Combining dynamic Bayesian prediction of the crop coefficient with automated lysimetry for highly accurate water-use control

Crop evapotranspiration is a key parameter for efficient irrigation, that can be estimated by means of the standardized FAO-56 technique based on a reference evapotranspiration and on a crop-specific coefficient (Kc). The latter, however, must be experimentally measured and varies with the environmental conditions and with the crop's leaf area development over time. The experimental assessment of Kc simulates intricate soil–plant-atmosphere interactions. This involves solving numerous equations that require extensive daily meteorological, crop, and substrate data over the entire growing season. Since this process demands significant time and resources, irrigation on the base of crop evapotranspiration is often disregarded. To make this method of irrigation feasible in soilless cultivation instead, we introduce a smart lysimeter composed of commercial off-the-shelf sensors, a cloud server and a lightweight IoT client that controls the irrigation duration such that the leaching fraction is kept constant. Sensor data acquired by the lysimeter is analyzed on the cloud server by means of an AI component based on a Dynamic Bayesian Network to produce a daily estimate of Kc. The AI component learns from sensor data over time to predict the evolution of Kc several days ahead and to adjust the irrigation threshold in the IoT client. Combining the Kc estimate with short-term weather forecast, the smart lysimeter can easily forecast the crop evapotranspiration and actuate an efficient irrigation, even across different geographical regions. The experiments conducted with the ornamental guelder rose (Viburnum opulus L.) under different weather conditions show that, during a Mediterranean summer, our lysimeter achieved 17% water savings in irrigation and a 64% reduction in runoff compared to manual irrigation by experienced farmers during periods of variable weather. The RMSE of Kc, averaged across all prediction horizons from one to seven days ahead, is 6.7% of full scale.