Estimation of actual crop evapotranspiration using artificial neural networks in tomato grown in closed soilless culture system

Actual crop evapotranspiration (ETc) is not only essential for irrigation management but also difficult to estimate and measure. The classical “two-step” approach of ETc estimation, using the reference evapotranspiration (ETo) and the crop coefficient (Kc), is utilised worldwide. In this study, we aimed to evaluate the potential of artificial neural networks (ANN) for ETc estimations in closed soilless culture systems. Tomato (cv. Duru) plants were grown in two substrates, perlite and coco fibre, in an unheated plastic greenhouse. During the growing seasons (2 spring and fall periods each) daily climatic values, ETc and Kc were determined. FAO Penman-Monteith (PM) was used as the general approach for ETc estimation in addition to Hargreaves (HG) for limited data conditions. ANN models were generated with 4 data sets: (1) all inputs, (2) without radiation (-Rs), (3) without Kc (-Kc), and (4) minimum data (-Rs -Kc), and all were compared implementing PM and HG equations. The Nash–Sutcliffe model efficiency coefficient (E) was used to assess the predictive skill of the models. Average Kc values of 0.40–1.31–0.61 and 0.40–1.76–0.77 were obtained for perlite and coco fibre, respectively. The ANN models for all input conditions produced successful results, improving estimation accuracy over traditional equations. For all data sets, the average RMSE and model efficiency values (RMSE = 0.540 mm day−1 and E = 0.817) of ANN models for perlite and coco fibre, were better than those obtained from the PM (RMSE = 0.667 mm day−1 and E = 0.759). The ANN (+1) models forecasted ETc with E = 0.861 and 0.532 RMSE for coco fibre, and with E = 0.797 and 0.522 RMSE for perlite. It was concluded that ANN models used for site-specific ETc predictions perform better than classical methods in soilless culture, and can be implemented into automation systems with forecasting ability to assist irrigation management.