Abstract: In this paper, a total of 115 lime concretion black soil samples collected from the northern Anhui Plain, China, were used as research objects to obtain hyperspectral data. Competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), and random forest feature selection (RFFS) were used to select the characteristic wavelength of soil total nitrogen content from 224 wavelengths of the hyperspectral data. Partial least square regression (PLSR), support vector regression (SVR), and least absolute shrinkage and selection operator (LASSO) were applied to establish the spectral regression model of soil total nitrogen content. It was shown that all of the wavelength-selecting models outperformed the full-wavelength models except for the CARS-PLSR model. By comparison of all the prediction models built by different combinations of wavelength-selecting methods and regression algorithms with respect to the prediction performance, it was found that the RFFS-LASSO model with 20 characteristic wavelengths got the best prediction results. The coefficient of determination (R²) and relative percent deviation (RPD) value of the model prediction set were 0.787 1 and 2.130 1, respectively. The results illustrated that RFFS-LASSO model was simple and effective for the prediction of soil total nitrogen content, and it had certain guiding significance for the development of proximal sensor of soil total nitrogen content.

Key words: precision agriculture, mathematical modelling, soil chemistry