Reference evapotranspiration time series forecasting with ensemble of convolutional neural networks

• Published in: Computers and electronics in agriculture 2020-10, Vol.177, p.105700, Article 105700
• Main Authors: de Oliveira e Lucas, Patrícia, Alves, Marcos Antonio, de Lima e Silva, Petrônio Cândido, Guimarães, Frederico Gadelha
• Format: Article
• Language: English
• Subjects: Agricultural management; Artificial neural networks; Convolutional neural networks; Crops; Ensemble learning; Evapotranspiration; Forecasting; Irrigation; Mathematical models; Neural networks; Population growth; Reference evapotranspiration; Statistical analysis; Sustainable development; Time series; Time series prediction; Water consumption; Water management; Water resources management
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105700

•Reference evapotranspiration forecasting is relevant to irrigation management and smart farming.•Ensembles of convolutional neural networks achieved high accuracy and low variance of predictions.•Estimation of uncertainty is provided by probabilistic forecasting. The population growth and climate change are making the agricultural sector to seek more accurate and efficient approaches to ensure an adequate and regular supply of food for society with less water consumption. Irrigation management, an essential practice for the development of sustainable agriculture, seeks, through the forecast of Reference Evapotranspiration (ETo), to know in advance the water requirements of crops to plan and manage water resources. However, there is still a gap in the literature regarding the application and evaluation of deep learning models and ensemble models to forecasting reference evapotranspiration in irrigation management. In this context, this paper aims to explore the use of Convolutional Neural Networks (CNNs) in the prediction of ETo time series. Three CNNs with different structures were employed to predict a daily time series. Also, in order to evaluate the ensemble forecast of ETo time series, we apply four ensemble models composed of these CNNs in order to produce a probabilistic forecast. This information can be useful for planning and controlling irrigation of crops. Through experimental tests on a real database, the results showed the feasibility of the CNN models for forecasting ETo and that ensemble models were better than the well-known Seasonal ARIMA and Seasonal Naive and improved predictions in terms of variance, precision and computational cost in relation to the individual CNN models, in addition to allowing the estimation of uncertainty, as their outputs are probability distributions. In order to promote reproducibility of this research, all data and codes are publicly available.