Abstract
Remote sensing images are primary data sources for land use classification. High spatial resolution images enable more accurate analysis and identification of land cover types. However, a higher spatial resolution also brings new challenges to the existing classification methods. In the low-level feature spaces of remote sensing images, it is difficult to improve classification performance by modifying classifiers. Probabilistic topic models can connect low-level features and high-level semantics of remote sensing images. Latent Dirichlet allocation (LDA) models are representatives of probabilistic topic models. However, at present, probabilistic topic models are mainly adopted for scene classification and image retrieval in remote sensing image analysis only. In this study, multiscale segmentation was employed to construct bag-of-words (BoW) representations of high-resolution images. The segmented patches were then utilized as “image documents.” A structural topic model was used with an LDA model to import spatial information from the image documents at two levels: topical prevalence and topical content in the form of covariates. In this way, latent topic features in image documents can be more accurately deduced. The proposed method showed more satisfactory classification performance than standard LDA models and demonstrated a certain degree of robustness against the changes in the segmentation scale. Acknowledgement for the data support from “Yangtze River Delta Science Data Center, National Earth System Science Data Center, National Science & Technology Infrastructure of China ( http://nnu.geodata.cn:8008 )”.
Highlights
Automatic land use or land cover classification using remote sensing images has been receiving much attention from researchers and is expected to be studied extensively in the future
To more effectively use Probabilistic topic models (PTMs) to extract latent topic features from remote sensing images and thereby to realize pixel-level land use classification, this study proposed a framework consisting of three tasks: 1) BoW representations of highresolution remote sensing images based on multiresolution segmentation; 2) latent topic feature extraction based on a structural topic model (STM); 3) supervised classification of image documents (Fig. 1)
During land use classification based on remote sensing images, because objects of the same type may exhibit different spectral features and objects of different types may have similar spectral features in the low-level feature spaces, it is difficult to improve the classification performance by directly modifying the classifiers
Summary
Automatic land use or land cover classification using remote sensing images has been receiving much attention from researchers and is expected to be studied extensively in the future. Due to the rapid development of electronics and information technology as well as sensor technology, high spatial resolution images are more available to researchers and practitioners. It becomes necessary to improve existing automatic classification methods based on remote sensing images. Because of higher spatial resolution, there are greater spectral differences between the pixels of the same object type whereas the objects of different types may have similar spectral features. These lead to new challenges in the conventional classification methods [1]
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
