Prediction Method for Soil Heavy Metal Content Based on Covariates over Proximally Sensed Spectra

Abstract:

Sampling, analyzing, and assessing soil heavy metal pollution requires significant manpower and resources. Access to easily obtainable environmental covariate information is crucial for enhancing the efficiency of soil heavy metal pollution monitoring. The spectra from proximal soil sensing are a comprehensive response of soil properties, and they have great potential to reveal heavy metal concentrations in soil. Near-infrared spectral characteristics of heavy metals in the surface soil of 109 samples were analyzed, and spectral information closely linked to soil Ni was extracted. This data was then utilized as auxiliary information to develop a co-Kriging model. Subsequently, co-Kriging models were constructed by using soil mechanical composition, and its combination with spectral information as auxiliary variables to compare the accuracy of spatial prediction mapping of Ni concentrations in the soil. The results indicated that the model incorporating silt concentrations in addition to the absorbance at 2 380 nm as auxiliary variables outperformed the model by using only silt concentrations. The cross-validated coefficient of determination R2CV was increased from 0.49 to 0.68, while the cross-validated root mean squared error (RMSECV) was decreased from 11.3 mg/kg to 9.5 mg/kg. These findings suggested that NIR spectra, as readily accessible auxiliary information, can be used with soil mechanical composition to develop spatial prediction models and further enhance the precision of regional soil heavy metal surveys. The research result can offer a cost-effective solution for the spatial prediction of heavy metals in soil.