Degraded land detection by soil particle composition derived from multispectral remote sensing data in the Otindag Sandy Lands of China

Abstract

Remote sensing technology has shown considerable potential for the estimation of soil properties. In this paper, we proposed a method by using the measured hyperspectral and BJ-1 multispectral image to estimate the silt content in soil quantitatively, and to develop a soil-based model which could be used in detecting desertification or land degradation. The test site was located in the Otindag Sandy Lands in Xilingol League, Inner Mongolia, China. Soil sampling was carried out to analyze the soil particle composition, and the spectral properties of soil samples were also examined in the laboratory. The differences in soil particle composition between the degraded lands and others were distinguished statistically, and the correlation of soil particle contents with spectral reflectance was analyzed based on the Partial Least Squares Regression (PLSR) to determine the sensitive band needed to calibrate the prediction model. The validity of the models was assessed by using the Prediction Residual Error Sum of Squares (PRESS). The results showed that silt content was an important factor to indicate land degradation, and it descended gradually with the development of land degradation. There was significant difference in silt content between degraded lands and the others, the silt content in undegraded lands was nearly three times greater than that in degraded lands. R2 of PLSR model based on the BJ-1 multispectral image (R2=0.504) was lower than that of the measured spectral model (R2=0.725), and validated PRESS (PRESS=0.753) was also higher than that based on the measured spectra (PRESS=0.562). However, the BJ-1 multispectral image did show a considerable ability to predict the soil silt content. Compared with the normal regression method, PLSR was not only effective in dependent or independent variable selection, but it was also reliable to determine the regression model with higher stability and lower error. By combining statistics and manually interactive adjustment, the thresholds of silt contents for categorizing different degraded lands were determined as 3.5% and 8.8%, and it can effectively detect the degraded areas from others; total classification accuracy of this approach reached 86.21%. It verified that the silt content was the best indicator indeed to detect the desertification or land degradation in drylands.