E²AlertNet: An explainable, efficient, and lightweight model for emergency alert from aerial imagery

Deep learning-based algorithms have shown significant state-of-the-art accuracy in aerial image classification. Besides, the nature of these algorithms is a black box, which puts the question of why a particular output is produced. Therefore explainability is one kind of solution to improve the transparency of DNN network's decision. In this paper, we have designed a lightweight and explainable convolutional neural network (CNN) architecture for emergency monitoring from aerial imagery. We interpret the outcomes of the proposed architecture with a newly designed explainable algorithm which is the improved version of the model-agnostic methods such as Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME). We have used a dedicated dataset named as Aerial Image Database for Emergency Response (AIDER) for the experiments and explained the decisions of the proposed CNN classifier to ensure reliability. The proposed classifier achieves 96% accuracy with minimal memory requirements on a benchmark set with known ground truth and explains their outcomes with the newly proposed explainable algorithm.