Integration of MRMR algorithm with advanced neural networks for modeling long-term crop water demand in agricultural basins

Abstract

Assessing actual evapotranspiration (AET) remains a key challenge in the design of efficient irrigation systems, strategies, and schedules. This complexity arises from the nonlinear nature of AET, which varies with crop type, growth stage, agroclimatic conditions, soil type, and irrigation method. In regions with limited weather data, such as parts of China, a major agricultural nation, precise AET estimation is crucial for optimizing the use of available irrigation water. Therefore, this study aims to achieve more accurate AET predictions through i) evaluating the performance of five artificial neural network (ANN) models optimized with the minimum redundancy maximum relevance (MRMR) algorithm to estimate monthly AET across diverse agroclimatic zones in China and ii) selecting the model with the highest accuracy based on performance metrics and minimal error between estimated and actual AET values. The analysis utilized weather data from Jinzhou, Anshan, Harbin, Shenyang, and Changchun from 1958 to 2021, with 75% of the data allocated for training and 25% for testing. AET was estimated by using five ANN architectures with the MRMR method ranking the AET predictors before modeling. The wide neural network (Wi-ANN) outperformed the other methods in both training and testing, achieving high accuracy across all metrics in the testing stage: R² (0.977), root mean square error 6.423 mm, and mean absolute error 3.371 mm. Overall, these findings underscore the robust capacity of the Wi-ANN model to forecast long-term AET at the studied sites. This approach offers a promising solution for enhancing irrigation practices and boosting agricultural productivity.