آشکارسازی ساختمان‌‌های با پوشش خاص در تصاویر فراطیفی با استفاده از الگوریتم هرمی مبتنی بر نشانه

Hyperspectral remote sensing technology has witnessed remarkable progress in the last two decades. One of the analyzes performed on the hyperspectral images is target detection. In this research, the detection of roofs with special cover has been done as a target in an urban environment. Simultaneously with the growth of urbanization and the development of urban areas, the need of managers and planners for very accurate maps of urban areas has increased significantly. Since an urban environment has complex characteristics in terms of physical, geometrical and elements used in buildings, hyperspectral data effectively help to identify, extract and produce a map of the constituent elements of an urban environment. Regarding the spectral detection of the target, continuous and numerous researches have been carried out since the last two decades. According to the studies carried out, until now, the hierarchical algorithm has achieved the best results in comparison with other algorithms for extracting spatial information in hyperspectral images, Therefore, in this research, it is tried to reveal buildings with special cover in hyperspectral images by presenting a new and accurate method.
Material and methods: The image data of the CASI sensor has been used to carry out this research. The images processed in this research include images with 32 spectral bands and a resolution of 2 meters, which were taken in May 2001 from the urban area of ​​Toulouse located in the south of France. In the proposed method, two classification algorithms of multilayer perceptron neural network (MLP) and support vector machine (SVM) are implemented on the hyperspectral image. Then, the map resulting from the combination of the two mentioned algorithms is used to select the marker for the marker-based hierarchical segmentation algorithm. Finally, with the help of the majority vote decision rule, the marker-based hierarchical segmentation map is combined with the map resulting from the integration of MLP and SVM classifications.
Results and discussion: In this research, Gaussian radial basis kernel was used to implement the SVM algorithm. The values ​​of two parameters, penalty (C) and width of Gaussian function () were determined in SVM algorithm with the help of cross validation technique. The MLP classification algorithm was implemented with 3 hidden layers that include 5, 6 and 8 neurons and its evaluation was done with 500 repetitions and to select markers, the analysis of the labeling of connected components was done based on 8 neighborhood pixels on the map resulting from the combination of MLP and SVM. Based on the obtained results, the map obtained from the proposed method includes uniform regions and has more interconnected structures to reveal buildings, which shows the importance of using spatial information along with spectral information.
Conclusion: In this research, the strategy of using spatial information along with spectral information to improve target detection in the analysis of hyperspectral images was examined. For this purpose, the spectral-spatial marker-based hierarchical algorithm, which is used in the image classification process, was used to reveal the roofs of the buildings. In the proposed method, two classification maps were used in the selection of markers and the decision rule of the majority vote in the case of the initial hierarchical segmentation algorithm. In the combination of MLP and SVM classification maps, conditional probability and selection of the highest probability of each pixel belonging to a class are used in the selection of markers and majority vote decision rule.
Material and methods: The image data of the CASI sensor has been used to carry out this research. The images processed in this research include images with 32 spectral bands and a resolution of 2 meters, which were taken in May 2001 from the urban area of ​​Toulouse located in the south of France. In the proposed method, two classification algorithms of multilayer perceptron neural network (MLP) and support vector machine (SVM) are implemented on the hyperspectral image. Then, the map resulting from the combination of the two mentioned algorithms is used to select the marker for the marker-based hierarchical segmentation algorithm. Finally, with the help of the majority vote decision rule, the marker-based hierarchical segmentation map is combined with the map resulting from the integration of MLP and SVM classifications.
Results and discussion: In this research, Gaussian radial basis kernel was used to implement the SVM algorithm. The values ​​of two parameters, penalty (C) and width of Gaussian function () were determined in SVM algorithm with the help of cross validation technique. The MLP classification algorithm was implemented with 3 hidden layers that include 5, 6 and 8 neurons and its evaluation was done with 500 repetitions and to select markers, the analysis of the labeling of connected components was done based on 8 neighborhood pixels on the map resulting from the combination of MLP and SVM. Based on the obtained results, the map obtained from the proposed method includes uniform regions and has more interconnected structures to reveal buildings, which shows the importance of using spatial information along with spectral information.
Conclusion: In this research, the strategy of using spatial information along with spectral information to improve target detection in the analysis of hyperspectral images was examined. For this purpose, the spectral-spatial marker-based hierarchical algorithm, which is used in the image classification process, was used to reveal the roofs of the buildings. In the proposed method, two classification maps were used in the selection of markers and the decision rule of the majority vote in the case of the initial hierarchical segmentation algorithm. In the combination of MLP and SVM classification maps, conditional probability and selection of the highest probability of each pixel belonging to a class are used in the selection of markers and majority vote decision rule.