Multiscale Meteorological Drought Spatial Reconstruction in North-Central Urban Core of Mexico City: An Explainable Deep Learning Approach

Mexico City experiences severe water stress driven by aquifer overexploitation and recurrent droughts. Effective water management requires operational spatial monitoring systems capable of spatially reconstructing meteorological anomalies across multiple temporal scales. In this work we developed an explainable deep learning framework using Long Short-Term Memory (LSTM) networks to spatially reconstruct three drought indices—the Standardized Precipitation Index (SPI), Standardized Precipitation Evapotranspiration Index (SPEI), and Reconnaissance Drought Index (RDI)—across five accumulation scales (3, 6, 12, 18, and 24 months). To strictly isolate genuine meteorological deviations, we adopted a hybrid statistical approach: SPI was computed following the standard WMO methodology using Gamma distribution fitting, while SPEI and RDI were computed using empirical monthly standardized anomalies to ensure robustness in non-stationary urban climates without forcing distributional assumptions. Model generalization was evaluated using a leave-one-microsite-out validation strategy, training on two stations and testing on a spatially isolated third station, with inter-station distances ranging from 1.8 to 6.7 km, sufficient to capture urban microclimatic heterogeneity while remaining within the same regional climate zone. We quantified feature importance using SHapley Additive exPlanations (SHAP) to provide mathematical transparency. The LSTM achieved predictive performance at long-term scales by effectively capturing deep sequential memory, while short-term reconstructions reflected the inherent noise of urban convective precipitation. The framework demonstrates reliable intra-urban spatial generalization capacity, supporting the development of diagnostic tools for metropolitan water stress assessment.