Regional soil moisture prediction system based on Long Short-Term Memory network

Abstract

In the context of climate change, drought has been recognised as one of the most severe threats for agricultural production since absence of water is one of the most limiting factors for the growth of plants. In this study, a regional system for soil moisture prediction based on ERA5 climate reanalysis dataset, an open-source meteorological dataset issued by Copernicus Climate Change Service, was developed. It consisted of the relevant meteorological parameters for Serbia, during the period 2011–2020. Daily values of maximum and minimum air temperature, precipitation and vapour pressure deficit were used as features, and they were fed to recurrent neural network in order to predict volumetric soil moisture for three days ahead. A Long Short-Term Memory (LSTM) network was designed and trained at a regional scale using the data from the 2011–2016 period, at 28 locations that cover four major soil types. Validation was done on 2017–2018 data and LSTM was compared with a statistical forecasting technique and a classical machine learning approach. Evaluation was done with error measures commonly used in literature. The resulting network yielded the lowest errors and proved to have good generalisation properties. The system was tested from September 2019 to April 2020 using short-range weather forecasts of Yr service. It will present the cornerstone of the irrigation scheduling service in AgroSense.rs, the Serbian national platform for digital agriculture.