Estimating the spatial distribution of soil total arsenic in the suspected contaminated area using UAV-Borne hyperspectral imagery and deep learning

Abstract

The total arsenic (TAs) content in the soil is commonly used as an important indicator for evaluating soil pollution. However, the traditional methods for investigating TAs concentration in soil over a large area are always labor-intensive and costly. As a rapid and convenient technique, unmanned aerial vehicle (UAV) equipped with hyperspectral camera offers a promising way for estimating the distribution of TAs. In this study, we utilized UAV-borne hyperspectral data over the Daye city of China mining suspected contaminated area to establish the deep model for retrieval of TAs. Specifically, 74 soil samples were collected in situ from the study area, and their TAs contents were measured by using atomic fluorescence spectrometry(AFS). Meanwhile, use UAV captured hyperspectral imagery of the study area. We propose a novel method which deep neural networks with competitive adaptive reweighted sampling (DNN-CARS) for the estimation of soil TAs content and the spatial distribution. For two testing areas, the values of R2 are 0.90 and 0.87, and the value of RMSE are 0.33 and 0.52. Experiments demonstrated, that UAV hyperspectral imagery combined with DNN-CARS is an effective tool for the evaluation of TAs content and mapping its spatial distribution.