Integrating Ultraviolet and Solar Radiation Monitoring Data with an MMS-GRU Framework for Multi-Scale Dengue Forecasting

Dengue fever continues to expand across tropical and subtropical regions, straining public health surveillance systems, partly because dengue time series exhibit non-stationarity, multi-periodic structure, and climatic drivers that vary geographically. This study proposes MMS-GRU, a hybrid multi-scale framework that decomposes dengue and climate series using the Maximal Overlap Discrete Wavelet Transform (MODWT) to separate seasonal trends from short-term dynamics, applies a two-stage Lasso–SHAP procedure to select informative climate predictors, and feeds each frequency component into a Stacked Bidirectional Gated Recurrent Unit (SBi-GRU) to capture temporal dependencies. The framework was evaluated at three endemic sites (Iquitos, San Juan, and the Philippines) over 13-, 26-, and 52-week horizons, using conventional climate records and, for the Philippines, satellite-derived NASA POWER variables. MODWT produced lower errors than EEMD and CEEMDAN at all three sites, whereas the benefit of feature selection varied with site, horizon, and architecture, and UV index and solar radiation variables ranked among the most influential inputs for the Philippines. These findings indicate that combining multi-scale decomposition with interpretable feature selection is a promising basis for dengue forecasting, although repeated runs and comparison with simple benchmarks under a common protocol are needed before the framework can support vector control planning.