A Rapid Self-Supervised Deep-Learning-Based Method for Post-Earthquake Damage Detection Using UAV Data (Case Study: Sarpol-e Zahab, Iran)

Immediately after an earthquake, rapid disaster management is the main challenge for relevant organizations. While satellite images have been used in past two decades building-damage mapping, they rarely utilized timely damage monitoring required rescue operations. Unmanned aerial vehicles (UAVs) recently become very popular due to their agile deployment sites, super-high spatial resolution, and relatively low operating cost. This paper proposes a novel deep-learning-based method post-earthquake building detection. The detects damages four levels consists of three steps. First, different feature types—non-deep, deep, fusion—are investigated determine optimal extraction method. A “one-epoch convolutional autoencoder (OECAE)” extract deep features from non-deep features. Then, rule-based procedure designed automatic selection proper training samples by classification algorithms next step. Finally, seven famous machine learning (ML) algorithms—including support vector (SVM), random forest (RF), gradient boosting (GB), extreme (XGB), decision trees (DT), k-nearest neighbors (KNN), adaBoost (AB)—and basic algorithm (i.e., multi-layer perceptron (MLP)) are implemented obtain maps. results indicated that auto-training feasible superior manual ones, with improved overall accuracy (OA) kappa coefficient (KC) over 22% 33%, respectively; SVM (OA = 82% KC 74.01%) was most accurate AI model slight advantage MLP 73.98%). Additionally, it found fusion using OECAE could significantly enhance damage-mapping efficiency compared those either (by average improvement 6.75% 9.78% OA KC, respectively) or (improving 7.19% 10.18% on average) alone.