Evaluating Performance of Multiple Machine Learning Models for Drought Monitoring: A Case Study of Typical Grassland in Inner Mongolia

Abstract

Driven by continuously evolving precipitation shifts and temperature increases, the frequency and intensity of droughts have increased. There is an obvious need to accurately monitor drought. With the popularity of machine learning, many studies have attempted to use machine learning combined with multiple indicators to construct comprehensive drought monitoring models. This study tests four machine learning model frameworks, including random forest (RF), convolutional neural network (CNN), support vector regression (SVR), and BP neural network (BP), which were used to construct four comprehensive drought monitoring models. The accuracy and drought monitoring ability of the four models when simulating a well-documented Inner Mongolian grassland site were compared. The results show that the random forest model is the best among the four models. The R2 range of the test set is 0.44–0.79, the RMSE range is 0.44–0.72, and the fitting accuracy relationship could be described as RF > CNN > SVR ≈ BP. Correlation analysis between the fitting results of the four models and SPEI found that the correlation coefficient of RF from June to September was higher than that of the other three models, though we noted the correlation coefficient of CNN in May was slightly higher than that of RF (CNN = 0.79; RF = 0.78). Our results demonstrate that comprehensive drought monitoring indices developed from RF models are accurate, have high drought monitoring ability, and can achieve the same monitoring effect as SPEI. This study can provide new technical support for comprehensive regional drought monitoring.

Keywords

• environmental remote sensing
• drought monitoring
• machine learning
• model construction