A near-standard soil spectral library to improve stability and transferability of hyperspectral remote sensing models for soil heavy metal prediction.

Limited sample sizes and regional differences in pollution characteristics and soil properties often obscure the characteristic bands of soil heavy metal (SHM), thereby reducing the stability, accuracy, and generalizability of prediction models. These limitations ultimately hinder the engineering application of hyperspectral remote sensing in SHM detection. To overcome this, a Near-standard Soil Spectral Library (NSSL) was innovatively developed by producing near-standard soil samples for cadmium (Cd), lead (Pb), and arsenic (As). Furthermore, the effectiveness of characteristic bands identified from NSSL was assessed through different machine learning models using both laboratory and field in-situ spectra from mining and agricultural scenarios. Additionally, the capability of NSSL as prior spectral knowledge was further examined to enhance the accuracy of prediction models and to improve model transferability. The results indicated the characteristic bands of SHM were identified as 405-501, 769-994, 1130-1142, 1196-1278, 1376-1468, 1670-1782 and 1949-2374 nm for Cd, 409-540, 1813-2060 and 2214-2354 nm for Pb, and 403-578, 1890-1898, 2053-2067, 2263-2266 and 2307-2356 nm for As. As prior spectral knowledge, NSSL significantly improved model accuracy based on field in-situ spectra under limited sample sizes, improving Rv2 values from 0.68 to 0.79 for Cd, from 0.69 to 0.78 for Pb, and from 0.72 to 0.80 for As. The performance (Rv2) of model transferability was enhanced from 0.57 to 0.79 for Cd, from 0.53 to 0.73 for Pb, and from 0.29 to 0.57 for As. These results emphasize the reliability of NSSL in supporting hyperspectral remote sensing for SHM detection and constitute a milestone in advancing the field.