Transferable field-scale daily soil moisture forecasting in cotton with a minimal-input LSTM for irrigation scheduling

This is the authors' abstract. We don't add key points for this paper.

Accurate daily, field-scale soil moisture forecasts are essential for data-driven irrigation scheduling, yet machine learning–based operational frameworks for this task remain limited. Using cotton field measurements and on-farm weather data from 2023 (training/validation) and an independent 2024 season (transfer learning), we benchmarked random forest (RF), XGBoost, support vector machine (SVM), and long short-term memory (LSTM) for sequential soil moisture prediction. Among these, LSTM consistently achieved the highest accuracy, capturing temporal dependencies that traditional machine learning models could not. Feature importance analysis identified precipitation, solar radiation, and vapor pressure as the three dominant drivers. A minimal-input LSTM trained with these three drivers maintained near-baseline performance (R 2 =0.95). Transfer learning of the minimal-input LSTM model to an independent 2024 growing season data showed moderate predictive skill via fine-tuning (R 2 =0.56). The operational deployment of this compact LSTM model enabled field-scale irrigation scheduling using a simple soil moisture threshold, which could be enhanced by hysteresis, short-memory smoothing, and water balance checks. The results show that: (i) sequence-based LSTM architectures are robust for daily, field-scale soil moisture forecasting, (ii) a compact, physically informed driver set (precipitation–radiation–vapor pressure) maintains high model accuracy with low data requirements, and (iii) forecasts can be directly translated into clear irrigation decisions. The approach provides an interpretable and scalable framework that links machine learning models with practical water management and offers a low-burden, high-accuracy workflow for precision irrigation in sandy soils.