Early Forecasting of Crop Irrigation for Sustainable Water Use with Satellite Data and Machine Learning

Early forecasting of crop irrigation demand improves water efficiency and supports food security. Accurate forecasts allow farmers and policymakers to schedule irrigation based on crop type and climate, ensuring timely water application throughout the growing season. Traditional biophysical models rely on fixed delivery schedules and broad hydrological trends, often lacking precision for specific crop stages. While machine learning (ML) has been widely used for yield and biomass prediction, its application in forecasting crop water requirements remains limited. This study bridges that gap by integrating remote sensing with ML to align water supply with crop demand under changing climate conditions, promoting sustainable irrigation. The study focuses on Pakistan’s Indus Basin, specifically Chaj Doab, located between the Jhelum and Chenab rivers. The region features flat terrain with coarse-textured alluvial soils and high evapotranspiration with wide variety of crops. The proposed methodology for estimating irrigation demand uses actual evapotranspiration (ET_act) derived from satellite-based biophysical and climatic variables. Landsat imagery with a spatial resolution of 30 m resolution from 2015 to 2025 was used to calculate normalized difference vegetation Index (NDVI), soil adjusted vegetation index (SAVI), land surface temperature (LST), and net radiation (Rn).

The dataset was split with 80% used for training and 20% for validation, to simulate a continuous forecasting scenario. The primary objective was to evaluate model performance in an unseen future period, reflecting irrigation forecasts in practice rather than re-learning from shuffled segments through temporal cross-validation. By contrast, the 80-20 split ensured a long historical record for vigorous training and a sufficiently large, continuous block of unseen data that spans entire season irrigation requirement which validation against observed evapotranspiration (ET) and local climate data from the Eddy Covariance Flux Tower, providing reliable ground-truth checks across entire cropping cycle.

The machine learning models tested included CNN, XGBoost, and Random Forest. Among these, CNN achieved the highest performance with an R² of 0.89, followed by XGBoost (R² = 0.81) and Random Forest (R² = 0.76) on the testing samples. The short-term irrigation forecasting model was evaluated across two cropping seasons Kharif and Rabi, using observed ET values and local climate data from an Eddy Covariance Flux Tower. For rice during Kharif, CNN predicted 6.798 mm/day compared to the flux tower's 6.99 mm/day. During Rabi, the model predicted wheat ET at 2.041 mm/day, closely matching the observed 1.86 mm/day. During the growth phase of wheat in Rabi season, CNN forecasted ET at 29.87 mm/day, closely matching the flux tower measurement of 33 mm/day. Similarly, during early April, the model estimated 12.12 mm/day versus an observed 13.15 m/day. The lowest deviation occurred during the week of December, with both CNN and flux tower ET values closely aligned (6.99 and 6.89 mm/day, respectively). Overall, CNN showed the highest correlation than other models across multiple crops (maize, potato, guava, and orchards), showing strong spatial accuracy and temporal relevance. The outcomes support a wide range of users including farmers, local organizations, and decision-makers by enabling proactive irrigation planning.