Extensive investigation and monitoring of lead (Pb) content of soil is significant for ensuring hazard-free agricultural production, protecting human health, and ecosystem security, especially in a mining area. One temporal period of a hyperspectral image is usually used to estimate the spatial distribution of Pb and other heavy metals, but hyperspectral images are usually difficult to obtain. Multispectral remote-sensing images are more accessible than hyperspectral images. In this study, a deep learning-based model using 3D convolution is proposed to estimate the Pb content from the constructed multi-phase, multispectral remote-sensing images. Multi-phase multispectral remote-sensing images were stacked to generate a data set with more spectral bands to reduce the atmospheric absorptive aerosol effect. At the same time, a neural network based on 3D convolution (3D-ConvNet) was proposed to estimate Pb content based on the constructed data set. Compared with partial least-squares regression (PLSR), random forest regression (RFR), support vector machine regression (SVMR), and gradient-boosting regression (GBR), experimental results showed the proposed 3D-ConvNet has obvious superiority and generates more accurate estimation results, with the prediction dataset coefficient of determination (R2) and the mean normalized bias (MNB) values being 0.90 and 2.63%, respectively. Therefore, it is possible to effectively estimate heavy metal content from multi-phase, multispectral remote-sensing images, and this study provides a new approach to heavy metal pollution monitoring.
Read full abstract