Development of a Deep Neural Network Model for Predicting Reference Crop Evapotranspiration from Climate Variables

The reference evapotranspiration (ET0) is driven by various climate parameters such as temperature, wind speed, humidity, and solar radiation, which is an acute parameter in irrigation scheduling and many hydrology-related issues. Modeling ET0 in various regions including data-short regions is an effective way to estimate reliable and precise ET0 values. The current work aims to develop a model based on a deep neural network that forecasts the ET0 on a daily scale using the historical climate datasets of the Udupi station in Karnataka. The study investigated the potential of the deep learning regression model on a Keras framework built on top of the TensorFlow platform on a graphical processing unit (GPU). The model developed using complete climate variables predicted highly accurate results compared to the model with a single meteorological datum. The model's coefficient of determination (R2) value of 0.9979 with the five climate variables explained that 99% of the data fit the regression line. The model training and accuracy can be inferred from the number of epochs considered during the execution.