High Spatial Resolution Remote Sensing Image Association Classification Method By Fusing Spatial Predicates

Accompanied by the emergence and application of high spatial resolution remote sensing images, the phenomenon of ’same material with different spectra’ and ’different materials with same spectrum’ becomes more common, it makes classification more difficult. The most common object-oriented classification method is mainly based on feature distances, that is, judging the similarity according to the distance, so as to implement the pattern classification. However, high spatial resolution images always have spectrum confusing problems, which make it hard to distinguish between each other only depending on feature distance. So we need new classification methods to solve this problem.The spatial predicate, depicting the geographic spatial information of the classification objects, shows the spatial relationships between objects in remote sensing image. The fusion of the spatial predicate and non-spatial predicate, mining spatial association rules based on spatial predicate, can establish the association between spatial and non-spatial predicate of the remote sensing image and object categories. Introducing the rules of spatial association into the classification of the remote sensing image, would have high application value and potentiality of research. We orient to the information extraction of high spatial resolution RS images, combine with object-oriented image supervised classification method, and do the research on the remote sensing image classification method by fusing spatial predicates. The main contents include three parts as follows:(1) Get characterized primitives’ spatial and non-spatial predicates. Firstly, a multi-resolution image segmentation which combines spectral features with shape features is used to divide the RS image into small primitives. Meanwhile, non-spatial characters are extracted from characterized primitives. The K-Means++clustering algorithm which has been optimized with initial cluster centers are used to carry on attribute section towards characterized primitives. Use the information gain algorithm to do dimensionality reduction for characterized primitives. Then, obtain land categories and simplified spatial information of characterized primitives. Finally, we need to combine the spatial predicates, the non-spatial predicates and land categories to the same transactional database for the subsequent spatial association rule mining. (2) Mine frequent and reliable spatial association rules. First of all, with the aid of FP-Growth algorithm, to obtain a number of rules generated in remote sensing image characteristics of primitive spatial predicate and non-spatial predicate and object classes, a derived from the transaction database support set is not less than the minimum support of the frequent pattern set. Secondly, using the method which can overcome the characteristics of mass and disorder in frequent patterns, in order to mine spatial association rules; Finally, use the method which can eliminate redundant and conflicting rules to prune spatial association rules.(3) High spatial resolution remote sensing image spatial association classification and supervised classification fusion classification. Firstly, we describe the spatial association rules extended to image classification method. Secondly, we look for the combination of spatial association classification and K-Nearest Neighbor (KNN) classification, spatial association classification and support vector machine (SVM) classification. Build the fusion of spatial association classification and these two common supervised classification, use it in remote sensing images terrain classification. Finally, we compared the experimental result accuracy of fusion classification method with traditional supervised classification, and justified the method this paper proposed.In this paper, we use high spatial resolution remote sensing images from GeoEye-1, with2meters resolution, nearby Mount Lu, as the experimental data. After that, based on the mined rules of spatial association, we carry out classification experiment. The obtained result, comparing with the result of traditional objected-oriented supervised classification method, shows that the method, introduced in this paper, avoids the problems of the classification feature set and the manual setting of classification rules. And this method, learning the knowledge of the classification of the spatial association automatically, can achieve classification results with high accuracy.