Abstract:
Heavy Metals (HMs) pollution in agricultural soils has become of great concern in Lebanon. Detailed and accurate spatial soil information became essential for environmental risk assessment. Hyperspectral remote sensing for the detection of heavy metal contamination in soils and vegetation has been widely studied as an alternative technique to traditional heavy metals measurement methods such as atomic absorption spectrometry (AAS) and was found to be cost-effective, efficient, and less time-consuming. However, the potential of remote sensing data in detecting HMs contamination in soils has not been explored and tried in Lebanon. In addition, factors affecting the use of hyperspectral imaging (HSI) in soil contamination detection were not thoroughly studied. This study investigated the impact of the Lebanese agricultural soil types on HSI at different contamination levels of Cadmium (Cd) and Nickel (Ni). Clay, silty, and sandy soils were tested being the three most prevalent Lebanese soil types. Soil samples were contaminated with Cd and Ni, considering the addition of one metal at a time, at different contamination levels and detection of the contamination was performed using AAS and HSI. Data from the AAS served as ground truth to confirm the efficiency of HSI in heavy metals detection and quantification. To fulfill the aim of the study, four artificial intelligence (AI) prediction models – partial least square regression (PLSR), support vector regression (SVR), neural networks (NN), and convolutional neural networks (CNN) – were tested and compared. Nonetheless, model performance statistics showed that the SVR and CNN performed better, providing high accuracy for almost all soil types. Clayey, silty, and sandy soils achieved 85%, 83%, and 79% accuracy respectively for Ni and 93%, 93.1%, and 92.5% for Cd contamination prediction using SVR. As for the CNN regressor, the prediction of HMs in the silty soil achieved 89.9% and 94.9% for Ni and Cd respectively and sandy soils achieved up to 87.2% and 93.8% accuracy for Ni and Cd respectively. Clay soil represented higher accuracy with the NN-based model achieving 93% and 95% for Ni and Cd respectively. The developed SVR and CNN models were then validated with newly contaminated soil samples. A particularly good fit was obtained with values of R2 ranging between a minimum of 0.79 for sandy and 0.81 for the clayey soil in predicting Ni content, and between 0.82 and 0.92 for sandy and clayey soil respectively describing Cd metal for SVR-developed models. CNN models also excelled in their validation results representing a range of R2 between 0.82 and 0.841 for Ni metal, and between 0.911 and 0.94 for Cd-developed models.
The results further revealed that soil type does have a strong impact on the reflectance of the HSI, affecting the reflectance intensity and spectral shape, in terms of physical properties, and representing different absorption features, which are generally related to the specific chemical properties of each soil type. This was also highlighted and justified by validating the same SVR and CNN-developed models but this time with different soil type data. The results showed that each soil type has its specific properties and cannot predict data related to other soil types (R2 ranged between 0.107 and 0.242 for SVR models and between 0.11 and 0.26 for CNN models).
The overall results indicated that the hyperspectral imaging technique can be used for HM detection in soils, and information about Lebanese agricultural soil can be accessible with relatively few human and financial resources.