Prediction of Soil Available Boron Content in Visible-Near-Infrared Hyperspectral Based on Different Preprocessing Transformations and Characteristic Wavelengths Modeling

Abstract

The trace element boron (Boron, B) is an important factor in crops’ development, pollination, and fertilization. Available boron (AB) in soil is the main source of boron nutrient absorption for crops. Rapid detection of AB is of great significance for crop nutrition diagnosis, soil testing and fertilization, precision agriculture development, scientific production management, and guarantee of stable yield and high quality. In this study, we propose a new method to predict soil available boron content using handheld nonimaging hyperspectroscopy in the visible-near-infrared range (350–1655 nm). As boron content is one of the fewest soil chemical elements, a rapid and accurate method has yet to be developed to detect and quantify the soil available boron. Visible-near-infrared ray (VIS-NIR) spectroscopy is widely utilized in the detection and quantification of soil available nutrients. There is, however, scant research on the detection of soil boron based on NIR data, and the performance of current regression model is still far from satisfactory. Our soil samples were collected from southern Anhui, China, with their NIR spectroscopy examined and the NIR data pretreated by 29 transformations and modeled with 10 regression algorithms. Of all the tested methods, SVM_RBF, BPNN, and PLS_RBF algorithms demonstrated the best performance and gave 0.80∼0.82 coefficient of determination value. At the same time, Random Forest algorithm (RFA), Successive Projection Algorithm (SPA), and Variable Importance in Projection (VIP) were used to extract the spectral characteristic wavelength data of soil available boron, and then the characteristic wavelength data were modeled with three regression algorithms: SVM_RBF, PLS_RBF, and BPNN. A comparative analysis of the prediction performance ( $R^2$ , RPD, RMSE, and RPIQ) of the models established at the full band showed that the RFA-MSC/BPNN model achieved the best performance. Compared with the best full-wavelength model DT/SVM_RBF, the test set achieved a 3.06% increase in $R^2$ , a 7.12% drop in RMSE, a 7.71% gain in RPD, and a 7.78% increase in RPIQ. Our work sheds lights on how to achieve rapid quantification of the soil available boron concentration.