Quantitative Detection and Long-Term Monitoring of Settle Dust Using Semisupervised Learning for Spectral Data

Classification of spectral data has recently drawn more and more attention in environmental and geoscience applications. In the past decade, this attention has been translated into an interest in employing unmixing techniques to retrieve accurate quantitative information suppressed in spectral data. The main task in real applications is to detect potential information regarding the physical and chemical nature of ground targets in different spectral data sources (point field and laboratory spectroscopy, hyperspectral imagery, etc.). Recently, semisupervised classification techniques have been proposed for spectral data by combining ground-truth and laboratory measured spectral signatures and advanced signal processing algorithms based on posterior probability support vector machine and Dempster-Shafer evidence theory. In this paper, the sensitivity of this combined classification method to extract and identify a small amount of settle dust over green vegetation canopy using field spectral data is examined and reported. The results are compared with the performance of selected semisupervised unmixing classification techniques.