Estimating soil heavy metal from hyperspectral remote sensing images base on fractional order derivative

Hyperspectral imaging technology has the unique potential for the low-cost, large-scale, and rapid monitoring of soil heavy metals. For hyperspectral images, the number of soil image elements differs greatly from the number of soil samples, so the problem of small samples is prominent. In this paper, a soil heavy metal estimation method based on Fractional-Order Derivative (FOD) for hyperspectral images is proposed. First, the neighboring pixels of soil samples were extracted to expand the samples and increase the spectral variability. Second, FOD was used to highlight the spectral features. Then, the bands were selected by Competitive Adaptive Reweighted Sampling (CARS), and partial least squares (PLSR) was used to construct the model. Seventy-two soil samples and aerial hyperspectral images obtained from the Huangshan South mine in Hami, Xinjiang were used to estimate three heavy metals, namely, lead (Pb), zinc (Zn), and nickel (Ni). After sample expansion, the estimation accuracy of the test set was improved for three heavy metals, the test set R2 improved from 0.6128 to 0.7974 for Pb, from 0.8178 to 0.8690 for Zn, and from 0.6969 to 0.8303 for Ni, while the R2 of the training set was above 0.8. The accuracy of estimation model for three heavy metals with the best fractional-order differentiation was better than that using integer-order differentiation. CARS+PLSR obtained higher estimation accuracy than the modeling approaches of GA+PLSR and CC+PLSR. The estimation accuracies R2 were 0.7974, 0.8690, and 0.8303 for Pb, Zn, and Ni, respectively. Sample expansion alleviated the overfitting phenomenon and improved the estimation accuracy. The FOD of the optimal order could effectively enhance the spectral features and improve the estimation accuracy. CARS was more accurate than CC and GA. © 2023 Science Press. All rights reserved.