A systematic evaluation of advanced machine learning models for nickel contamination management in soil using spectral data

Abstract

Soil nickel (Ni) contamination attributes a crucial environmental concern because its adverse effects on people health and ecosystem. Numerous studies have estimated Ni concentrations in soil, however, previous studies face several limitations, such as limited sample size and restricted spatial coverage, which impede their practical application. In this study, comprehensive and reliable dataset was utilized and 12 machine learning models were trained to predict soil Ni contamination at large-scale. After hyperparameter tuning, the light gradient-boosting machine (LGBM) method showed the optimal performance with values for area under the curve, accuracy rate, precision rate, F1 score, and recall rate metrics of 0.8024, 0.8218, 0.6818, 0.7561, and 0.7183, respectively. Accordingly, the LGBM model was employed for feature importance analysis, with the top three most sensitive bands identified within the wavelength ranges of 2214–2215 nm, 2214.5–2215.5 nm, and 2215–2216 nm, with feature importance scores of 159, 147, and 119, respectively. The results validate the effectiveness of machine learning techniques in detecting Ni concentrations in soils, which can directly inform the regulation of soil Ni levels and contribute to the promotion of soil management, crop cultivation, and disease prevention.