On the use of machine learning methods to improve the estimation of gross primary productivity of maize field with drip irrigation

Abstract

As an important part of the energy-matter cycle, gross primary productivity (GPP) reflects the ability of plants to absorb carbon dioxide from the atmosphere. The accurate estimation of GPP is critical to understanding the regional carbon cycle. With the development of machine learning (ML) theory, machine learning models are increasingly used to study complex phenomena with high variability in time and space. In this study, three machine learning models (Support vector regression (SVR), artificial neural network (ANN), and long short-term memory networks (LSTM)) were investigated for predicting GPP in northwest China and compared them with the traditional physical models. Carbon flux, various environmental factors, and maize growth indices were measured in the maize field over five years in northwest China. The rigorous analysis which included statistical comparison and cross-validation for the prediction of GPP confirmed that the machine learning models performed better than traditional physical models. And the SVR model performed the best among the considered ML models with the highest nash-efficiency coefficient and the lowest root mean squared error. The machine learning model also outperformed the traditional physical models on cloudy days and after irrigation. The SVR achieved good prediction accuracy and high stability. With different training data sets, the ANN and LSTM were relatively more sensitive to the training data set. When the training data was sufficient, SVR, ANN and LSTM could achieve similar prediction accuracy, but SVR was slightly higher. When the training data was small, the simulation accuracy of SVR was better than ANN and LSTM. The performance of ANN and LSTM was more sensitive to parameter selection, and the relationship between model performance and parameter selection had no obvious regularity. Based on this comprehensive comparison study, it was elicited that the SVR model can be successfully applied to GPP simulation of maize fields, which provided a new perspective for the application of machine learning modeling in GPP simulation.