Integration of machine learning models and remote sensing for estimating the Water Deficit Index

Irrigation management based on accurate estimation of crop water stress is essential for improving water use efficiency in arid agricultural systems. This study developed Machine Learning (ML) models for the direct estimation of the Water Deficit Index (WDI) from satellite imagery. Actual and potential evapotranspiration (ETa and ETp) were derived from Landsat 8 data using the METRIC model to generate WDI reference images. Then, the Landsat 8 spectral and thermal bands were used as predictors, and the METRIC-derived WDI was used as the predictand to train and test three ML models: Artificial Neural Network (ANN), Support Vector Regression (SVR), and Random Forest (RF). The study area included pistachio orchards with diverse tree ages, cultivars, and irrigation practices, providing a wide range of agronomic and environmental conditions for model development. Results showed that the RF model achieved relatively better accuracy (DC = 0.850, RMSE = 0.075, KGE = 0.853) compared to the ANN and SVR models, indicating its suitability for estimating crop water stress under heterogeneous conditions. The proposed approach enables the generation of spatially explicit WDI maps in near-real time, providing a cost-effective decision-support tool for precision irrigation and sustainable water resource management in arid regions. • WDI (Water Deficit Index) is estimated using Landsat 8 imagery and ML models. • Landsat 8 imagery was processed via METRIC to derive WDI in pistachio orchards. • ML models were trained and tested using spectral bands and WDI raster data. • The Random Forest model achieved the highest accuracy in WDI estimation. • The selected model enables sustainable irrigation and water management.