Classification of Tropical Disease-carrying Mosquitoes Using Deep Learning and SHAP

Abstract

In this paper, we present a novel technique for identifying mosquitoes that carry tropical diseases using Deep Learning and SHAP for model interpretability. We propose an end-to-end deep (E2E) Convolutional Neural Network (CNN) architecture that leverages mosquito wingbeat sounds to extract relevant features. To achieve high-performance audio processing, we integrate Kapre, an audio processing library optimized for GPU execution. Our approach also incorporates SHAP to provide a transparent explanation of the model’s predictions, enabling us to identify and characterize the time-frequency patterns that the model emphasizes. Ultimately, our research aims to support disease control initiatives by providing an automated means of identifying disease-carrying mosquito species, which has the potential to improve public health in tropical regions.