Hyperspectral estimation of desert soil organic matter content based on gray correlation-ridge regression model

Abstract: Organic matter content in soil is one of the most significant indicators evaluating the soil fertility, and its dynamic monitoring is good for further development of accurate agriculture. In recent years, obtaining Vis-NIR (visible - near infrared) continuous spectrum data of soil through hyperspectral technique and realizing accurate inversion prediction according to organic matter spectrum reflection characteristics have become a hot topic in current remote sensing field. However, in the hyperspectral inversion process of desert soil organic matter, there exists the problem of "low organic matter content, weak spectrum response and low model precision". The research collected different soil samples in Seder Boker region, south of Israel, divided the experimental soil samples into sandy soil and clay loam after particle size analysis in the lab, and applied potassium dichromate external heating method to measure the organic matter content in the soil. The raw hyperspectral reflectance of soil samples was measured by the ASD FieldSpec 3 instrument. After data preprocessing and different mathematical manipulation, 6 spectral indicators were obtained, i.e. reflectivity (REF), inverse-log reflectance (LR), continuum removal reflectance (CR), standard normal variable reflectance (SNV), first-order differential reflectance (FDR) and second-order differential reflectance (SDR). Then, gray correlation degree (GCD) between different spectral indicators and organic matter content was calculated, and SNV, FDR and SDR through gray correlation (GC) test (GCD>0.90) were chosen as the sensitive spectral indicators. Moreover, hyperspectral inversion model of soil organic matter was built based on sensitive spectral indicator using partial least squares regression (PLSR) method and ridge regression (RR) method, and the precision of inversion result was verified and compared. And then, the performances of these models were evaluated by the determination coefficient for calibration set (Rc2), determination coefficient for prediction set (Rp2), root mean squared error (RMSE) and relative percent deviation (RPD). The results indicated that: Soil particle size has a certain impact on the spectral response of organic matter, and the inversion effect of hyperspectral model on the organic matter content in sandy soil is superior to clay loam; after comparing and analyzing the models built according to different spectral indicators, Rc2, Rp2 and RPD of SNV-PLSR soil model and SNV-RR soil model built according to SNV are the highest and RMSE is the lowest, so SNV is the optimal spectral inversion indicator of soil organic matter; SNV-RR model has the most ideal inversion effect on organic matter content of these 2 kinds of soil: For sandy soil, Rc2 is 0.887, Rp2 is 0.866, RMSE is 0.610 g/kg and RPD is 2.72; for clay loam, Rc2 is 0.889, Rp2 is 0.863, RMSE is 0.898g/kg and RPD is 2.37. After analysis, it is known that SNV-RR model has extremely strong forecast ability for organic matter of sandy soil, and very good quantitative forecast ability for organic matter of clay loam. In addition, compared with PLSR model, Rc2 and Rp2 of RR model decline slightly. However, on the premise of ensuring precision, the number of band section used in modeling only accounts for about 4% of total spectrum. Not only does it simplify the model greatly, but also realizes "dimensionality reduction" and "optimization" of hyperspectral data. Through band selection function effect of RR method, the significant band section of soil organic matter is analyzed: The sensitive band of organic matter of sandy soil is mainly concentrated at 820-860 and 940-970 nm, but the sensitive band of organic matter of clay loam is concentrated at 730-790 and 800-820 nm. The united application of gray correlation analysis and RR method in the modeling analysis of soil organic matter provides a new approach to optimize the hyperspectral model and quickly measure the organic matter content in soil. GC-SNV-RR organic matter inversion model of 2 kinds of soil is simple and has good prediction. It provides support for remote sensing analysis on desert soil organic matter, which realizes the speedability and accuracy in monitoring the desert soil organic matter.