Research On Semantic Segmentation Of Remote Sensing Image Based On Deep Learning

Semantic segmentation of remote sensing images is a key technology in the field of processing remote sensing images,and its segmentation results can be widely used in environmental change monitoring,disaster assessment,urban planning and other fields.At present,remote sensing image semantic segmentation networks based on deep learning are developing rapidly,and how to use neural networks to effectively improve the accuracy of remote sensing image segmentation has become a common research goal for scholars.Aiming at the problem that the accuracy of segmentation and the complexity of the model are difficult to reconcile in the research of semantic segmentation of remote sensing images at this stage,this dissertation proposes an optimized semantic segmentation model which is based on ICNet.First of all,in order to solve the current problems of the lack of remote sensing data sets and the lack of targeted experimental samples,this dissertation constructs a remote sensing data set.We collected Google satellite remote sensing images and set the semantic segmentation target features into 5 categories,as well as used the arcgis to make corresponding label map.At the same time,in order to expand the data set and improve the generalization ability of the model,the remote sensing image is processed by data augmentation.Secondly,this dissertation proposes to use ICNet as the backbone network,on the premise of not increasing the complexity of the model,it achieves the efficient extraction of effective features in the input image through adding the ECA module to the feature extraction network;And in view of the problems of computational complexity and memory occupation of the expanded convolution in the network,we propose to replace the expanded convolution with the JPU module.While JPU expands the model’s receptive field,it can also extract the characteristics of multi-scale semantic information across multi-level feature maps,which improves the performance of the model.Finally,in order to verify that the optimized ICNet has a better semantic segmentation effect,this dissertation uses the semantic segmentation evaluation index as the criterion.Through the experimental results can show that the improved ICNet’s m Io U has increased by 5%,and the OA has increased by 4%.What’s more,it’s prediction results are more refined,and the misclassification situation has also been effectively reduced.In order to put the improved ICNet model into practical application,we made a remote sensing image semantic segmentation software based on deep learning.The software can directly call the trained model,and is equipped with a variety of remote sensing image processing tools,which can realize the semantic segmentation of remote sensing images more conveniently and intelligently.