Estimating Soil Arsenic Content with Visible and Near-Infrared Hyperspectral Reflectance

Abstract

Soil arsenic (AS) contamination has attracted a great deal of attention because of its detrimental effects on environments and humans. AS and inorganic AS compounds have been classified as a class of carcinogens by the World Health Organization. In order to select a high-precision method for predicting the soil AS content using hyperspectral techniques, we collected 90 soil samples from six different land use types to obtain the soil AS content by chemical analysis and hyperspectral data based on an indoor hyperspectral experiment. A partial least squares regression (PLSR), a support vector regression (SVR), and a back propagation neural network (BPNN) were used to establish a relationship between the hyperspectral and the soil AS content to predict the soil AS content. In addition, the feasibility and modeling accuracy of different interval spectral resampling, different spectral pretreatment methods, feature bands, and full-band were compared and discussed to explore the best inversion method for estimating soil AS content by hyperspectral. The results show that 10 nm + second derivative (SD) + BPNN is the optimum method to predict soil AS content estimation; R 2 v is 0.846 and residual predictive deviation (RPD) is 2.536. These results can expand the representativeness and practicability of the model to a certain extent and provide a scientific basis and technical reference for soil pollution monitoring. Sustainability 2020, 12, x FOR PEER REVIEW 3 of 19 polluting enterprises in the suburb, such as chemical plants, funeral homes, medical waste treatment 95 plants, and so on. These polluting enterprises are also direct discharge areas of wastewater, waste 96 gas, and waste slag in Weinan City. In recent years, heavy metal pollution caused by city expansion 97 has drastically affects the soil ecological environment and human health. 98 2.2. Acquisition and Processing of Soil Data 99 The results of previous field investigations indicated that the soil AS content in the mining area 100 is mainly affected by human activities and shows a large difference, and the suburban soil AS 101 content is mainly affected by land use types and shows a small difference. Therefore, from 102 November to December, 2018, 90 soil samples were collected in accordance with the uniform grid 103 method with a higher sampling density in the mining area (50 m) and a lower density in the suburb 104 (200 m). The 50 soil samples were collected from the mining area and the rest form the suburb. There105 were six members in the research group with unified training on the sampling method. First, sample 106 points ware selected on the satellite map, and then the six members were divided into two groups to 107 finish the sampling job in each area. A real-time kinematic (RTK) was used to precisely locate every 108 sampling location. Figure 1 shows the position of the sampling points. The sampling was conducted 109 at a depth of 0-20 cm. Stones, plant residues, and other large debris were removed from each fresh 110 sample, which were then mixed tho...

Estimation of Soil Arsenic Content with Hyperspectral Remote Sensing

“…Soil arsenic content and spectral reflectance may often have a complex nonlinear relationship [38], so the PLSR method has poor performance in some cases. The neural network algorithm has excellent performance and efficiency, and has excellent performance in solving complex nonlinear problems [39][40][41][42], but it is prone to lack of generalization ability. Therefore, it is combined with SFLA's excellent global search ability to optimize the initial parameters of RBFNN, so as to obtain a model with better fitting ability and prediction accuracy.…”

Detecting Arsenic Contamination Using Satellite Imagery and Machine Learning

“…SD magnifies smaller differences between wavelengths and reduces some sources of random error caused by external factors through this focus on differences. MSC can correct for certain errors caused by light scattering [26].…”

“…For the purposes of this study, each of the SG, FD, and SD filters use a window length (the number of coefficients) of 17 and a polynomial order/degree of 4, as consistent with past hyperspectral research [18,26]. The predictive ability of these four noise reduction algorithms was tested by combining their processing of the spectra with the PLSR ML model.…”

“…Multiple research studies have been conducted in the past few years to explore relationships between hyperspectral data and heavy metal contamination in bare soil, including As, Chromium (Cr), Cadmium (Cd), Lead (Pb), and Zinc (Zn), using regression-based machine learning approaches on small-field scales with lab-processed soil data and hyperspectral measurements [11,12,[18][19][20][21][22][23][24][25][26][27]. This approach is advantageous because it offers a remote yet effective technique for scalable heavy metal detection.…”

“…Third, it was hypothesized that the application of spectral preprocessing, data augmentation, and dimensionality reduction would serve to increase the accuracies of the baseline models. Note that dimensionality reduction has some disagreement in literature regarding its effectiveness and necessity; while some studies's models benefit from it [21], some have found it ineffective [26].…”