Remote Sensing Image Knowledge Mining Based On Change Detection For Specific Region

The development of satellite platform and sensor technology provide maturingplatforms and data base for military and civil applications widely. The graduallyimprovement of data spatial resolution, spectral resolution and time resolution providesmore adequate information support for related studies. The dynamic monitoring of keyregions belongs to the domain of change detection in remote sensing. There are manyclassical algorithms based on analyzing two-temporal images. But there are a lot ofproblems unsolved such as fully historical data mining and intelligent changingdetection with target-oriented methods. In need of dynamic monitoring for someappointed harbor and ships, this paper focuses on studying the data mining methods andknowledge representation strategies using multi-temporal or hyper-temporal historicalremote sensing images.The main research works are as follow：First, a hyper-temporal historical data mining method is presented for harbor’schange detection and targets recognition. According to the characteristics of interestedtargets, the process is divided into two steps: static background analysis and dynamictargets clustering，with different data mining algorithms and knowledge organizations.In the case of static background analysis, three approaches are designed as medianmethod, average method and nearest-time method. The background’s knowledge isorganized through statistical characteristics and structural characteristics by data mining.In the case of dynamic targets clustering, the ships’ pattern information is analyzed byfuzzy clustering methods based on a set of feature vector.Second, a real-time process is designed and realized depended on knowledge fromthe hyper-temporal historical data mining. The key steps are: rapid segmentation for seaand bank based on the static background mask; rapid targets ROI extraction based onthe structural features of the harbor, and complete the ship regions based on scan linesmethod. In order to distinguish ships those stand side-by-side, an algorithm is proposedby tracing region contour for adhesion targets partition; based on the clusteringknowledge and the typical feature vector, the fuzzy C-means clustering methods isapplied in ship recognition.Third，the whole dynamic monitoring flow for ship targets is designed andimplemented. Include: ship detection and pseudo-target elimination; ship targetsclustering; collection the changes of the classes and amount of targets; abnormitydetection and alarm; and update knowledge base on the analysis results throughhuman-computer interaction according to the data mining process frame.We use commercial remote sensing data to simulate some dynamic changes. The preliminary experimental results verify the validity of above methods.