A regional machine learning method to outperform temperature-based reference evapotranspiration estimations in Southern Spain

Abstract

Accurate estimations of reference evapotranspiration (ET0) are crucial for determining crop water requirements and designing an adequate irrigation scheduling to optimize the use of water. In this work, a new clustering method to outperform the accuracy of ET0 estimations only using temperature variables has been developed and assessed, based on the multifractal properties of the Diurnal Temperature Range (DTR). Thus, a more accurate weather stations’ grouping method has been evaluated, regardless of their geographic location. All the datasets were collected from 89 automated weather stations in the period 2000–2018 and pooled into two main regions (1 and 2). In each region, an iterative procedure has been carried out: 1) selection of all the stations except the candidate one for the training procedure and 2) test procedure using the candidate station. The results showed that Machine Learning models (ML) highly outperformed Hargreaves-Samani (HS) in most of the cases, being Multilayer Perceptron (MLP) the most accurate over Extreme Learning Machine models (ELM). On average, the results obtained by MLP using the best configuration in the first region were better than those obtained in the second region. Specifically, the first region got an Root Mean Square Error (RMSE) = 0.657 mm/d, Nash–Sutcliffe Efficiency (NSE) = 0.897, Coefficient of Determination (R2) = 0.931 and Mean Bias Error (MBE) = |0.04|mm/d while the second region obtained an RMSE = 0.703 mm/d, NSE = 0.867, R2 = 0.897 and MBE = |0.045|mm/d. Regarding the seasonal performance, spring and autumn obtained the best NSE and R2 results, whereas winter carried out the lowest RMSE values. This study provides a new and more accurate methodology to improve ET0 estimations on a regional basis and only using temperature data in the whole process.