Detection of Heavy Metal Contamination in Agricultural Soils in Lebanon: Comparative Assessment Between Atomic Absorption and Hyperspectral Imaging

Abstract

The levels of heavy metals (HMs) in agricultural lands in Lebanon have been witnessing a noticeable increase in the past few years, due to increased anthropogenic pollution sources. HMs pose a serious threat to the environment for being non-biodegradable and persistent, accumulating thus to dangerous levels in the soil. A constant monitoring of the occurrence and levels of these contaminants in agriculture soils is thus essential. In Lebanon, a continuous environmental monitoring, including the assessment of levels of HMs in agricultural soils on a national level, is lacking. This is due in part to the high cost of analysis. Indeed, the traditional laboratory and chemical analysis methods used for the detection and quantification of HMs are costly and time consuming. Recently, hyperspectral imaging (HSI) emerged as an automated, objective, sensitive and rapid approach in the assessment of the soil contamination status due to its ability to capture and analyze a broad range of spectral information. Accordingly, the proposed research aims at assessing HMs contamination in major agricultural areas in Lebanon, and evaluate the effectiveness of using HSI in the detection of HMs in contaminated agricultural fields. Additionally, the study aimed at assessing the use of HSI and machine learning for the classification of soil based on its percent particle size distribution. Soil samples were collected from thirty-nine different locations across the country and were analyzed for Copper (Cu), Nickel (Ni), Lead (Pb), Cadmium (Cd), Chromium (Cr) and Zinc (Zn) using Atomic Absorption Spectrophotometry (AAS). Sources of HMs contamination were assessed and the extent of soil contamination and potential ecological risk were determined using the geo accumulation index and the ecological risk factor. The collected soil samples were also scanned using the Hyspex SWIR-384 to study the soil spectral behavior, and developed machine learning (ML) algorithms (partial least square regression; PLSR, and Support Vector Regression; SVR) were applied to detect and quantify the HMs present in the soil. In addition, the performance of four ML algorithms (support vector machine SVM, random forest RF, linear discriminant analysis LDA and neural network NN) in accurately defining different soil types based on their distinct spectral characteristics and features was evaluated. The findings from the study revealed alarming soil contamination with Ni and Cd in different agricultural areas in Lebanon. The HM concentrations exceeded to EU permissible limits in soils in 30% and 18% of the tested samples for Ni and Cd respectively. Zinc and copper were both below the EU permissible limits. However, Zn exceeded the WHO target limits in all tested samples, while Cu exceeded the target WHO limits in some locations. HM values above WHO target limits indicate soil pollution without the need for immediate intervention. No soil pollution was detected in the case of Pb and Cr which were below both the EU permissible limits and WHO target values. The calculated I geo index (I geo) revealed moderate to heavy Ni contamination (I geo values between 2 and 4), and heavy Cd contamination (I geo>4 for most of the samples) of the tested soils. Moreover, the contamination factor index (CF) was the highest for Cd showing considerable to very high contamination (CF exceeding 6 in several locations). The ecological risk factor showed very high to extreme contamination for most of the Northern locations, specifically for the case of Ni and Zn. HSI scans showed that the soil type influences the soil reflectance, where the clayey content and organic matter showed to significantly decrease light reflectance. Mostly, water absorption bands (expressed at 1400 nm and 1900 nm) and adsorption of HMs on iron oxides and organic matter (bands ranging between 2200-2300 nm) were observed. Furthermore, HSI was not successful in the quantification of HMs in agricultural soils. Both PLSR and SVR models exhibited low accuracy (R2 < 0.35) in predicting HM concentrations in the tested soil samples, namely due to the relatively low concentrations measured in these soils. However, HSI and ML were successfully used in soil classification, with the Neural Network model showing the best accuracy in predicting the soil type (R2=0.805). In conclusion, the study sheds light on the contamination levels of HMs in the Lebanese agricultural soils and the need to address sources of contamination. In addition, the study highlights the potential of using HSI in soil classification while demonstrating the limitation of this technique in determining HM content in soils at relatively low contamination levels.