as the most crucial factor in cross-data damage detection.The conclusion supports domain-adaptive techniques for automated damage assessment in practical disaster response.Source: §V · p. 6
  • Future research will explore semi-supervised and unsupervised domain adaptation to reduce annotation requirements and support dependable HMS-integrated outputs.Source: §V · p. 6
Loading 2603.14694v2…