Abstract:

In the present study, the airborne hyperspectral data were selected as the data source, and the spectra in the study area were transformed by continuum removal(CR), inversion recovery(IR), logistic regression(LR), first derivative reflectance(FDR), second derivative reflectance(SDR), inversion first derivative reflectance(IFDR), logarithm first derivative reflectance(LFDR), inversion logarithm regression(ILR), respectively,to construct normalized difference spectral index (NDSI), and these constructed NDSI data were denoted as NDSI-CR, NDSI-IR, NDSI-LR, NDSI-FDR, NDSI-SDR, NDSI-IFDR, NDSI-LFDR, NDSI-ILR, respectively. The correlation between NDSI and humic acid content was analyzed to identify the characteristic spectra. On this basis, multiple linear regression (MLR), partial least squares (PLSR), back propagation neural network (BPNN) and support vector machine (SVM) models were introduced to construct prediction models. The coefficient of determination (R²), root mean squared error (RMSE) and ratio of performance-to-deviation (RPD) were used as model evaluation indexes to select the best modeling method for the estimation of humic acid content at the field scale. It was shown that NDSI-FDR, NDSI-SDR, NDSI-IFDR, NDSI-LFDR had a higher correlation with humic acid content. In 396-1 000 nm, there were three sensitive band intensive regions with the humic acid content, which were located in the coordinate regions of 480-550 nm and 510-570 nm, 730-790 nm and 740-800 nm, and 880-930 nm and 880-930 nm.For the established BPNN model based on NDSI-LFDR, its R ² on the modeling set and validation set was 0.916 and 0.805, respectively, its RMSE on the modeling set and validation set was 0.799 and 1.107, respectively, and its RPD was 2.189, which could satisfy the requirement of field-scale estimation of humic acid content.

Key words: humic acid content, airborne hyperspectra, characteristic spectrum