Research map / Detection and segmentation / Open-set and few-shot detection
Generalizable Disaster Damage Assessment via Change Detection with Vision Foundation Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DAVI, a test-time adaptation method that fuses a source-trained change detection model with SAM text-prompted building confidence differences, refined by augmentation-consistency and coarse image-level masking, to detect building damage in unseen disaster regions without target labels.
Problem
Building damage detection from bi-temporal satellite imagery fails when models trained on limited xBD regions and disaster types are applied to unseen geographies and disaster types, and target labels are unavailable.
Approach
Combines task-specific pseudo labels from a source change detection model with task-agnostic pre-/post-disaster SAM building-confidence differences (Diff-SAM), then refines pixel-level via augmentation-consistency and image-level via source-model presence masking, and adapts the target model with pseudo-label CE plus entropy minimization.
Key results
- On the Sulawesi tsunami target, the method reaches F1 0.4234 with SNUNet-CD versus 0.0369 for the source-only fine-tuned model and 0.0488 for CoTTA.
- On Santa Rosa wildfires, F1 0.7476 (SNUNet-CD) exceeds the source-only baseline (0.7256 for source-star) and SHOT (0.7376) among the adaptation baselines.
- On the Türkiye 2023 earthquake, human-annotated subset F1 of 0.68 versus 0.36 for Source* and 0.42 for SHOT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DA-VFM with SNUNet-CD rather than the strongest listed baseline, SHOT | Wildfires (California, USA) · Wildfires (California, USA) | F1-score: 0.7376 -> 0.7476 | Helped |
| Use DA-VFM with BIT-CD rather than the strongest listed baseline, Source | Hurricanes (Texas, USA) · Hurricanes (Texas, USA) | F1-score: 0.0888 -> 0.5947 | Helped |
| Use DA-VFM with SNUNet-CD rather than the strongest listed baseline, Source | Tsunamis (Sulawesi, Indonesia) · Tsunamis (Sulawesi, Indonesia) | F1-score: 0.1491 -> 0.4234 | Helped |
| Use DA-VFM rather than SHOT | 2023 Türkiye earthquake · 2023 Türkiye earthquake | F1-score: 0.42 -> 0.68 | Helped |
| Add coarse-grained pseudo labels to Source Only (ablation) | Source Only · Wildfires (California, USA) | F1-score: 0.7376 -> 0.7396 | Helped |
| Use Source Only rather than Diff-SAM Only (ablation) | Wildfires (California, USA) · Wildfires (California, USA) | F1-score: 0.4693 -> 0.7376 | Helped |
| Add coarse-grained pseudo labels to Diff-SAM Only (ablation) | Diff-SAM Only · Wildfires (California, USA) | F1-score: 0.4693 -> 0.4647 | Hurt |
| Add Diff-SAM to the source model (ablation) | Source Only · Wildfires (California, USA) | F1-score: 0.7376 -> 0.7415 | Helped |
| Add coarse-grained pseudo labels to Source + Diff-SAM (ablation) | Source + Diff-SAM · Wildfires (California, USA) | F1-score: 0.7415 -> 0.7463 | Helped |
| Add refinement to Source + Diff-SAM without coarse-grained labels (ablation) | Source + Diff-SAM · Wildfires (California, USA) | F1-score: 0.7415 -> 0.7432 | Helped |
| Add refinement to Source + Diff-SAM with coarse-grained labels (ablation) | Source + Diff-SAM with coarse-grained labels · Wildfires (California, USA) | F1-score: 0.7463 -> 0.7476 | Helped |
Surprising findings
- Source-only models transfer poorly to hurricanes and tsunamis, with F1 dropping to near zero for some source-model baselines, while the Diff-SAM component is key for those unseen disaster types.
Limitations
- Relies on a source model trained on a different disaster and region; performance depends on the source-target disparity.
- Diff-SAM threshold is chosen by F1 against source-model maps on the target data, which still requires some target-side choice.
- Wildfire gains over strong source-only baselines are small, and the Türkiye evaluation uses only a 10% human-annotated subset.
Experimental setup
Datasets: xBD 2023 Türkiye earthquake dataset (Maxar)
Benchmarks: xBD
Metrics: precision recall F1-score accuracy
Baselines: CVA IRMAD PCAKmeans ISFA DCVA SCCN TENT SHOT CoTTA
Builds on: SNUNet-CD BIT-CD SAM TENT SHOT CoTTA
Closest papers by meaning
- DisasterM3: A Remote Sensing Vision-Language Dataset for Disaster Damage Assessment and Response (NeurIPS 2025 · similarity 0.87)
- RSCC: A Large-Scale Remote Sensing Change Caption Dataset for Disaster Events (NeurIPS 2025 · similarity 0.83)
- DISCount: Counting in Large Image Collections with Detector-Based Importance Sampling (AAAI 2024 · similarity 0.83)
- Cracks in the Foundation: A Civil Infrastructure Dataset to Challenge Vision Foundation Models (NeurIPS 2026 · similarity 0.82)
- MONITRS: Multimodal Observations of Natural Incidents Through Remote Sensing (NeurIPS 2025 · similarity 0.82)
- The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.