Semantic Understanding Of High Resolution Remote Sensing Image

With the rapid development of remote sensing technology,remote sensing data can be obtained more and more easily.Besides,the quality of these remote sensing data also improves a lot.So how to effectively use these remote sensing data,which contains a wealth of information,is a problem worthy of research.The visual information of high spatial resolution(HSR)remote sensing image can reflect the main content of this image and the corresponding semantic information can describe the image more detailed.So how to combine these two kinds of information,in order to make full use of the HSR images then understand these images in the semantic level,is an urgent problem to be solved.At present,the researches on HSR images only use the visual information and a small amount of semantic information,such as object detection,image classification,image segmentation,scene classification and so on.However,these works could only recognize the objects in the images or get the class labels of the images.But they could not recognize the attributes of the objects and the relation between each object.Namely,the aforementioned works cannot understand the HSR images in a semantic level.And semantic understanding of HSR images is still blank in remote sensing analysis area currently.In addition,there is no dataset both contains the HSR images and the corresponding text descriptions.According to the above two problems,this thesis first constructed two high-resolution remote sensing image-captions datasets based on UCM dataset and Sydney dataset,which are two representative HSR remote sensing image datasets.These two datasets combine the visual and semantic information of HSR images.So that we can fully utilize the HSR data,making it plays a greater role in the field of city planning,disaster monitoring,military reconnaissance.This is one of the main purposes of this thesis.Then,this paper proposes a deep multi-modal neural network modal,using convolutional neural network(CNN)to extract the visual information of images and using recurrent neural network(RNN)to combine the visual information with semantic information of the HSR images.Then the two networks are integrated to construct the deep multi-modal neural network model and finally achieve high-level semantic understanding of HSR remote sensing images.In the end,this paper compares the results of different types of CNN combined with RNN on two HSR remote sensing images-caption datasets.The results of the experiments prove the validity of our approach from the aspects of qualitative and quantitative.