DETECTION AND ASSESSMENT OF HEAVY METAL CONTAMINATION IN SOIL USING HYPERSPECTRAL IMAGING – A MACHINE LEARNING APPROACH

Abstract

The accumulation of heavy metals in the soil can adversely affect its microbial biomass and enzyme activities. Heavy metals can be transferred to plants, causing phytotoxicity. Additionally, heavy metals can be transmitted to humans through the food chain, posing severe health problems such as cardiovascular disease, kidney and brain damage, fatigue and dizziness, and lung cancer. It is essential to detect the presence of heavy metal contamination (HMC) and assess its level in the soil. Various methods, including biosensors and chemical analysis, can quantify heavy metals in soil. However, these techniques have many limitations, including high cost, time-consuming processes, laboriousness, and limited to small areas. In this work, we propose non-contact sensing hyperspectral imaging as an alternative to overcome the above challenges to detect HMC in vast areas efficiently. Clay soil samples from Jeb Jannine in west Bekaa were artificially contaminated with Cadmium (Cd) and Nickel (Ni). The contaminated samples were then subjected to scanning using a HySpex SWIR-640 hyperspectral camera, which generated a reflectance curve over a range of wavelengths from 950 to 2500 nm. This reflectance curve provides insight into how heavy metals impact soil reflectance. To process the data, machine learning models such as Partial Least Square Regression (PLSR), Support Vector Machine (SVM), Decision Trees (DT), and Deep Neural Networks (DNN) were developed to predict the level of heavy metals in the soil. The results showed that hyperspectral data is highly sensitive to the presence of HMC (Cd and Ni) in soil. The 1D convolution neural network outperformed other machine learning models tested, including PLSR, SVM, and DT, as well as deep learning models such as Long-ShortTerm Memory Neural Networks and Artificial Neural Networks in predicting the presence of the mentioned heavy metals. Specifically, the Ni and Cd data demonstrated high regression accuracies of 0.98 and 0.94, respectively. The remarkable regression accuracy achieved by the 1D Convolution neural network underscores the potential of Hyperspectral Imaging (HSI) in predicting the concentration of various heavy metals in soil using DNN algorithms. This approach could offer policymakers an effective solution for monitoring HMC and making decisions regarding environmental management.