Research Of Remote Sensing Image Classification Algorithm Based On Superpixel Region Merging

Remote sensing image classification is currently one of the research priorities and the hot topics in pattern recognition and intelligent technology. No matter in the military or in the civilian, this research has favourable application value and important significance. This thesis makes a deep analysis based on the existing classification techniques, and summarizes these methods’advantages and disadvantages. By observing a large number of remote sensing images, some vital features of the ground objects are obtained. According to these features a novel algorithm to overcome the shortcomings of the current methods is proposed. The main achievements and contents are as follows:(1) The definition of ground objects is given and classified into six types such as buildings, shadows, barelands, rivers, vegetation and road. They have the following important characteristics respectively. The gray level of river is generally low, nearly dark green, and it has flat color distribution and widespread connectivity. Shadow’s gray level is rather low, with unsure texture information which depends on the ground object it covers. The characteristic of the building is of the most unobvious, with varied colors and rich texture information. Road has obvious shape features and plain grey level. The vegetation area looks darker than river, but it has richer texture information. The bareland has the obvious color features, and often there are a large number of connected areas. Color, shape, and texture features are used to classify the image in this thesis, and outstanding classification accuracy is achieved finally.(2) A set of effective classification algorithm framework for the remote sensing image is put forward. First, the SLIC algorithm is used to segment the image, and then the small and meaningless regions produced by SLIC are merged into their adjacent regions with the maximum similarity. This step is followed by color feature extraction from new superpixels, founding the CRF model to calculate every superpixel’s label. After that, the CRF model will merge every two superpixels which are adjacent and samely labeled to finish the image segmetation.(3) The LDA algorithm for feature dimension reduction is introduced in the classification process. It can make the largest separability between the features. After dimensions reducing, the color, shape, texture features are extracted as the characteristic vector of SVM classifier to classify every ground object successfully.(4) A software platform based on Google Earth and Matlab is designed in this thesis. It can automatically start Google Earth software, captureing a given-size image, cutting the image into several small parts to classify respectively, and joins them together to form the final result. Finally, experiments are performed to prove that the platform is effective and the algorithm is applicable.