Abstract

Soil temperature is a key determinant of soil health and agricultural productivity, especially in arid regions vulnerable to climate change. This study investigates the use of advanced machine learning models to predict soil temperature variations in Bustan, Uzbekistan, a region facing significant climatic stress. Using 16 years of meteorological data, including atmospheric temperature, humidity, and wind speed, eight machine learning models were evaluated for their ability to predict surface and subsurface (10 cm depth) soil temperatures. Among the models tested, the bi-directional long short-term memory (Bi-LSTM) algorithm demonstrated superior predictive accuracy with R² values exceeding 0.94 for subsurface temperatures. The two-step modeling approach utilized Bi-LSTM outputs from surface temperature predictions to inform subsurface estimates, reflecting a novel methodology for climate-sensitive agriculture. The results provide a practical framework for improving irrigation planning, crop yield forecasting, and sustainable land management in data-scarce arid environments. By demonstrating high accuracy and real-world applicability, this AI-driven model offers a scalable solution for enhancing agricultural resilience in Uzbekistan and similar contexts.