Machine learning models for actual crop coefficient estimation on sensor-less fields

Efficient freshwater management in agriculture is a critical global priority due to increasing pressures from climate change and a growing population. Irrigated agriculture, which accounts for 70% of global freshwater withdrawals, plays a vital role in food security but faces significant challenges due to water scarcity and changing climatic conditions. In this context, accurate estimation of the actual crop coefficient (Kc-act) is essential to optimise irrigation strategies, thus improving water use efficiency, and enhancing the resilience of agricultural systems. This study proposes an innovative machine learning framework to estimate Kc-act across different temperate and continental climatic zones, without relying on field sensors. By integrating ERA5-L reanalysis data, the model provides accurate Kc-act estimates even in the absence of complete datasets. The framework was tested on a wide range of crops, including both homogeneous herbaceous crops (e.g., wheat and maize) and tree crops (e.g., citrus and olives), demonstrating its adaptability to different agricultural contexts. The process includes data preprocessing, actual evapotranspiration (ETa) prediction (using a Random Forest model), and seasonal decomposition to regularise the data. Results demonstrate significant accuracy improvements over traditional methods, with RMSE values as low as 0.073 for citrus orchards and 0.143 for olive groves. The analysis reveals variations in the values Kc-act influenced by climatic and field management conditions, offering a practical tool to optimise irrigation strategies and improve the efficiency of water use. The model reduces the dependency on expensive equipment and provides a scalable solution to monitor and manage water resources in agriculture.