An LSTM approach to deciphering irrigation operations from remote sensing and groundwater levels records

Agricultural irrigation, the largest consumptive water user, significantly impacts terrestrial energy and water cycle, atmospheric boundary layer and the sustainability of water resources management. However, irrigation records usually lack the necessary detail in terms of amount, location, time and source with adequate spatial and temporal resolution that are required for understanding farmers’ irrigation behavior and representing irrigation in hydrologic models. This study addresses the irrigation scheduling gap by leveraging in situ groundwater level records of index wells and multi-source remote sensing observations. We used a Bi-directional Long Short-Term Memory (LSTM) network to capture the temporal relationship between groundwater fluctuations and land surface responses to irrigation. We trained the LSTM model to detect irrigation events based on groundwater level changes in the High Plains region of Nebraska and Kansas from 2001 to 2020. Using Integrated Gradients, an Explainable AI (XAI) technique, we identified that precipitation, MODIS evapotranspiration (ET), and Near-Infrared NIR reflectance are critical factors in detecting irrigation, with antecedent rainfall reducing irrigation likelihood. This framework enables allocation of long-term irrigation amounts to individual events, allows hydrologic models to assimilate irrigation dataset to assess irrigation impacts, and improves irrigation behavior representation in water resources management. • Daily groundwater depth at index wells indicates nearby irrigation events. • LSTM model predicts irrigation timing using remote sensing and hydroclimatic data. • XAI identified precipitation, ET, and NIR reflectance as key irrigation indicators.