Prediction of Ni and Cd contamination in agricultural soils in Lebanon using hyperspectral imaging and artificial intelligence.

This study evaluated hyperspectral imaging (HSI)-based prediction of nickel (Ni) and cadmium (Cd) contamination in three agricultural soils from Lebanon, classified as clayey, silty, and sandy, and developed machine learning (partial least squares regression (PLSR) and support vector regression (SVR)) and deep learning (neural networks (NN) and convolutional neural networks (CNN)) models to predict soil contamination levels. The tested soil groups exhibited distinct absorption features, reflectance intensities, and spectral shapes. Among the developed prediction models, SVR and CNN achieved the highest performance. Using SVR, the prediction performances ( R 2 ) for Ni contamination were 0.850, 0.830, and 0.790 for clayey, silty, and sandy soils, respectively, while the corresponding R 2 for Cd were 0.930, 0.931, and 0.945. For the CNN models, R 2 values in silty soil reached 0.899 for Ni and 0.944 for Cd, whereas in sandy soils, they were 0.872 and 0.938, respectively. Clayey soils exhibited the highest prediction performance with the NN-based model, achieving R 2 of 0.930 for Ni and 0.950 for Cd. SVR and CNN models were validated with newly contaminated soil samples and demonstrated a good fit ( R 2 > 0.790) across all tested soils. Both SVR and CNN showed poor transferability across the different tested soil types ( R 2 ranged from 0.107 to 0.242 for SVR models and from 0.110 to 0.260 for CNN models), demonstrating limited model transferability among the tested soil groups and emphasizing the need for soil-specific or matrix-informed calibration models. However, because only three soils with multiple differing physicochemical properties were investigated and detailed mineralogical characterization was not performed, the specific contribution of individual soil characteristics to model transferability could not be isolated.