Estimating evapotranspiration in screenhouses using artificial neural network models

Increasing crop production under screens and in screenhouses has increased the interest in estimating crop evapotranspiration (ET) in this semi-protected environment. This study examines using artificial neural networks (ANN), for ET estimation, based on measured meteorological variables. Trained ANNs can provide ET estimates based on simple meteorological data. The experiment was carried out in a large banana screenhouse in the Carmel coast region, Israel, during a period of 52 days, from June to August 2016. Direct measurement of evapotranspiration was estimated with an eddy covariance system. Five internal micro-meteorological variables, air temperature and humidity, net and global radiation and air velocity, were measured on the same tower. The ANN model was implemented using MATLAB. All data collected during 52 days were divided into two periods: learning and verification. The ANN model was operated in two modes. The first included as input all 5 micro-meteorological variables and in the second, input data included different subsets of these variables. Running the model with all 5 variables gave a fit with a deviation of 7% from direct ET measurement and R²=0.94. In the second mode of operation, results showed that using only air temperature and global radiation as input data provided a reasonable estimation with a deviation of 8% from the measurement (R²=0.95). In summary, the results show that trained ANN models can provide accurate estimates of banana ET in a screenhouse based on micro-meteorological measurements.