AgroML: An Open-Source Repository to Forecast Reference Evapotranspiration in Different Geo-Climatic Conditions Using Machine Learning and Transformer-Based Models

Abstract

Accurately forecasting reference evapotranspiration (ET0) values is crucial to improve crop irrigation scheduling, allowing anticipated planning decisions and optimized water resource management and agricultural production. In this work, a recent state-of-the-art architecture has been adapted and deployed for multivariate input time series forecasting (transformers) using past values of ET0 and temperature-based parameters (28 input configurations) to forecast daily ET0 up to a week (1 to 7 days). Additionally, it has been compared to standard machine learning models such as multilayer perceptron (MLP), random forest (RF), support vector machine (SVM), extreme learning machine (ELM), convolutional neural network (CNN), long short-term memory (LSTM), and two baselines (historical monthly mean value and a moving average of the previous seven days) in five locations with different geo-climatic characteristics in the Andalusian region, Southern Spain. In general, machine learning models significantly outperformed the baselines. Furthermore, the accuracy dramatically dropped when forecasting ET0 for any horizon longer than three days. SVM, ELM, and RF using configurations I, III, IV, and IX outperformed, on average, the rest of the configurations in most cases. The best NSE values ranged from 0.934 in Córdoba to 0.869 in Tabernas, using SVM. The best RMSE, on average, ranged from 0.704 mm/day for Málaga to 0.883 mm/day for Conil using RF. In terms of MBE, most models and cases performed very accurately, with a total average performance of 0.011 mm/day. We found a relationship in performance regarding the aridity index and the distance to the sea. The higher the aridity index at inland locations, the better results were obtained in forecasts. On the other hand, for coastal sites, the higher the aridity index, the higher the error. Due to the good performance and the availability as an open-source repository of these models, they can be used to accurately forecast ET0 in different geo-climatic conditions, helping to increase efficiency in tasks of great agronomic importance, especially in areas with low rainfall or where water resources are limiting for the development of crops.

Keywords

• machine learning
• transformers
• neural networks
• support vector machine
• reference evapotranspiration
• forecasting