AI Summary & Key Takeaways
AI InsightStacked Sentinel-1 + Sentinel-2 fusion improved pseudo-label quality and segmentation accuracy (semi-supervised Overall Accuracy (OA) ≈95.1%), while the PCA＋KCCA hybrid achieved the best unsupervised performance (OA ≈89.4%).
📌 3 Key Takeaways
Abstract
The catastrophic collapse of the Derna Dam created an urgent need for rapid and reliable mapping of flood extent and building damage to support disaster response and recovery. This study presents a multi-sensor, multi-method change detection framework integrating open-access Sentinel-1 Synthetic Aperture Radar (SAR), Sentinel-2 optical imagery, and Very High-Resolution (VHR) Maxar imagery to detect post-event changes and assess building-level damage. Two methodological approaches were evaluated: semi-supervised and unsupervised. The semi-supervised pipeline utilizes a Multilayer Perceptron (MLP)-based post-classification to generate pseudo-labels for training two U-Net variants: single-encoder early fusion and Siamese double-encoder mid-fusion architectures. The unsupervised methods include Principal Component Analysis (PCA) with Change Vector Analysis (CVA), Kernel Canonical Correlation Analysis (KCCA) implemented efficiently with Random Fourier Features (RFF), and a novel PCA with KCCA hybrid that combines dimensionality reduction with nonlinear correlation analysis. Rapid assessment using Sentinel-2 and OpenStreetMap (OSM) data identified changed areas through Change/No-Change and Normalized Difference Moisture Index (NDMI) masks. A detailed analysis employed VHR Maxar imagery and a pretrained Mask Region-based Convolutional Neural Network (Mask R-CNN) within ArcGIS Pro to extract pre- and post-event building footprints and evaluate structural impacts. Validation against manually labeled VHR samples (131 polygons; 3,371 pixels) confirmed the robustness of the framework. Stacked Sentinel-1 + Sentinel-2 fusion improved pseudo-label quality and segmentation accuracy (semi-supervised Overall Accuracy (OA) ≈95.1%), while the PCA＋KCCA hybrid achieved the best unsupervised performance (OA ≈89.4%). For the assessment of building damage, two methods were utilized: Quick Assessment (QA) and Deliberate Assessment (DA). A total of 2,193 flooded buildings were identified using the Deliberate Assessment, while 2,694 flooded buildings were identified through the Quick Assessment. These findings were validated using data from the United Nations Satellite Center (UNOSAT). The study contributes (i) a reproducible, end-to-end workflow for rapid post-disaster mapping, (ii) an efficient RFF-based KCCA implementation, and (iii) a novel PCA＋KCCA hybrid that optimizes the balance between accuracy, robustness, and computational efficiency for operational-scale change detection.