Multi-Scale Semantic Segmentation Of High-Resolution Remote Sensing Imagery Based On Deep Learning

With the development of high resolution remote sensing technology,high resolution remote sensing images have been widely applied to various application fields,such as national economy,national security,and so on.The extraction and recognition of geographic objects from high resolution remote sensing images are key steps for subsequent high-level processing,therefore,the segmentation of high resolution remote sensing images,especially the semantic segmentation has become one of the important issues in this field.In recent years,deep learning has developed rapidly,and the deep learning network can extract the semantic information related to the task layer by layer,which has surpassed the traditional machine learning methods in many fields.In this thesis,based on the study of semantic segmentation of deep learning,the multi-scale semantic segmentation of high-resolution remote sensing images has been explored combining the multiscale characteristics of geographic objects in remote sensing images.These studies have important theoretical and practical values.The main work and innovation in this thesis are as follows:(1)As different sizes of geographic objects always exist in high resolution remote sensing images,features from different sizes of receptive fields should be acquired for the accurate segmentation of geographic objects in high resolution remote sensing images.In this thesis,several semantic segmentation methods based on Fully Convolutional Networks are studied.Their characteristics and the segmentation results are analyzed and compared.Meanwhile,several types of segmentation evaluation criteria have been used to assess the accuracy of the results of semantic segmentation quantitatively.(2)The results of semantic segmentation based on deep learning usually show blobby and cannot retain accurate edges of geographic objects in high resolution remote sensing images.The traditional object-based methods can usually keep edges of geographic objects well,but the results of segmentation have no semantic information.The ability of extracting semantic features of deep learning networks and the advantage of keeping accurate edges information of traditional segmentation methods are combined together in this thesis.The segmentation results based on deep learning under the constraint of objects are obtained.The combination has improved both the edge accuracy and segmentation accuracy to some extent.(3)A large amount of labeled data are usually required for semantic segmentation based on deep learning,but remote sensing images always lack enough labels.A weakly-supervised learning method of semantic segmentation based on scribbles is proposed.Firstly,the condition random field is constructed on the whole remote sensing image.The appearances of pixels can be taken as the observation field and the labels of pixels can be taken as the label field.The problem of image semantic segmentation can be converted into pursuing a solution of a label field with the maximum posterior probability given a specific observation field.With the help of condition random field,the partly labeled image can be extended into the whole labeled image which can be considered as ground truth and be used to train the deep neural network.The trained deep neural network can be used to model different types of geographic objects and update the probabilities of pixels that belong to different classes.Then,labels of the whole image will be renewed by the condition random field again and will be used to train the deep neural network in the next iteration.The renewing of labels and the training of the deep neural network alternate until a stable segmentation result is obtained.The effectivity of the proposed method has been proved by the experiment.