Evaluating Machine Learning Models for GLDAS Air Temperature Prediction on North Western Coast of Egypt in a Changing Climate

Accurate air temperature forecasting is essential for climate risk management, resilient agricultural planning, and sustainable development, yet it remains challenging in regions with sparse meteorological station networks. In Egypt, the limited spatial coverage of ground-based observations constrains the reliable characterization of local climate variability and predictive capacity. This research investigates the potential of integrating NASA's Global Land Data Assimilation System (GLDAS) remote sensing product with machine-learning (ML) techniques to overcome data scarcity and improve daily air temperature forecasting along the Egyptian North-Western Coast. GLDAS v2.1, 0.25° resolution, Near-surface air temperature was evaluated against in situ observations for the period 2000–2020 at four locations: Alexandria, Marsa Matrouh, Sallum, and Siwa Oasis, confirming its high fidelity (R² ≥ 0.886) and reliability of reanalysis data for local-scale forecasting. A correlation-based feature selection strategy was applied to identify the most informative lagged temperature predictors, followed by the implementation of Linear Regression (LR), Support Vector Regression (SVR), and Long Short-Term Memory (LSTM) models. Results demonstrate that the SVR model consistently achieved superior performance across all sites, delivering the lowest Mean Squared Error (0.659–1.610°C) and highest Nash-Sutcliffe Efficiency (0.9428–0.9685). The research establishes a robust, scalable framework that transforms freely available global satellite data into precise local forecasts. This workflow provides a critical tool for enhancing climate adaptation strategies in agriculture, water management, and early warning systems in arid, data-deficient regions, overcoming the delay in the availability of climate models online.