Research On Classification Of Satellite Remote Sensing Image Based On Segmentation And Ensemble Learning

Remote sensing is a comprehensive sensing technology for detecting long-range objects through sensors,which has been widely used in many fields such as geographical conditions census,disaster prevention and flood control,urban planning and surveying and map-making,cultural heritage and ecological environment protection,natural resources exploration and so on.According to the key problems to be solved in remote sensing image processing for geographic conditions census and detection,two important technologies are studied,namely,image segmentation and image classification.Image segmentation is to form an object of characterizing the land information by merging the pixels according to similarity criteria,which is the prerequisite and basis of object-oriented image classification.Image classification is to analyze and mark the objects by the machine learning algorithm.Firstly,this thesis analyzes the background and significance of the research,the difficulties and the development trend of the future,and introduces several classic methods about remote sensing image segmentation and classification.In order to accurately evaluate the effect of image classification,this thesis compares the results from two aspects: subjective evaluation and objective analysis.In order to improve the generalization ability of classifier,this thesis puts forward an idea of Ensemble Learning(EL).Many classifiers are integrated to form a strong classifier,which can reduce the decision error rate by only using a single classifier.The main contents of this thesis are:1.There are some problems about selecting bandwidth artificially in multi-scale segmentation of remote sensing image,and generating over merge,under merge and error merging during object merging.In order to solve these problems,a new method of adaptive bandwidth is proposed in this thesis by combing image gray value and two-dimensional histogram,which not only improves the segmentation effect but also improves the processing efficiency.Then,an optimized algorithm of region merging based on RAG is proposed in this thesis,which is more reasonable during merging of image objects because of considering the spectrum information,texture and spatial location information of object,and the result can betterrly represent the information of the objects in the image.2.In order to solve the problem that it is easy to cause the error prediction of an object in the classification of image through a classifier,a new method of category marking for forecast objects by using ensemble learning to integrate multiple learning devices is proposed in this thesis.Firstly,image preprocessing is conducted,and then the typical features and thematic index of image object are extracted after image segmentation,which are normalized and compose vector space.Then,a strong classifier is produced by EL algorithm to identify and mark the forecast samples.Finally,the correctness and rationality of selective integration algorithm is verified in this thesis.The experimental results show that this algorithm can solve the salt and pepper effect of traditional classification method based pixel,reduce the error rate of single classifier and improve the classification accuracy.