Indicator Spectral Bands and Logistic Models for Detecting Diesel and Gasoline Polluted Soils Based on Close-Range Hyperspectral Image Data

Abstract

In this research, we derived indicator spectral bands and classification models for detecting diesel or gasoline pollution in soil using a near-and shortwave-infrared (NIR-SWIR) hyperspectral camera under a close range and laboratory condition. The soils samples were collected from temperate climate soil with spectral characteristics manifested by secondary minerals. The hyperspectral images show that the diesel and gasoline polluted soil samples have distinctive spectral differences from clean soil. Different from moisture soil, the spectral absorption features of petroleum hydrocarbons (PHCs) are preserved with an increase in gravimetric content. The more PHCs contents, the stronger the depths at the spectral absorption features. In diesel polluted soils, the absorption features were observed in various content levels. However, we found a detection limit for gasoline content in soil, because the absorption features by PHCs disappeared at 8 wt.%. To derive the indicator bands, the images were classified by the random forest algorithm (RF) with an accuracy and kappa coefficient of 94.3% and 0.92 using three groups of bands corresponding to ferric ion, C-H stretch/bending, and BTEX C-H absorptions. The detection models derived from a logistic regression achieved an overall accuracy of 91.82%. The field test of the models on unprocessed soils achieved an accuracy of 83.36%. Because of their simple forms, the logistic detection models can be transferred to remote sensing applications of soil PHCs pollution under a close-range condition such as drone-based projects.

Topics & Concepts