Prediction of lead in agricultural soils: An integrated approach using machine learning, terrain attributes and reflectance spectra

Abstract

The study aimed to predict lead (Pb) concentrations in agricultural soils by analyzing visible, near-infrared, and shortwave infrared (VNIR-SWIR) spectra and terrain attributes under three distinct approaches. Approach 1 relied solely on the VNIR-SWIR data, with the cubist_FD (cubist_first derivative) modeling technique as the most effective, achieving a coefficient of determination (R2) of 0.63, a concordance correlation coefficient (CCC) of 0.51, a mean absolute error (MAE) of 6.87 mg kg-1, and a root mean square error (RMSE) of 8.66 mg kg-1. Approach 2 combined VNIR-SWIR spectra and terrain attributes, where SNV (standard normal variate) preprocessing technique and extreme gradient boosting (EGB) yielded superior results with R2 of 0.75, CCC of 0.65, MAE of 5.48 mg kg-1, and minimal RMSE of 7.34 mg kg-1. Approach 3 solely employed terrain attributes, and the combination with cubist modeling demonstrated the best prediction results, with R2 of 0.75, CCC of 0.66, MAE of 6.18 mg kg-1, and RMSE of 7.71 mg kg-1. The cumulative assessment identified that the fusion of terrain properties, SNV-transformed VNIR-SWIR spectra, and EGB as the optimal method for estimating Pb concentrations in agricultural soil yielded the highest R2 value and minimal error. Comparatively, Gaussian process regression (GPR), artificial neural network (ANN), and support vector machine (SVM) techniques achieved high R2 values in approaches 2 and 3 but also exhibited higher estimation errors. In conclusion, the study underscores the significance of using relevant auxiliary datasets and appropriate modeling techniques for accurate Pb concentration predictions with minimal error in agricultural soils. The findings contribute valuable insights for developing successful soil management strategies based on predictive modeling.