Development and validation of a Geospatial eXplainable Artificial Intelligence (GeoXAI) framework for mapping mosquito density in metropolitans

Abstract

Accurately predicting mosquito density in metropolitans is challenging due to the nonlinear, spatially non-stationary relationships between explanatory variables and in situ monitoring data. These challenges are further exacerbated by the limited interpretability of traditional statistical and artificial intelligence models. To overcome existing limitations, we integrated the Shapley additive explanation (SHAP) method and geographically neural network weighted regression kriging (GNNWRK) model to develop the GNNWRK-SHAP framework for predicting mosquito density in Shenzhen and Central Guangzhou (C-Guangzhou), China. GNNWR is an enhanced spatial model derived from GWR, which combines ordinary least squares and spatial weighted neural network to capture spatial non-stationarity and complex nonlinearity. Our results demonstrated that the GNNWRK-SHAP framework achieved the lowest mean absolute error (1.72 in Shenzhen, 3.22 in C-Guangzhou), root mean square error (2.05 & 3.76), and mean absolute percentage error (42.48 % & 62.95 %) compared to alternative models. By integrating the SHAP, the framework quantified variable contributions, identifying precipitation and temperature as dominant explanatory variables shaping mosquito distributions. Moreover, spatial SHAP values highlighted the spatially varying impacts of these, offering site-specific decision support for mosquito density control. This study provides a novel framework for high-precision mapping while maintaining spatial interpretability, providing new insights for mosquito monitoring and management.