Wetland Detection from Multi-sources Remote Sensing Images Based on Probabilistic Latent Semantic Analysis

A novel wetland detection approach for multi-sources remote sensing images was proposed, which based on the probabilistic latent semantic analysis (pLSA). Firstly, spectral, texture, and subclass of wetland were extracted from high-resolution remote sensing image, and land surface temperature and soil moisture of wetland were derived from corresponding multispectral remote sensing image. The feature space of wetland scene was hence formed. Then, wetland scene was represented as a combination of several latent semantics using pLSA, and the feature space of the wetland scene was further described by weight vector of latent semantics. Finally, supporting vector machine (SVM) classifier was applied to detect the wetland scene. Experiments indicated that the adoption of pLSA is able to map the high-dimensional feature space of wetland to low-dimensional latent semantic space. Besides, the addition of subclass and quantitative environment features is able to characterize wetland feature space more effectively and improve the detection accuracy significantly. © 2017, Surveying and Mapping Press. All right reserved.