Prediction of Irrigation Needs at Farm Level Using Recurrent Neural Networks.

The efficient management of irrigation water is key to reduce crop production risks and protect the environment. Ensuring adequate soil moisture for plant growth requires the temporal monitoring of the soil water content at several soil depths, which represents a sequential problem. In this study, recurrent neural networks (RNN) are proposed to define irrigation prescriptions. RNNs have chain-like structures of repeating modules. An internal self-looped cell is created, which allows it to display dynamic temporal behavior that help RNNs to remember previous information and process arbitrary sequences. Long short-term memory (LSTM) is a novel RNN architecture designed to overcome vanishing problems in the training process. During 2017 and 2018, an on-farm soil-sensor based irrigation scheduling study was conducted in Samson, Alabama. Ten sensor probes located on a loamy sand and clay loam soils collected hourly soil water tension data at 15 cm, 30 cm, and 60 cm soil depths during 120 days. Soil water retention curves were used to convert hourly soil water tension data into soil moisture and to estimate hourly irrigation prescriptions. Several LSTM models were trained over different hyperparameters: hidden neurons, learning rate and training iterations. The prediction performance of these models was studied by calculating RMSE (Root Mean Square Error) and coefficient of determination . Experiments were defined using different train-validation-test splits with combinations of individual sensor data, soil type and years, with a maximum dataset of 17568 records and seven categorical attributes. Soil water tension at three depths, weather data and irrigation amounts were used as inputs to the LSTM networks. The results show that LSTM Neural Network approach demonstrate good prediction capability of irrigation prescriptions for the soil types studied (0.9 and mm). Therefore, LSTM could be a method used to determine irrigation scheduling strategies based in soil water tension sensors data.