Aerial Imagery of Natural Disaster-Affected Areas (AINDAA) Dataset for Semantic Segmentation and Natural Disaster Assessment

This study presents a new dataset of aerial imagery of natural disaster-affected areas called AINDAA for visual display in post-disaster scenarios, particularly flood disasters, and analyzes a sophisticated deep learning model for semantic segmentation. AINDAA is a high-resolution aerial imagery with fully annotated ground truth data, which is very suitable for accurate flood mapping and natural disaster assessment for developing countries, such as Indonesia, in assisting disaster emergency response and post-disaster management. AINDAA has complex background images and brings more clarity to disaster areas, important objects affected by natural disasters, objects with irregular shapes and sizes, and small objects. This study comprehensively proposes several scenarios to improve the model's performance. The results showed that the PSPNet(50) (bp) model achieved the most optimal performance for detecting and identifying objects in aerial imagery of natural disaster-affected areas, which used the best parameters, pre-trained on combined general datasets for transfer learning, and re-trained on AINDAA, which achieved the highest accuracy value of 96.98%. The experiments show that AINDAA has challenges in detecting and identifying objects with small shapes in post-disaster aerial imagery.