Identifying hurricane damage using explainable compact transformer with convolutional embedding

Abstract

Hurricanes are tropical storms that cause substantial loss of human life and property. Damage assessment after a storm is vital for emergency personnel to allocate resources suitably. Typically, a ground survey is conducted to evaluate the extent of the damage by estimating the number of flooded or damaged buildings, which is generally a time-consuming and tedious procedure. Due to the critical nature of the situation at hand, an effective and efficient strategy is necessary for its solution. In this study, we classify satellite images of damaged and normal areas by modifying an explainable Compact Convolutional Transformer (CCT) model to achieve high performance with comparatively less computational requirements. CCT is a Vision Transformer Variant with incorporated convolutions, enabling enhanced inductive bias and eliminating the need for positional embedding (as it is replaced with convolutions). We incorporate LIME, an Explainable Artificial Intelligence (XAI) framework, to offer interpretable and explainable predictions. Our proposed model achieves 98.79% accuracy, which outperforms other popular and existing models.