Scene Classification Based On The BOVW And Doc2Vec Models For High-spatial Resolution Remote Sensing Imagery

Owing to the rich spectrum,texture and structural information,high spatial resolution(HSR)remote sensing image are widely used in agriculture,forestry,geological survey and analysis,and urban planning.Scene classification,which aims to recognize and analyze the whole image scene,and obtain the scene semantic information of the image itself to explain the content of the image scene,is one of the basic techniques of the above various applications.The traditional remote sensing image classification technology is very mature,and it can achieve relatively high classification accuracy for both pixel and object-oriented classification.However,for some classes,with advanced semantic features,such as airports,business districts,residential areas,and parking lots,the composition may be a variety of feature types,in the meantime,the same type of features may also have complex spatial arrangements in different images.And the characteristics of spectral heterogeneity,as well as different lighting conditions and the effects of scale variation,make the traditional classification method difficult to perform the task of scene classification.The main reason is that traditional classification methods are difficult to establish the relationship between lowlevel visual features,such as spectrum,texture,structure,and high-level semantic information,leads to the so-called "semantic gap" problem.Bag of visual words(Bo VW)combined with the probabilistic topic model(PTM)is a commonly used method to solve the "semantic gap" problem.In view of the inherent shortcomings of the probabilistic topic model and the rapid development of natural language processing(NLP)technology,this paper introduces the natural language processing model Doc2 Vec and combines with Bo VW,proposes a scene classification method.In order to verify the effectiveness of the proposed method,this paper conducts a scene classification experiment based on the UC Merced dataset of 21 semantic scenarios and the RSSCN7 dataset of 7 semantic scenarios with 4 different scales.Firstly,combing the low-level features commonly used in previous literature about scene classification,combined with the technical route,data overview,application scenarios,etc.,select GLCM as the texture feature extraction algorithm,color moments as the spectral feature of the image,SIFT as the structural feature algorithm.Secondly,the sampling method was improved during the extraction of image features.According to the color moment,GLCM and SIFT,the sub-pixel segmentation method,evenly grid sampling method and dense key point sampling method are used to sample the three features in order to obtain the optimal feature sets.Finally,in order to verify the performance of Doc2 vec,this paper sets the experiment to compare with the PTM such as TF-IDF and LDA.The experiment on UC Merced focuses on the multi-category scene classification performance of the model,while the RSSCN7 focuses on verify the model's performance on different scales,Illumination and scale variation.The experimental results show that the proposed method has higher classification accuracy and robustness on the two data sets while superior to the traditional PTM and expresses the image content and features effectively.