Data-driven estimation of actual evapotranspiration to support irrigation management: Testing two novel methods based on an unoccupied aerial vehicle and an artificial neural network

Abstract

Recent advances in remote sensing and machine learning show potential for improving irrigation use efficiency. In this study, two independent methods to determine the irrigation dose in processing tomatoes were calibrated, validated, and tested in an irrigation experiment. The first method used multispectral imagery acquired from an unoccupied aerial vehicle (UAV) to estimate the FAO-56 crop coefficient, Kc. The second method used an artificial neural network (ANN) trained on eddy covariance measurements of latent heat flux and meteorological variables from a nearby meteorological station. An irrigation experiment was conducted, where the farmer was instructed through a mobile application with updated irrigation recommendations. Evapotranspiration estimated by the new methods was set as the irrigation dose for the UAV and ANN treatments. The best-practice irrigation, commonly used by the regional farmers, was set as the control treatment (100%), guided by an irrigation expert and soil sensors for feedback. Derivatives of this treatment at 50%, 75%, and 125% of the control irrigation dose were tested. Yield, water use efficiency (WUE), and Brix level were measured and analyzed. Results show that both methods, UAV and ANN, estimated evapotranspiration to derive the irrigation dose at a near-perfect agreement with best-practice irrigation, both in the total amount and irrigation rate. Furthermore, there were no significant differences between the best practice and the experimental treatments in yield (117 ton/ha), water-use efficiency (31.7 kg/m3), and Brix (4.5°Bx). These results demonstrate the potential of advanced machine learning techniques and aerial remote sensing to quantify crop water requirements and support irrigation management.

Topics & Concepts