Research On Automatic Labeling Method Of Remote Sensing Images Based On Multi-feature Fusion Latent Dirichlet Allocation

With the rapid development of remote sensing-related technology,labeling remote sensing images has become an important part of processing remote sensing images.But the accuracy of manual labeling will be affected by the differences in the subjective judgment of the labelers,and the tasks are heavy and the efficiency is not high.Therefore,the research of remote sensing image automatic labeling method has important significance.This paper proposes an automatic labeling method for remote sensing images based on multi-feature weighted fusion topic models,and verifies the effectiveness of the method through multiple sets of comparative experiments,which has good practical significance.This paper firstly divides the image into several non-overlapping regions,extracts color feature,texture feature,and spatial feature of each region,and combines the weighted features to obtain words that can accurately represent the stable features of each region,and then use the optimized clustering algorithm generates accurate visual words,and calculates the word frequency distribution according to the Euclidean distance between the words in each region and the visual words,and finally uses the LDA probabilistic topic model for training to classify remote sensing images.The specific research work includes the following two aspects:(1)A single feature cannot fully express the feature information of a remote sensing image,and splicing directly simple multi-feature cannot reflect the matching degree between each dimension feature and the remote sensing image.This paper adopts a multi-feature weighted fusion method,calculates the weight matrix according to the standard deviation between multiple features,realizes multi-feature weighted fusion,and makes full use of the extracted multi-feature vectors to obtain a better evaluation of each area.The fusion feature is used as a regional word,and a single feature,serial mode is combined with multiple features,and a weight mode is combined with multiple features as a regional document word.Comparative experiments have proved the necessity of multi-feature fusion and the superiority of weighted fusion.(2)In the process of generating visual words which use the traditional clustering algorithm,there are deficiencies in the processing of edge outliers and the selection of initial cluster centers.Therefore,this paper proposes an optimized clustering algorithm based on the density of feature points.According to the density of feature points and the Euclidean distance from the known initial cluster centers,the selection of the initial cluster centers is optimized.And in the iterative process,the elimination of edge isolated points is improved according to the deviation degree,and a more representative and stable visual word of the divided area is obtained.And through the visual words generated by traditional clustering algorithm and optimized clustering algorithm,the stability and accuracy of the visual words which are generated by the optimized clustering algorithm are proved.In order to evaluate the automatic labeling method proposed better in this paper,the traditional classifier is selected for automatic labeling on the same data set.The experimental results show that the overall labeling accuracy of the automatic labeling method of the multifeature fusion topic model proposed in this paper can reach 0.8595,which is 0.0293 higher than the overall labeling accuracy of the SVM classifier,which has relatively better labeling results in traditional classifiers.The method in this paper has a good effect of automatic labeling of remote sensing images.