Research On Remote Sensing Image Semantic Segmentation Based On Improved U-Net

With the rapid development of satellite remote sensing technology in China,the acquisition of remote sensing images has become increasingly convenient and fast.Remote sensing images are widely used in the fields of urban planning and development,rural arable land area detection,geological disaster forecasting and national defense.How to effectively extract useful information from a large number of remote sensing images has become a research hotspot.Accurate segmentation of remote sensing images is an important prerequisite for extracting effective information from remote sensing images.Deep learning techniques are currently developing rapidly and are widely used in several research fields.This thesis explores the application of convolutional neural networks in the field of semantic segmentation of remote sensing images based on convolutional neural networks,which are a popular deep learning technique nowadays.The main research elements include:(1)Aiming at the problems of small size,difficulty in learning image texture details and poor image segmentation fineness of some feature categories in remote sensing images,this thesis proposes a U-Net model based on multi-scale feature fusion,drawing on the idea of FPN networks,to effectively fuse the multi-scale semantic information extracted from the encoder side.To address the problems of poor segmentation accuracy and segmentation discontinuity of larger size and irregular objects,this thesis explores the problems such as data loss in the use of ordinary convolution operation,and proposes the use of null convolution instead of ordinary convolution operation in convolutional neural network,which increases the perceptual field of convolution operation and further improves the model segmentation accuracy.(2)To address the problem that the accuracy of the network model in predicting pixels of different categories is reduced due to the uneven data of feature categories in remote sensing images,this thesis explores the concentration loss function often used in the training process of deep learning network models,and after analysis and discussion,this thesis proposes a new loss function based on category weights,namely w DL.loss function can,to a certain extent,alleviate the segmentation accuracy problem caused by the uneven percentage of feature categories in remote sensing images.(3)In order to strengthen the feature extraction ability of the network model for remote sensing images,this thesis adds the channel attention mechanism SE block module to the process of jump connection.On this basis,this thesis also introduces the LSTM module and embeds the LSTM module into the SE block module to use the high-level semantic information in the network model to guide the extraction of the low-level semantic information and solve the problem of large differences between the low-level semantic information and the high-level semantic information.Through experimental verification,the feature extraction enhancement module proposed in this thesis achieves better improvement results.(4)In order to verify the effectiveness of the segmentation model proposed in this thesis,some classical semantic segmentation models are also reproduced in the experimental process for comparison with the method proposed in this thesis.The semantic segmentation model proposed in this thesis was validated on the2020 CCF BDCI remote sensing image parcel segmentation dataset.The experiments show that the method proposed in this thesis has strong practical value on real remote sensing scenarios.