Multi-scale Remote Sensing Image Segmentation Based On Granularity Theory

High spatial resolution remote sensing image (HSRI) provides both the opportunity and challenge for remote sensing application. With the spatial resolution refinement, the limitation of the traditional pixel-based image analysis method becomes obvious. Object-based image analysis technique can eliminate the'salt and pepper' effect and is quite efficient in using spatial or contextual information. It thus becomes the first choice for HSRI application recently. As a fundamental process, HSRI segmentation partitions the images into un-overlapping homogenous regions or objects. The segmentation quality has a direct influence on the latter image analysis. In HSRI, the various landscapes patterns exhibit multi-scale hierarchical and structural characteristics, which change depending on the scale of observation. Consequently, there often does not exist a single scale of segmentation that could be deemed appropriate for analysis of the entire image. Clearly, a multi-scale analysis of the image is necessary which naturally entails a segmentation technique that is capable of generating a multi-scale representation of the image data. Till now, a lot of multi-scale segmentation algorithms have been developed. However, most of them are not aiming at HSRI segmentation and have difficulty in multi-scale information transferring. In view of these limitations, this dissertation introduces the granularity synthesis and granularity stratification techniques into the segmentation process to explore an effective multi-scale segmentation approach for HSRI based on the granular theory.The dissertation starts from the multi-scale characteristics of gournd objects on HSRI, proposes a new approach for multi-scale image segmentation based on the gradular theory. Both the granularity stratification and granularity synthesis techniques are implemented in the synthesization process of multi-scale segmentation results. The involved key theories, processing algorithms and applications are researched and the major works are listed as follows:1. A regional-adaptive marker-based watershed algorithm integrating edge information is developed. The marker image is firstly extracted by a regional-adaptive threshold segmentation of the gradient image to solve the over-segmentation problem. Instead of using a fixed single threshold, a threshold image is firstly estimated. For each pixel, the threshold value is determined by the gradient distribution of the local region and the fractile value of the image histogram. As a result, the threshold values of the textured regions are relatively high and the threshold values of the spectral homogenous regions are relatively low. The extracted markers are more coincide with the inner regions of the ground objects. Then, to retain the weak object boundaries and improve the boundary location accuracy, the edge detected by the confidence-embedded method is integrated into the proposed algorithm. Both the marker image and the gradient image are rectified according to the edge information to ensure the edge pixels are labeled lastly as the object boundary pixels. 2. Two multi-scale segmentation methods are developed based on the scale space theory and the Mumford-Shah model respectively. In the first method, the image is transformed into multi-scale images by nonlinear filters to construct a scale space firstly. Then, the multi-scale images are segmented by the proposed marker-based watershed algorithm to achieve multi-scale segmentation results. Based on the scale space theory, image segmentation and edge detection can be connected naturally. For each image scale, the detected edge is integrated into the segmentation process to achieve the corresponding scale of segmentation result. In the Mumford-Shah model method, the initial homogenous elements are extracted by the proposed watershed algorithm at the beginning. Then the neighboring homogenous elements with the lowest merging prices are merged with the Mumford-Shah function value minimized. Each merging operation generates a single scale of segmentation. After many times of object merging, a series of multi-scale image segmentation results can be achieved. Here, the merging price is defined as the ratio of the region fitting error and the neighboring objects' common boundary length. In this method, the edge information can be integrated by reducing the merging probability of the common boundary that has high overlapping rate with the edge information.3. Scale selection based on supervised segmentation quality evaluation is studied. Classification samples are used as the reference data for segmentation quality assessment to find the optimal segmentation scale. Considering that the samples reflect the classification target, the selected scale will better fit the application requirements. Furthermore, because the objects appear with different inherent scale, it is necessary to select multiple optimal segmentation scales for different objects. With consideration that objects belonging to the same land cover class are usually with the same optimal scale, an optimal segmentation scale can be determined for each land cover class using the corresponding samples.4. The scale synthesis theory and methods are studied under the concept of granular theory. Both the granularity stratification and granularity synthesis techniques are researched and implemented for scale synthesis. In the granularity stratification method, the image segmentation is divided into two levels:coarse segmentation and fine-grained segmentation. The coarse segmentation partitions the image into multiple large regions according to the land cover types in advance. Then, each region is segmented with the optimal segmentation scale determined by the land cover type in the fine-grained segmentation. In the granularity synthesis method, the optimal segmentation result for each land cover type is selected as the medium granular space, then these granularity spaces are combined to realize the scale synthesis.The main innovative points are as follows:1. A regional adaptive marker-based watershed segmentation method integrating edge information is proposed. Based on the scale space theory and the Mumford-Shah model, two multi-scale image segmentation algorithms are proposed. With the corresponding edge information integrated, these methods are capable in eliminating the small undesired objects and retaining the weak object boundaries in the final segmentation result. Moreover, the proposed methods are of high efficiency for large image data.2. The scale selection technique based on the supervised segmentation evaluation is proposed. The classification samples are used as reference data, and the discrepancy measures are used to evaluate the segmentation result. Because the segmentation assessment result reflects the similarity between segmentation result and the classification samples, the selected scale can fit the application requirement well.3. The granularity theory is studied and two kinds of scale synthesis methods are developed using the granularity stratification and granularity synthesis techniques. The granularity stratification method can transfer the information of the coarse level segmentation to the fine-grained level of segmentations. The information of the coarse leve segmentation is used to direct the selection of the finer segmentation scales. The optimal scales of segmentation results are synthesized to achieve the final segmentation result. The granularity synthesis method combines different scale of segmentation results to achieve the segmentation result. It can simplify the image segmentation multi-dimensional tasks into multiple single dimensional tasks and get the final result by synthesis methods.This dissertation introduces the granular theory into the segmentation of HSRI. Experiments show the proposed methods can adaptively produce good segmentation results for images with different land cover types and meet the segmentation requirements of different classification purposes.Further research is needed on the following issues:1. To enhance the self-adaptiveness of the scale selection techniques, segmentation assessment measures should be further studied to standardize the segmentation assessment procedures.2. The difference of segmentation scales among different gound objects should be further studied. The image can be divided into more land cover types to ensure that the optimal segmentation scale of each land cover type is nearly the same.3. In the granularity stratification method, further research on how to use the different information of the coarser level (such as the degree of fragmentation) is needed.4. The optimal features and methods to be used for segmentation of different ground objects should be further studied.