Extracting Residential Information In Mountainous Areas By A Combined Use Of Semantic Constraints And Object-Oriented Method

The spatial distribution information of residential land in mountainous areas not only has strong spatial relationship with the rural population change, economic development and the relevant planning in mountainous areas, but also is an important basic data for mountain disaster emergency response and assessment. Therefore, timely grasping of the spatial distribution information of mountainous residential land has important significance to rural development and construction, disaster prevention and disaster relief in mountainous areas. However, this distribution information is difficult to obtain, due to the vastly territory of our country, and the complexity of the mountainous terrain. What’s more, the distribution of residential land is affected by many factors, and the rural residential land are discrete in mountainous areas. Fortunately, the development of remote sensing technology, especially the development of high resolution remote sensing technology, provides the possibility for obtaining the distribution information of residential land in mountainous areas rapidly.Object-oriented image analysis (OBIA) has become one of the main research methods of extracting information from high resolution remote sensing images. This method utilizes homogenous objects as the basic unit for information extraction, which is totally different from those method using single pixels. Such processes act as human brain when doing analysis. However, the OBIA method still has some drawback, such as determination of segmentation scale, feature selection and other issues. Coupled with the complexity of the high resolution remote sensing image itself and the ground object itself, the extraction results from the OBIA may include a large number of misclassified objects, so that the results cannot meet the needs of practical usage, especially when processing images covering a wide range. To solve this problem, this thesis proposes to add some geological expert knowledge(e.g., the spatial distribution law of residential land and the logic connection information with other features) as the semantic constraints to further refine the classification results of residential land. For testing purpose, the GF-1 data are utilized to carry out mountainous residential land extraction experiments. The specific research work mainly includes the following contents:(1) Extraction of residential land in mountainous areas based on the object-oriented method. Firstly, the basic principles of the OBIA are expounded. Then, for the selection of the optimal segmentation scale, we analyze the image segmentation scale by taken the rates of change of local variance (ROC-LV) as an index, and use it as a reference for the selection of the optimal segmentation scale. The residential land in mountainous areas are influenced by many factors, so that the features of the residential land are difficult to be unified. Therefore, we do a system analysis on its spectral features, geometrical features, and texture features, and then compare all these features with those of other typical objects. The experimental results show that the texture features are relatively uniform, which can be used as the main feature type for extracting residential information in mountainous areas. On the basis of this, a GF-1 data covering Tagong, Kangding county area is selected to extract the residential land. Moreover, we also analyzed the misclassification in the results.(2) Extraction of residential land in mountainous areas based on semantic constraints. In the particular geographical environment of mountain areas, there is a clear distribution pattern of residential land. Therefore, firstly, we analyze the distribution pattern of residential land in mountainous areas, then the pattern is used to establish the semantic constraints, including valley lines, terrain slope, and snow lines. We also expound in detail the factors that need to be considered when establishing semantic constraints. Finally, the proposed method based on semantic constraints is applied to the residential land extraction experiment in Tagong, Kangding County area. The accuracy is enhanced by 1.2 times as compared with the results by the conventional object-oriented method. The experimental results show that the proposed method can improve effectively the extraction reliability of the residential land in mountainous areas.(3) Extracting residential land in mountainous areas with large extent. The expansion of the research field scale will lead to a reduction of detail information. Another aspect is that the increase in the range of research area will bring the problem that the local rules cannot be applied to other regions, which makes it more difficult to extract the information from the high resolution remote sensing images in a large area. Considering of this problem, we propose to use the features of the object itself and the global features to extract information, in which the features between objects should be satisfied with the overall situation. Actually, the semantic constraints are used as geological expert knowledge, and is taken as the global feature. Then the proposed method is used to extract residential land in mountainous areas with very large extent. In this part, Shimian Country, Yaan City is taken as the research area for the experiment, which has an area of about 3000 km2.6 scenes of GF-1 images is spliced and used for data processing. From the analysis, the rate of reliable classification is 82.06%, and the leakage rate is 6.02%. The experimental results show that the method act better in the experiment of extracting residential land in mountainous areas under very large extent, and can effectively avoid the misclassification caused by many factors, such as the snows and clouds on mountains, the bare land on hillside, and the beach land in some regional areas. It is worth noting that the method also provides a reference for extracting other information from high-resolution remote sensing images covering large areas. In addition, we also analyzed the impacts to information extraction in large areas from image mosaic, DEM precision, and semantic constraints, when considering that they will become more complex.Theoretical and experimental studies show that, the proposed method of combining semantic constraints and object-oriented method can effectively extract the residential information in complex mountainous areas, exclude most unreasonable misclassified objects, make the extraction results consistent with the actual circumstances, and provide a reference for the application of high-resolution remote sensing images in practical producing procedure. In addition, the experiments also show that it can help to improve the automatic interpretation accuracy by extracting of a variety of image features, including multi-source data, combining with GIS, adding geological expert knowledge, optimizing the combination of these features, and complementing to each other among different features.