关键词: 土壤盐分, 光谱变换, 特征波段选择, 相关性分析, 最小角回归, 竞争性自适应重加权采样

Abstract:

【Objective】Soil salinization is a key environmental problem threatening the sustainable development of agriculture in arid areas, leading to the deterioration of soil structure, crop yield reduction and ecosystem degradation. The purpose of this study is to use spectral transformation, band selection and a variety of machine learning methods to build a soil salinity prediction model, which can quickly and accurately estimate soil salinity, and provide technical support for the scientific management of salinized farmland.【Method】Taking farmland soil in Dalate Banner as the research object, soil samples were collected systematically and their electrical conductivity (EC) and spectral reflectance data were measured. Firstly, Savitzky-Golay (S-G) filter was used to smooth the original spectrum (R). On this basis, 12 kinds of spectral transformation processing including reciprocal, logarithmic, first-order differential and second-order differential were carried out to mine the hidden spectral features. Then, the correlation analysis (CA) and least angle regression (LAR) methods were used to reduce the feature dimension, and the competitive adaptive reweighted sampling (CARS) algorithm was combined to further screen the sensitive feature bands. Finally, partial least squares regression (PLSR), support vector machine (SVM), back propagation neural network (BPNN) and random forest (RF) models were constructed based on the optimal features. The performance of the model was comprehensively evaluated by determination coefficient (R²) and root mean square error (RMSE), and the modeling effects of feature sets in different algorithms were compared.【Result】After spectral transformation, the correlation coefficients of the original spectrum were improved in varying degrees, indicating that spectral transformation could significantly enhance the correlation between soil salinity and spectral characteristics; When CARS was used for feature band optimization, LAR had better feature dimension reduction effect than CA; The reciprocal logarithmic first-order differential (ATFD) combined with PLSR model performed best, and its validation set accuracy was R²=0.81, RMSE=2.04 dS·m^-1; The comparison of different modeling methods showed that the performance of PLSR prediction model was better than the other three models (BPNN/RF/SVM), indicating that PLSR model was more suitable for the prediction of soil salinity in this region.【Conclusion】The hyperspectral prediction model of soil salinity based on ATFD-LAR-CARS-PLSR has high accuracy and optimal prediction ability, which proves that hyperspectral technology combined with multi-dimensional feature optimization can effectively realize the prediction of soil salinity in arid areas.

Key words: soil salinity, spectral transformation, characteristic band selection, correlation analysis (CA), least angle regression (LAR), competitive adaptive reweighted sampling (CARS)