Comparation of rice yield estimation model combining spectral index screening method and statistical regression algorithm

Abstract: A crop yield is one of the most important parameters in agricultural production. An accurate estimation of regional crop yield can greatly contribute to agricultural production management and national food policy. However, only a few studies have been focused on the combined effects of different exponential screening and statistical regression at present, even though there are various models of crop yield estimation. In this study, a comparative investigation was performed on the three types of index screening and three regression models, in order to explore the coordinated effect of the estimation model for the rice yield. The influence mechanism was also proposed to achieve an optimal yield estimation model suitable for the local production conditions. An important rice-producing area, the Sanjiang Plain in the Heilongjiang Province of China was taken as the study area. The rice unit yield and MOD09A1 remote sensing data were collected in the Bielahong River basin of the study area in 2019. After preprocessing, a total of 36 remote sensing variables were obtained, where four original bands and five vegetation indices of rice at the four growth stages, including the tillering, booting, heading, and milk ripening stage. Subsequently, the remote sensing variables were screened for the high sensitivity to the rice yield using the correlation coefficient (r), Variable Importance in Projection (VPI), and Out-Of-Bag (OOB) data importance analysis. After that, nine estimation models of rice yield were constructed to combine with the Random Forest (RF), Support Vector Machine (SVM), and Partial Least Squares (PLS) regression, such as the r-RF, r-SVM, r-PLS, VIP-RF, VIP-SVM, VIP-PLS, OOB-RF, OOB-SVM, and OOB-PLS. Several experiments were carried out for each model. Thus, the best input data was achieved for the optimal model. The determination coefficient, Root Mean Square Error (RMSE), and normalized Root Mean Square Error (nRMSE) were also used to evaluate the model. The results showed that the same index screening was fitted the different models with different degrees, where the OOB was more suitable for RF, the VIP was more suitable for r and PLS, and the r was more suitable for SVM. Specifically, the PLS and SVM model performed better in the three modelings, whereas, the RF model performed the best, among which the combined OOB-RF model was the best, with the model determination coefficient of 0.742, RMSE of 206 kg/hm2, and nRMSE of 3.10%. Therefore, the index screenings varied greatly with the regression, where the OOB-RF model presented the best yield estimation in the study area. This finding can provide a strong theoretical reference to integrate the exponential screening and regression for the rice yield estimation model.