Time Series Forecasting to Support Irrigation Management

Abstract

Irrigated agriculture is the most water-consuming sector in Brazil, representing one of the main challenges for the sustainable use of water. This study has investigated and evaluated popular machine learning techniques like Gradient Boosting and Random Forest, deep learning models and univariate time series models to predict the value of reference evapotranspiration, a metric of water loss from the crop to the environment. The reference evapotranspiration ET0, plays an essential role in irrigation management since it can be used to reduce the amount of water that will not be absorbed by the crop. We performed the experiments with two real datasets generated by weather stations. The results show that the deep learning models are data-hungry, even when we increased the training set it was not enough to outperform multivariate models like Random Forest, Gradient Boosting and M5’ which indeed execute faster than the deep learning models during the training phase. However, the univariate time series model as the evaluated deep learning models (stacked LSTM and BLSTM) is a viable and lower-cost solution for predicting ET0, since we need to monitor only one variable.