From Aerial Imagery to Label-Conditioned Disaster Reports

Abstract

This study proposes a hybrid approach for disaster damage detection and automated explanatory reporting using the LADI-v2 dataset, which comprises aerial imagery. The system is structured upon a two-stage architecture: in the first stage, a ResNet-18 model trained to identify 12 distinct damage categories exhibited high classification performance with an F1 score of 81.88%. In the second stage, Qwen2-VL and BLIP-2 Vision-Language Models (VLMs) were optimized through LoRA-based fine-tuning to translate detected findings into natural language. Performance analyses revealed that the BLIP-2 model outperformed in fidelity to reference texts and key concept extraction, achieving scores of 53.29 in sacreBLEU and 71.03 in ROUGE-1. Conversely, the Qwen2-VL model excelled in semantic flexibility and grammatical richness with a METEOR score of 0.81. Findings indicate that BLIP-2 excels in high-precision reporting, whereas Qwen2-VL provides superior contextual depth. This framework demonstrates significant potential as a rapid, interpretable, and collaborative decision-support tool for disaster management within experimental settings. © 2026 IEEE.