On retrieving the chromium and zinc concentrations in the arable soil by the hyperspectral reflectance based on the deep forest

Abstract

It has drawn growing attention to retrieve heavy metal concentration in naturally contaminated arable soil by its spectral reflectance. However, such a spectral reflectance is generally affected by various heavy metal elements, posing a significant challenge to ensure the inversion accuracy for a specified heavy metal concentration. The Deep Forest 2021 (DF21) algorithm shows excellent performance in the deep learning, which provides a potential method for hyperspectral inversion with high precision. This paper takes the Chromium (Cr) and Zinc (Zn) concentrations as examples, it explores the DF21 for retrieving heavy metal concentration by the spectral reflectance of the naturally contaminated arable soil. By so doing, various spectral pretreatments and Principal Component Analysis (PCA) are applied to the spectral reflectance. Subsequently, the obtained spectral data, combined with the DF21, are used to establish inversion models. Further, the performances of established models are compared to identify the optimal one for retrieving the Cr and Zn concentrations. The results show that the spectral pretreatment could potentially impact the inversion accuracy, but it exerts little or even a negative impact on the inversion performance once the PCA is applied. As a result, the DF21, together with the original spectra processed by the PCA, i.e., the ORI-PCA-DF21 model, has the optimum performance for retrieving both Zn and Cr concentrations. It is also found that the Cr concentration, which shows a relatively lower degree of heterogeneity, has higher inversion accuracy, suggesting that the spatial heterogeneity could potentially affect the performance of the ORI-PCA-DF21 model.