Open Set Segmentation For Remote Sensing Image Based On Multidimensional Features

With the global warming effect and the continuous aggravation of environmental pollution,extreme severe weather occurs frequently all over the world,and the natural disasters caused by it have caused huge economic and property losses for human beings.Remote sensing satellites generally operate in geosynchronous orbit,so they can provide large-scale,multi-temporal monitoring information to the earth,and have been widely used in the field of natural disaster loss assessment.Owing to the development of deep learning,semantic segmentation has also achieved significant improvements,which has been successfully applied in remote sensing image processing and recognition.Commonly,image semantic segmentation is based on Closed Set assumption: it assumes that all training and testing pixels come from the same label space which does not contain unknown label.However,due to the variety of geological and geomorphic changes caused by natural disasters,and the different geological structure and biological composition in the affected areas,the texture and spectral characteristics of remote sensing images show obvious differences,which make the assumption does not hold in real-world scenarios.In practical application,geological and geomorphic changes caused by natural disasters(such as fires,landslides,landslides,etc.)can be regarded as unknown ground features compared with common ground features(such as water bodies,forests,buildings,etc.).This requires the algorithm model to have both the ability to classify known categories and the ability to identify unknown categories,which belongs to the open set identification problem.In this thesis,we proposed a novel approaches for Open Set semantic segmentation of remote sensing image.We mainly focus on two aspects: Firstly,we construct a network structure based on the depth parallel mining of multi-dimensional features to improve the semantic segmentation effect,aiming at the problem that the existing models are insufficient in mining the shape and spectral information of remote sensing images;Secondly,we a construct semantic segmentation open set recognition algorithm based on multivariate Gaussian distribution to realize the recognition of unknown categories of ground objects.The main contributions of this work are:(1)This thesis reviews the research status of common open set recognition methods and analyzes the latest progress of open set semantic segmentation.Open set identification methods can be mainly divided into discrimination method and generation method.The discriminant method uses machine learning algorithm to distinguish the feature space of known samples and unknown samples.The generation method realizes the recognition of unknown samples by introducing the reconstruction error into supervised learning.(2)Through the research on the application of the above six semantic segmentation techniques in the field of remote sensing images,a semantic segmentation method of remote sensing images based on multi-dimensional feature parallel mining is selected.According to Lovasz Loss,this thesis improves the Loss function of the generated model,uses the gate convolution structure to obtain the shape information of the image,uses the attention mechanism to obtain the waveband relationship information of the remote sensing image,and introduces a learnable sampling structure into the codec structure to improve the learning ability of the model.(3)Through the research of open set recognition method,the open set semantic segmentation algorithm of remote sensing image based on multivariate Gaussian distribution is build.In the model testing stage,PCA algorithm is used to fit different Gaussian distributions of known categories in the low-dimensional space.When a sample does not conform to the above Gaussian distribution,it is recognized as an unknown category.