Detection and Quantification of Nickel and Cadmium in Clay Agricultural Soil Using Hyperspectral Imaging and Artificial Intelligence-A Case Study in Lebanon

Abstract

Heavy metal contamination in agricultural soils poses serious environmental and health problems. Intensive efforts are employed to improve existing quantification methods of heavy metals in contaminated environments. Conventional detection techniques are time-consuming, tedious, and costly. The application of hyperspectral remote sensing in this field is possible and promising as a fast, nondestructive, and reliable detection technique. However, factors impacting the efficiency of image acquisition in detecting and quantifying heavy metals in agricultural soils were not thoroughly studied. This study proposes to assess the use of hyperspectral imaging and artificial intelligence for the detection of nickel (Ni) and cadmium (Cd) in agricultural clay soil collected from the Bekaa Valley, a major agricultural area in Lebanon, under different contamination levels and soil moisture content. The novelty of this study relies on the incorporation of HSI and AI for environmental monitoring in Lebanon.

Soil samples were contaminated with Ni and Cd individually, with concentrations ranging from 150 mg/kg to 4000 mg/kg and 2.5 mg/kg to 4000 mg/kg, respectively. The moisture content of raw and contaminated soil was varied from 5% to 75% based on soil water thresholds. Hyperspectral imaging was used to detect Ni and Cd contamination in the soil at different contamination and moisture content levels. The spectral curves showed an inverse correlation between Ni and Cd concentration and spectral reflectance. Based on continuum removal, Ni presence was well expressed near 2190 nm with a Pearson correlation factor of -0.79 and that of Cd near 1700 and 1750 nm with a correlation factor of -0.88 and -0.95 respectively. In addition, spectral changes due to the variation in soil moisture content were detected near 1400 and 1900 nm with a Pearson correlation of -0.88 and -0.81, respectively.

Machine learning and deep learning algorithms were used to develop univariate and multioutput models to predict the concentration of Ni and Cd in contaminated soil and assess the effect of soil moisture content on metals quantification. The models were constructed using partial least square regression (PLSR), support vector regression (SVR) and random forest regression (RFR) machine learning algorithms. In addition, artificial neural networks (ANN), convolution neural networks (CNN) and long short-term memory (LSTM) were used as deep learning algorithms. RFR had the highest prediction accuracy of 0.85 and 0.95 for Ni and Cd respectively. The prediction and validation results using other models were favorable except for PLSR due to data nonlinearity. In addition, the results showed that Cd was more predictable than Ni. Furthermore, to assess the effect of soil moisture content on heavy metal detection, multioutput heavy metal-moisture content machine and deep learning models were developed. The overall accuracy of all Ni and Cd model algorithms has decreased with the introduction of moisture to the model along with elevated RMSE values. With respect to RFR, the accuracy of Ni prediction was unchanged following water addition, whereas that of Cd has only slightly decreased from 0.95 for univariate to 0.90 for multioutput.

The results show the potential of using HSI and AI as a reliable and cost-effective approach for heavy metal pollution assessment in contaminated soils. The findings from this study may involve further investigations to examine its ex-situ applicability considering actual field conditions.