Parameter comparison for linear spectral unmixing in field hyperspectral sampling of rocky desertification

Abstract

Rocky desertification is one of the most serious problems in environmental deterioration, and its accurate mapping is of many implications for maintaining Earth’s sustainability. Compared to traditional field surveying approaches, remote sensing (RS), particularly hyperspectral RS, has proved to be a more efficient solution plan. Yet, hyperspectral RS also suffers from the problem of spectral mixing, since rocky desertification may correspond to various fractions of vegetation, bare soil, and exposed rock. Although linear spectral unmixing (LSU) is an effective method for resolving their ratios, different constraint conditions involving the selections of pure objects and end-member spectra may influence the result. To better overcome this basic problem, the key point of parameter comparison for LSU in field hyperspectral sampling of rocky desertification was investigated. In the typical rocky desertification areas in southwest China, an experiment of field hyperspectral RS sampling various scenarios of ground object mixing was conducted. Then, LSU was operated, by firstly classifying the digital photos of the sample plots to derive the proportion of each object. The specific LSU operations were classified into four types. The selection of end-member spectra were classified into three kinds of cases. With the 12 combination cases of the above-listed scenarios compared, we found that the results under the conditions of ANC and full constraint were better than the ASC and unconstrained conditions, and the performance for the end-member selection scenarios from case A to case C was dropping but could handle more complex situations. These inferences can supply a more solid theoretical basis for better implementing spectral unmixing in hyperspectral RS of rocky desertification.