Search

[Objective] The present paper aimed to realize the rapid estimation of soil pH in tobacco field on the regional scale. [Method] 296 tobacco field soils in Sichuan province were studied. The spectral reflectance of 390 - 2500 nm band of farmland soil was obtained by hyperspectral imaging technology. The effects of 12 spectral preprocessing methods, 2 characteristic band screening methods and 4 modeling methods on the accuracy of the hyperspectral estimation model of soil pH in tobacco field were systematically studied. [Result] In the range of 600 - 2500 nm, the spectral reflectance curves of different pH were obviously different. The accuracy of the estimation model was improved after single and combination preprocessing of the original spectrum. Among them, the first derivative (Dl) combined with standard normal variate (SNV) was the best spectral preprocessing method. A total of 93 soil pH characteristic bands were selected by competitive adaptive reweighted sampling (CARS) algorithm, which were mainly concentrated in the near infrared band 800 - 2500 nm. Both kernel ridge regression (KRR) and support vector machine (SVM) could obtain better estimation accuracy than partial least square regression (PLSR) and ridge regression (RR), regardless of spectral features obtained by all-band spectrum or principal component analysis dimension reduction. However, the estimation accuracy of PLSR and RR models was higher than that of KRR and SVM models in the feature bands screened by CARS. Based on Dl-SNV preprocessing method, the PLSR estimation model of soil pH established by using CARS screening characteristic bands had high accuracy, the coefficient of determination of model validation set was 0. 758, the root mean square error was 0. 555, and the residual prediction deviation was 2. 034. [Conclusion] It is feasible to use hyperspectral imaging technology to estimate the pH of tobacco field with high accuracy at regional scale, and the pH can be estimated by D1-SNV-CARS -PLSR model. [ABSTRACT FROM AUTHOR]