A multimodel hybrid framework for long-term maximum temperature prediction and environmental management in the Chapada Diamantina Plateau of Brazil

Abstract Accurate maximum temperature (Tmax) prediction is critical for water security and agriculture in the Chapada Diamantina, Brazil. This study evaluates a multimodel SARIMAX (mmSARIMAX) and a hybrid mmSARIMAX-LSTM approach to forecast monthly Tmax across sixteen sites using NASA POWER reanalysis data. Contrary to expectations, the deep learning hybridization yielded only marginal performance gains compared to the standalone mmSARIMAX. Using novel ad-hoc metrics, namely Scarcity-Adjusted Lag Importance (SALI) and Scarcity-Adjusted Variable Importance (SAVI), we identified three distinct physical control axes governing regional predictability: Radiative (ensuring long-term stability), Hydro-Advective (modulating short-term anomalies), and Spatial (driving local discrepancies). Results indicate that Tmax predictability is a mosaic of regimes, where longwave radiation and wind dynamics act as primary drivers depending on the sub-region. Our study shows that parsimonious statistical ensembles, when incorporating appropriate temporal memory, provide a robust, low-cost framework for regional climate intelligence, supporting evidence-based decision-making in data-scarce environments.