This study presents an integration of Long Short-Term Memory (LSTM) networks with IoT-ready environmental monitoring to forecast hourly Vapor Pressure Deficit (VPD) for precision irrigation in tropical fruit orchards. Historical temperature and humidity data representing the commercial durian orchard in Chanthaburi, Thailand (2021–2025) were used to train and evaluate three forecasting configurations with lookback windows of 24, 168, and 336 h. The 14 day model (L = 336) achieved the highest accuracy (MAE = 0.0901 kPa, RMSE = 0.1283 kPa, \(R^2\) = 0.7957), while the 1 day model (L = 24) provided near-equivalent performance (MAE = 0.0936 kPa, \(R^2\) = 0.7854) with substantially reduced computational cost (approximately 3.4 \(\times \) faster than the 14 day configuration), minimal overfitting (8.5%), and rapid convergence (best validation at epoch 21). The L = 24 configuration was selected for deployment due to its superior trade-off between predictive accuracy and operational efficiency, while requiring the smallest memory footprint among all configurations and thus remaining suitable for CPU-only. The 24 h forecast output was integrated into a smart irrigation decision system that classifies atmospheric conditions into physiological zones including Green (optimal), Yellow (suboptimal), and Orange (stress) to recommend real-time irrigation schedules. Field simulations demonstrated that the system effectively identifies afternoon-to-evening irrigation windows (14:00–17:00) aligned with agronomic best practices, enhancing water-use efficiency and reducing stress exposure. This research establishes a practical, scalable framework for deploying machine learning–driven VPD forecasting within an IoT-integrated analytical framework, supporting sustainable water management under tropical climate variability and directly contributing to the United Nations Sustainable Development Goals, particularly SDG 2 (Zero Hunger), SDG 6 (Clean Water and Sanitation), and SDG 13 (Climate Action).
