AI-Based Irrigation Demand Forecasting for Climate-Adaptive Crop Water Management in Irrigated Agriculture

Accurate and timely forecasting of crop irrigation demand is a critical agrometeorological challenge, particularly in regions experiencing intensifying climate variability, water scarcity, and growing food insecurity. Weather conditions, including temperature extremes, precipitation deficits, and elevated evaporative demand, directly drive crop water requirements and can significantly reduce yields when irrigation scheduling is poorly timed. Existing studies on crop water demand have employed machine learning (ML) for actual evapotranspiration (ETa) prediction, incorporating feature selection to reduce data dimensionality and mitigate multicollinearity, thereby improving model performance. However, existing ETa models have limited spatial continuity and restricted applicability under changing climate conditions.

This study addresses these limitations by developing a machine learning-based Irrigation Demand Forecasting Model (AI-IDFM) that integrates multi-sensor satellite data, spectral indices, and various climatic variables to determine crop water requirements under changing climate conditions. The actual crop evapotranspiration (ETact), a key agrometeorological variable linking atmospheric demand and crop water use is estimated through a surface energy balance framework using Landsat satellite data with 30 m spatial resolution, complemented by high-resolution (3 m) land cover information, capturing fine-scale microclimatic variability relevant for field-level irrigation decision-making. The key variables, including NDVI, SAVI, LST, and net radiation (Rn), along with climatic variables such as precipitation, temperature, and humidity, are incorporated into the model. Deep learning and machine learning approaches, including Convolutional Neural Network–Multilayer Perceptron (CNN-MLP), Random Forest (RF), and Gradient Boosting (GB), were used for training and validation. The CNN-MLP model performed best, achieving an R² of 0.90, followed by RF and GB with R² values of 0.87 and 0.80, respectively. From this study, the short-range irrigation advisories are generated by coupling AI-IDFM with daily meteorological outputs from the ICON ensemble model, providing 7-day forecasts of key atmospheric drivers, including temperature, total precipitation, and relative humidity. Forecasted inputs for April 2025 indicated temperatures ranging from 36.2-42.8°C, average humidity between 18.5-43.7%, reflecting intense thermal stress during the pre-Kharif transition period that directly amplify crop water demand.

Model performance showed that CNN predictions closely matched observed ETact for rice (6.798 vs. 6.99 mm/day) and wheat (2.041 vs. 1.86 mm/day), validated against eddy covariance flux tower observations across Kharif and Rabi cropping seasons in Okara district, Punjab. The model consistently outperformed ensemble methods across diverse crop types. During the Rabi wheat growth phase, CNN forecasted ET at 29.87 mm/day compared to a flux tower measurement of 33 mm/day, with the lowest deviation recorded in December (6.99 vs. 6.89 mm/day). The results demonstrate that AI-IDFM provides a robust, weather-informed irrigation advisory framework capable of supporting farm-level decision-making across diverse cropping systems. The approach links weather data to actionable irrigation decisions, enabling scalable management from field to regional levels, improving water use efficiency and supporting climate-resilient agriculture in data-scarce systems.