Study On Estimations Of Soil Heavy Metal Available Copper Content Based On Hyperspectral

In recent years,with the influence of human mining activities,soil heavy metal copper pollution has become increasingly serious.Copper is one of the heavy metal pollutants in soil.There is a close relationship between available copper content and soil heavy metal pollution.Hyperspectral remote sensing technology can achieve an efficient and large-scale estimation of soil available copper content without destroying soil samples.In this study,61 surface soil samples in Daye city were collected.The Fractional-Order Derivative(FOD)was used to process reflectance spectral data and constructed five estimation models after correlation analysis.The feature band selection methods were used to optimize the model.It is of great significance to soil pollution assessment and ecological environment control in mining areas.The main conclusions are as follows:(1)The prediction ability of the models based on the FOD and the first-order derivatives is compared,and the results show that the 0.7-order model performs the best effect.The FOD algorithm can capture the change of gradient inclination,remove the baseline drift and amplify the weak spectral characteristic.The FOD algorithm can overcome the insensitivity of the traditional first-order derivative to subtle changes,and improve the correlation between the available copper content of heavy metals and the spectral reflectance.(2)Based on the three feature band selection methods of Interval Random Frog(i RF),Iterative Retained Information Variable(IRIV),and Variable Combination Population Analysis(VCPA)were optimized models.The model performance of full spectrum,i RF,IRIV and VCPA with different feature bands was compared and analyzed.The results show that the VCPA algorithm can eliminate the collinearity between spectra,shorten the model running time,and effectively improve the prediction accuracy of the model.(3)Compared with five modeling methods:RFR,Light GBM,Adaboost,Catboost,and DFR.The results show that the prediction ability of DFR model is better than other four traditional models,and has excellent performance in small sample data,the R~2,RMSE and MAE of the validation set were 0.891,0.443 and 0.355,respectively.