URBAN FEATURES EXTRACTION FROM HIGH RESOLUTION SATELLITE IMAGE USING DEEP LEARNING BASED SEMANTIC SEGMENTATION ALGORITHMS

摘要

Automated extraction of urban features from high resolution satellite imagery is a challenging task, especially extracting buildings and roads since such man-made features exhibit high intra-class variations and low inter-class variations. The conventional image processing techniques such as grey-level thresholding, surface-based segmentation, iterative pixel classification, edge detection, based on the fuzzy set, when used for high resolution satellite image segmentation, are not able to properly retain high-level features and details such as feature boundaries. The extensive range of applications for urban feature extraction includes automated map making, urban planning, and change detection for real-estate management, land use analysis, and disaster management. For this research, a deep learning-based semantic segmentation algorithm is discussed. In this study, several semantic segmentation models using convolutional neural network (CNN) are compared to achieve optimal accuracy for extracting urban features from high resolution worldview-2 satellite imagery. Semantic segmentation of satellite and aerial images encounter-difficulties due to factors such as relief displacement of high rise buildings, shadows of tall building and trees, etc. Accurate segmentation of buildings and roads into distinct classes from high resolution images becomes challenging because such man-made features have similar reflectance patterns over the visible range of the electromagnetic spectrum. To resolve these issues, an appropriate semantic segmentation model is to be chosen that will give optimal accuracy in extracting urban features (especially buildings and roads) from satellite imagery. Experimental results show that the developed model improves the accuracy of segmentation in comparison with conventional image processing techniques. Predicted results show more than 80% of overall accuracy is achieved using the proposed algorithm. The semantic segmentation model used in this research automatically extracts urban features especially buildings and roads with optimal accuracy.