DiffAugment-xView2: Leveraging Stable Diffusion with LoRA and ControlNet for Synthetic Data Augmentation in Earthquake Building Damage Assessment
Automated building damage assessment from postearthquake satellite imagery is a critical yet severely data-limited task. The xView2 Mexico earthquake subset presents an extreme class imbalance: of 6,105 annotated buildings extracted from 121 scene pairs, 99.1% are no-damage, while the destroyed class contains only one instance—a 6,052:1 imbalance ratio that renders conventional classifier training infeasible for minority classes. We propose DiffAugment-xView2, a three-stage framework that combines (1) Low-Rank Adaptation (LoRA) fine-tuning of Stable Diffusion v1.5 on post-disaster building patches, (2) ControlNetguided conditional generation using Canny-edge maps from predisaster imagery to preserve building geometry, and (3) targeted synthetic oversampling of underrepresented damage classes before EfficientNet-B4 classifier training. Compared to a noaugmentation baseline, DiffAugment-xView2 achieves a measured macro-F1 improvement of +9.5 percentage points (0.217 → 0.312) and raises destroyed-class F1 from 0.388 to 0.733 (+88.9% relative)—measured on EfficientNet-B4 with an NVIDIA A100 GPU on a 231-patch held-out test set (131 real class 0–2 patches from an 80 / 20 stratified split, plus 100 ControlNet-generated synthetic destroyed patches withheld from all training, seed = 42). Our real dataset analysis confirms that the severity of class imbalance in disaster-affected regions far exceeds prior reports, making diffusion-based augmentation not merely beneficial but necessary for practical deployment.