| The information of optical and SAR images can be complement for improving the urban land cover classification accuracy. However, because of the different imaging principles, optical and SAR images have different geometric and radiative characteristics. In addition, with the improvement of spatial resolution, feature extraction is increasing complex and difficult, and optical and SAR images cannot be registrated accurately, which brings challenges to information fusion. This paper focuses on the fusion strategy for urban land cover classification using high spatial resolution optical images, full polarization SAR images and x-band airborne SAR images.(1) we proposed a new strategy for urban land cover classification by fusing high spatial resolution optical images and quad-polarization SAR images. 1m RGB airborne images and 5.2×7.6m quad-polarization RADARSAT-2 images in Zhangye City, Gansu Province were used. Both of the images were acquired in July 2012.The novelty of strategy was in twofold: introduce an object-based classification method to alleviate the negative impact of inner class variability in high spatial resolution images and geometric differences between the SAR and optical image; construct a fuzzy model to fuse features with different properties and ranges. Two groups of experiments were carried out to validate the strategy. Results of first experiment indicated that the optical image shows good performance at extracting land covers with distinct spectral features such as natural vegetation, while the SAR image is better at differentiating several other land covers with similar spectral identities. For example, the introduction of polarimetric SAR features clearly improve the classification accuracy of bare soil and buildings by about 5%-10%, and also help the separation of artificial and natural vegetation to some extent. The overall classification accuracy increases from 85% to 88.18%. In addition, the second experiment not collected training samples but directly used the parameter settings of first experiment. The results of second experiment are consistent with the first experiment and the overall classification accuracy increases from 81.82% to 88.04%.Based on object-based classification and fuzzy deduction, the proposed strategy proves a promising tool in fusing SAR and optical images to extract urban land cover.(2) This research is based on 1m RGB airborne images and 1m x-band airborne SAR images of Tientsin acquired in July 2014. Through analysis on the characteristic of typical urban features, with the use of object-oriented morphological segmentation, Support Vector Machine classification and hierarchical classification, two experiments were conducted. The first experiment was based on object-oriented segmentation using SVM classification(RBF kernel function). Four groups experiments were set up for comparsion. The pixel-based classification results are shattered and mixed. On the contrary, object-oriented classification results have clear boundaries, and the overall accuracy has improved by about 5%. Classification only by optical images shows serious building and road confusion, and the classification for water, shadow, vegetation is often mixed. The introduction of SAR images greatly improves the classification accuracy of all kinds of urban features, especially road, vegetation, water and shadow. Moreover, the overall classification accuracy has increased about 10%. The second experiment introduces the hierarchical classification strategy, with the training sample n-D visualizer analysis, simpler SVM classifier(Linear), less valid data and less characteristic bands, the classification result is very close to the SVM(RBF) classification result, the overall accuracy difference is 0.47%, the classification accuracy difference between all kinds of features is within 5%. Specifically, the extraction of road, vegetation, bare land and shadow is very good, and only parts of water is divided into the shadow. |
PDF Full Text Request