Aerial Image Classification in Post Flood Scenarios Using Robust Deep Learning and Explainable Artificial Intelligence

Flood is a natural disaster that very often causes disastrous effects on the human populace and infrastructures. In case of floods, aerial images enable one to analyze the extent of the damage and affected areas. This paper describes an approach for aerial image classification in the post-flood scenario by using efficient deep learning architectures such as MobileNet and DenseNet for precise classification of regions within aerial images. The proposed models are intended to classify within six distinct classes - building, flooded, forest, mountains, sea, and street. For every model, Explainable Artificial Intelligence (XAI) techniques are inputted to provide such interpretability and transparency. The methods used are Class Activation Maps generated through Grad-CAM. The results show the deep learning models' capability to differentiate between flooded regions and other prominent features for post-flood assessment and recovery planning. In addition, the trustworthiness of model predictions, which is most important for real-world disaster management and response applications, is enhanced via interpretability through Grad-CAM. The performance of this model has been assessed with various metrics and has been found to contribute to the already growing pool of studies being directed towards improving post-disaster recovery efforts through AI-based solutions. Keywords: Aerial Image Classification, Post-Flood Analysis, MobileNet, DenseNet, Explainable Artificial Intelligence (XAI), Grad-CAM, Deep Learning, Flooded Areas, Damage Assessment, Disaster Management, Image Classification, Computer Vision, Remote Sensing, Disaster Recovery, Class Activation Mapping.