The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLVIII-3/W2-2022
https://doi.org/10.5194/isprs-archives-XLVIII-3-W2-2022-23-2022
https://doi.org/10.5194/isprs-archives-XLVIII-3-W2-2022-23-2022
27 Oct 2022
 | 27 Oct 2022

ASYMMETRIC FUZZY CLASSIFICATION NETWORKS FOR CONSTRUCTION LAND DETECTION IN HIGH RESOLUTION REMOTE SENSING IMAGES

R. Fang, Z. Wu, and X. Song

Keywords: Construction land detection, urban expansion, scene classification, neural network, remote sensing image, urban dynamic cognition

Abstract. Urbanization is an essential phase of a nation's economic development. A very effective way to examine urban growth is to look at how impervious surface changes over time, however impervious surface can only show the current situation in terms of urban development. Compared to the existing methods, the use of construction bare land to monitor urban growth has the following benefits over currently used techniques. First, it is possible to track the progress of structures being constructed as part of urban expansion. The second is to assess the city's development intensity and identify the inward expansion. Therefore, the detection of construction bare land is of great significance for the development of a more sophisticated urban dynamic perception technology. This paper proposes an asymmetric fuzzy classification network (AFCNet) for detection of construction bare land scenes. The generation of fuzzy sample sets, the backbone network, and the proposed fuzzy classification module make up the method's three primary components. The deep features of the scene are then extracted using the fuzzy classification network and converted into ambiguity. Finally, the ambiguity is converted into predicted probability using the fuzzy weight vector. The fuzzy sample set is generated to introduce more prior information into the network. High-level features are extracted using the backbone network. Fuzzy classification methods based on spectral features are used to improve the performance of scene classification. The results demonstrate that the OA of our method is higher than all other comparison methods.