%A JING Xia, ZHANG Jie, WANG Jiao-jiao, MING Shi-kang, FU You-qiang, FENG Hai-kuan, SONG Xiao-yu %T Comparison of Machine Learning Algorithms for Remote Sensing Monitoring of Rice Yields %0 Journal Article %D 2022 %J SPECTROSCOPY AND SPECTRAL ANALYSIS %R 10.3964/j.issn.1000-0593(2022)05-1620-08 %P 1620-1627 %V 42 %N 05 %U {https://www.gpxygpfx.com/CN/abstract/article_12669.shtml} %8 2022-05-01 %X Rice is China’s major grain food crop, grown mainly in the Yangtze River valley and southern China and on the Yunnan-Guizhou Plateau. The use of hyperspectral remote sensing technology to monitor rice yield before it matures is important. It can promptly adjust cultivation management methods and guide reasonable fertilization, and accurately grasp rice yield information to help the government make decisions in advance. In this study, the nitrogen fertilizer gradient experiment was carried on 2019—2020 at Zhongluotan experimental base in Baiyun District, Guangzhou City. The rice canopy hyperspectral data, crop population growth parameters (plant above-ground biomass (AGB) and leaf area index (LAI)) and crop nitrogen nutrient intakes at the rice differentiation and heading stage were obtained. Three machine learning algorithms, Bayesian Ridge Regression (BRR), Support Vector Regression (SVR), and Partial Least Square Regression (PLSR), were used to establish yield estimate models based on the different data sources, including rice canopy spectrum data, spectrum data combined with crop growth parameters, and spectral data, crop growth parameters, and crop nutrient intake data. The estimation accuracy of BRR, SVR and PLSR models, were evaluated and compared, then the best estimation model and optimal estimation growth period for rice yield were determined. The results showed that among the three methods, the BRR and SVR methods were more suitable for yield monitoring, with better performance in different periods and different parameter combinations (R2>0.82, NRMSE<8.22%). Based on the 2019 and 2020 data, the best monitoring model for yield monitoring using full-band spectral information was the BRR model at the differentiation stage with R2 of 0.90 and the SVR model at the tasseling stage with R2 of 0.87. When full-band spectral information was used for yield monitoring, the best monitoring model for both periods was the BRR model with R2 of 0.90 and 0.92, respectively, compared with the BRR and SVR models, the highest R2 of the PLSR model was only 0.75 for different periods and different combinations of parameters; based on the 2020 data, when three different parameter combinations were used as inputs, the BRR model was the best in both periods, and the modeling accuracy was higher in the differentiation period than in the tasseling period (R2 increased by at least 0.02 and NRMSE decreased by at least 0.61%), and when the input parameter combinations were full-band spectra with crop population growth parameters and crop nutrient uptake, the accuracy of the BRR model for yield estimation reached Through the experimental study, it was concluded that the optimal monitoring period for yield is the differentiation period. The optimal monitoring model is the BRR model. The study results can provide a reference for early remote sensing monitoring of rice yield.