Development of Soil Moisture Model Based on Deep Learning

Abstract

Soil moisture is an essential hydrologic parameter in agricultural studies, as most services in agricultural ecosystems, such as water storage, food and fiber supply, and water and climate regulation are associated with its spatial and temporal variability. Furthermore, soil moisture is considered to be the most important soil factor for runoff and flash flooding, due to its significant variation even on a daily time scale. The research area is the plain of Arta, which is located in the Epirus region, in the westernmost part of Greece, and includes the lower part of the hydrological basins of the rivers Aracthos and Louros. They provide water to an extensive irrigation network, while on the top edge of the plain a massive hydroelectric dam has been constructed on the Aracthos riverbed, with an annual production of 300 MW. The distinctive characteristic of the area is the elevated precipitation height, reaching a value of 2000 mm/a. A few agro-meteorological weather stations have been installed within the research area, continuously providing precipitation, humidity, solar radiation, wind direction, wind speed, temperature, soil temperature, and soil moisture data since 2015. Using the abundance of available data, which are prerequisites for the construction of a soil moisture deep learning model, a prediction model was created. As the aim of this study is the development of a feasible and efficient model for soil moisture prediction, several classes of artificial neural networks were applied, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The achievement of good data fitting and high accuracy in predicting the trends and actual values of soil moisture data led to the creation of a theoretical basis for precision agriculture, in terms of water usage and potential drought or flood control.