Research On Remote Sensing Image Semantic Segmentation And Change Detection Based On Fully Convolutional Neural Network

With the rapid development of aerospace technology,a variety of satellite remote sensing images emerge in an endless stream.High-resolution remote sensing images have become important data sources for humans to quickly achieve global or large-area earth observations.However,the remote sensing image is disturbed by clutter and the contour of the target in the image is blurred.At the same time,with the exponential increase in the amount of remote sensing image data,higher-level requirements and challenges are placed on the processing technology of remote sensing images.In recent years,deep learning has been widely used in image understanding and visual analysis with its excellent algorithm performance.It breaks through the constraints of traditional computer vision algorithms and provides a new method for the processing of remote sensing images.This paper mainly studies and analyzes the semantic segmentation of remote sensing images and the detection of changes in SAR images.The full convolutional neural networks is used to achieve pixel-level end-to-end classification.The main of this paper is as follows:This paper presents a reduced Seg Net network(R-Seg Net)for semantic segmentation of high-resolution aerial images.The pixel-level classification is achieved by Encoder-Decoder network.The encoding network extracts image features and then the decoding network expands feature maps to the same size as the original image.The R-Seg Net network can implement semantic segmentation of aerial images.In order to further improve the segmentation accuracy,the following three strategies are used for experimental research: a)Combining the feature information of each convolution block in the R-Seg Net encoding network with the feature information of the corresponding convolution block in the decoding network.b)Taking into account the complex diversity of surface features in high-resolution aerial images and the recognition effect of small target objects,we convert the 6 classification model into 6 binary models.Then separately learn from different categories to get segmentation results.c)Using the advantages of multiple-classifier ensemble learning,we fuse the predicted results of the above different models and give the final segmentation results.Experiment results show that these measures can effectively improve the segmentation effect of the image and can accurately detect the different categories of target areas in aerial images.Finally,the higher segmentation accuracy is obtained.This paper converts the SAR image change detection problem into a binary classification image semantic segmentation problem.Combining the supervised deep neural network model and the unsupervised SAR image change detection problems,this part proposes a new SAR image change detection method based on full convolutional neural networks.Sparse auto encoder is used to learn the features from difference image,and it can suppress the impact of noise and provide the training sets for the following supervised semantic classification.The initial segmentation result will be obtained as a semantic label through the traditional clustering methods.The SAR image change detection based on the full convolutional neural networks mainly extracts more abstract semantic features from the feature maps generated by unsupervised feature learning,and makes use of these features to detect the changed region and unchanged region more accurately.Compared with convolutional neural network,this method simplifies the training step while improving the accuracy,and can realize end-to-end SAR image change detection at the pixel level.We have experimented the proposed method on many data sets,and the results are satisfactory.