关键词: 煤矿区, 复垦农田, 土壤有机质, 高光谱, 偏最小二乘回归

Abstract: 【Objective】 In terms of the problems in the Loess Plateau, such as many hills, complex topography, low soil organic matter content (SOMC), sampling difficulties, large areas of land damage caused by mining activities and so on, the object of this study is to provide an alternative method for the rapidly quantitative monitoring and evaluation of the SOMC in the process of land reclamation and comprehensive renovation. 【Method】 Taking the cropland soil in the coal mining areas in Xiangyuan County, Shanxi Province was picked as research object, 152 soil samples were collected from the intermediate strip area of land destruction region in a north to south direction. The physical and chemical properties of the soil samples were analyzed. At the same time, the raw hyperspectral reflectance (R) of the soil samples was measured by the standard procedure with an ASD FieldSpec 3 instrument equipped with a high intensity contact probe under the laboratory conditions. The raw spectral reflectance (R) were pretreated by the smoothing or denoising methods of multiplication scatter correction (MSC), baseline offset correction (BOC) and Savitzky-Golay filter in the ParLes 3.1 software. And the raw spectral reflectance (R) was transformed into two types of spectra, which were first order differential reflectance (D (R)) and inverse-log reflectance (lg (1/R)), to analyze the correlation coefficients between the three spectra and their SOMC. Then the significant bands were extracted by the significant correlation coefficients (P=0.01) of the three spectra with the SOMC. Finally, based on the full bands (400-2 400 nm) and significant bands of the three spectra, the hyperspectral predicting models of the SOMC were established by the method of partial least squares regression (PLSR). The optimal models were determined by the assessing indices of predicting accuracies, including coefficient of determination (R²), root mean square error (RMSE), and residual prediction deviation (RPD). 【Result】 The spectra in the bands of 400-1 800 and 1 880-2 400 nm for the raw spectral reflectance (R), 420-790, 1 020-1 040, and 2 150-2 200 nm for D (R), and 400-1 830 and 1 860-2 400 nm for lg (1/R), were significantly correlated with SOMC (P=0.01). And the maximum correction coefficients between the three spectra and their SOMC were 800 nm of the raw spectral reflectance (R), 600 nm of D (R), and 760 nm of lg (1/R). After the transformation of D (R), there were prominent differences among the absorption peaks of the spectral curves in different soil samples, and their correlation coefficients were improved from the value of 0.72 to that of 0.82 in the range of visible bands (400-800 nm). The models of significant bands could obtain better predicting accuracies compared with that of full bands by the method of PLSR. Among the three spectra, the predicting accuracy of lg (1/R) was the best, and R², RMSE of the calibration dataset were 0.95 and 7.64, while R², RMSE, and RPD of the validation dataset were 0.85, 3.00, and 2.56, respectively. For the models of R-PLSR and lg (1/R)-PLSR of full bands, the predicting abilities were good. The R², RMSE, and RPD of R-PLSR were 0.79, 3.64, and 2.10, respectively. And the coefficient of R², RMSE and RPD of lg (1/R)-PLSR were 0.79, 3.53, and 2.17, respectively. However, for the model of D (R)-PLSR, the predicting SOMC were only roughly estimated, and the indices of the predicting accuracies were not satisfying. R², RMSE and PRD of the D (R)-PLSR were 0.61, 5.43, and 1.41, respectively. Finally, by analyzing the predicting accuracies of the three spectra in both full bands and significant bands, it was found that the models of R-PLSR, D (R)-PLSR and lg (1/R)-PLSR in significant bands achieved desirable predicting effect. 【Conclusion】 In the study area, soil spectral reflectance has a high correlation with SOMC, and PLSR is a good method to establish the predicting model of SOMC.

Key words: coal mining areas, reclamation cropland, soil organic matter content (SOMC), hyperspectral, partial least squares regression (PLSR)