Remote Sensing Estimation Of Heavy Metal Cd And Pb In Cultivated Soil Based On HSI Hyperspectral Data

How to get the information of soil heavy metal content and its spatial distribution is the foundation of soil pollution prevention and control and has become a hot topic in the research field of soil heavy metal pollution. Hyperspectral analysis technology can obtain the fine spectral information of ground objects because of its characteristics of high spectral resolution, multi band and strong continuity. The hyperspectral analysis technology provided a new way for non destructive soil heavy metal monitoring and provided a theoretical basis for the monitoring of soil heavy metals in solis with hyperspectral remote sensing. In this research, authors extracted the original soil spectral from the HSI hyperspectral remote sensing data by preprocessing. This paper have analysised the correction between cadmium content, lead content and the spectral variables, confirmed the spectral characteristic bands of cadmium and lead, established the inversion model of cadmium and lead preliminarily based on the preprocessing of the original soil spectral. The terrain factors were chosen to corrected the inversion model for the mapping and prediction of the cadmium and lead content in soils of the study area. The main results were as follows.(1) The the pappy soil and purple soil spectral curves which were extracted from the HSI hyperspectral remote sensing data in the study area possessed similar shape, spectral absorption and reflection characteristics. A wide absorption band of the soil spectral curves was observed near 660 nm. The slopes of the spectral curves were larger in the band of 890-900 nm. The feature bands of soil spectral variables were mainly located in 466 nm,480 nm,526 nm,552 nm,580 nm,601 nm,620 nm,640 nm,660 nm,678 nm,690 nm,701 nm, 716 nm,807 nm,827 nm,862 nm and 900 nm.(2) The continuum removal spectral, reciprocal logarithm spectral, first derivative and second derivative spectral has highlighted the characteristics of the bands, and enhanced the correlation with heavy metals. The original spectrum was negatively correlated with the content of cadimium and lead and the reciprocal logarithm spectrum was positively correlated with the content of cadimium and lead, while the correlation between continuum removal, first and second derivative spectral variables and cadimium, lead contents was both positive and negative correlation. The derivative spectral variables were the most strongly correlated with the cadimium and lead content. The maximum correlation band between cadimium and the first derivative spectral variables was 561nm and they have a eextremely significant negative correlation. The maximum correlation band between lead and the second derivative spectral variables was 598nm and and they also have a eextremely significant negative correlation.(3) The modeling accuracy of partial least square regression model was better than that of the multiple stepwise regression model for the inversion of cadmium and lead content based on the five spectral variables. The modeling accuracy of derivative spectral variables were better than that of other spectral variables for the inversion of cadmium and lead content based on the two kinds of modeling methods.The optimal model for the inversion of cadmium content was the partial least square regression model based on the second derivative of spectral variables. The modeling decision coefficient of the optimal model for cadmium was 0.353 and the root mean square error was 0.046. Besides, the optimal model for the inversion of lead content was the partial least square regression model based on the first derivative of spectral variables. The modeling decision coefficient of the optimal model for lead was 0.388 and the root mean square error was 0.003.(4) In the cadmium content prediction model, the test precision of the multiple stepwise regression model based on first derivative of spectral variables was 84.30%, which was higher than other models. The test accuracy of cadmium content prediction model was more than 82%. In the lead content prediction model, the test precision of the multiple stepwise regression model based on inverse logarithms of spectral variables was 92.21%, which was the highest. Great predictions had gotten for all lead inversion models with the test precision more than 90%. According to the inversion models, the optimal models for cadmium was partial least square regression model based on the second derivative spectral, and the optimal models for lead was partial least square regression model based on the first derivative spectral.(5) After considering the topographic factors such as slop and curvature, the stability and accuracy of the corrected inversion model for cadmium and lead had been improved. The modeling decision coefficient reached 0.370 and 0.406, and the testing accuracy reached 84.72% and 91.87%, respectively. The average prediction errors were 0.0006 and 0.0001 respectively, and the trend of model prediction was low. The corrected inversion models of cadmium and lead content were used for parameter mapping. The cadmium content of the cultivated soil was 0.184 - 534 mg/kg and the lead content of the cultivated soil was 20.360 - 34.840 mg/kg. The spatial distribution of the mapping was similar to the ordinary Kriging interpolation in the middle of the study area and the spatial distribution of the mapping was lower than the ordinary Kriging interpolation in the west and southeast of the study area.