Hyperspectral Remote Sensing Approach for Establishment of Heavy Metal Predictive Models

Hyperspectral remote sensing is widely used in environmental monitoring, which not only saves labor time for sampling and analysis, but also reduces use of chemicals for sample pretreatment. These advantages signify the concept of green chemistry and water sustainability. The demand for good quality, realtime, non-destructive and high resolution water quality information therefore has been growing in research areas of environmental monitoring. This study aims to investigate the feasibility of using hyperspectral remote sensing technique by visible-near infrared spectroradiometer (VNIR, FieldSpec HandHeld 2) for rapid water monitoring of heavy metals, followed by establishment of their prediction models. Heavy metals of lead (Pb), zinc (Zn) and copper (Cu) in the range of concentration between 100 to 2000 mg/L were selected as the target samples in this study. The influences of environmental parameters such as chlorophyll a (Chl-a) or turbidity on the spectral data were also studied. The sensitive bands for the target metals were characterized in the range from 800 nm to 1075 nm, based on the reflectance spectral data. Spectral data for developing of the quantitative predictive model was first preprocessed with first derivative and logarithm transformation, followed by establishing of the prediction model using multivariate linear regression (MLR). The models were then optimized by removing outliers and screening of suitable concentration ranges. Logarithm transformation was determined as the best method for predicting Pb and Zn, whereas raw data can be used for predicting Cu. The optimized prediction model for Cu was found to have the highest coefficient of determination (adjusted R² ) of 0.84 and least normalized root mean square error (NRMSE) of 0.12, while using the outputs of 7 sensitive bands (975, 980, 1016, 1023, 1041, 1057, 1065 nm) at range of 100-1700 mg/L. This result could be attributed to the blue color characteristic of the solution, whereas the others remain clear. In addition, the addition of environmental substances of Chl-a (0.22-0.88 μg/L) and turbidity (51-204 NTU) was found to have little intervention to the prediction models. All the information would be useful for future practical applications for identification of pollution sources of the study metals.