| Abstrakt: |
Taking the lakeside oasis on the west bank of Bosten Lake as the study area, and using measured soil organic carbon content data and soil hyperspectral data, the original spectrum R was mathematically transformed and differentially transformed. The competitive adaptive reweighted sampling (CARS) was used to screen the characteristic variables under different spectral forms, and the selected characteristic band was used to build a BP neural network model to estimate the carbon content of soil organic. The results showed that the soil surface organic carbon content in the study area ranged from 0.80 to 63.15 g/kg, with an average value of 17.57 g/kg, and the coefficient of variation was 71.48%, which was medium variability. The CARS algorithm compressed the modeling input bands to less than 2.76% of the full band number, and R, R', 1/R, (1/R)', log(1/R), log(1/R)', 1/logR, and (1/logR)' spectral forms were more concentrated in the NIR long-wave 1 500-2 500 nm and visible band 380-760 nm; R", (1/R)", log(1/R)", and (1/logR)" spectral forms were more concentrated in the NIR band 760-2 500 nm. The accuracy of CARS-BP estimation model constructed by second-order differential transformation was better than that of first-order differential transformation, with the best estimation effect as R"-CARS-BP. The R² of the training set and validation set were 0.81, 0.83, RPD were 2.30, 2.45, and RMSE were 5.75 g/kg, 4.89 g/kg, respectively. [ABSTRACT FROM AUTHOR] |
