Research On Remote Sensing Cross Domain Scene Classification With Unknown Categories

Remote sensing cross domain scene classification studies how to use the existing labelled data to facilitate the current classification task,in which the labelled data and the data to be classified usually come from different distributions(known as different domains).As remote sensing technology enters the multi-sensor era,it takes time and effort to acquire a large number of annotations for each kind of data.It is of great value to reuse existing data and improve data utilization rate.This dissertation focuses on remote sensing cross domain scene classification with unknown categories.In practical,for a remote sensing dataset to be classified,1)it is usually hard to find an existing annotated dataset that covers all categories of the target domain dataset,and the problem of incomplete categories needs to be solved;2)because the categories of existing annotated dataset and target dataset are not exactly the same,unknown categories need to be separated to solve the problem of negative migration in distribution alignment caused by category mismatch;3)it is desired to classify unknown categories given a few labelled samples,the problem of performance degradation caused by domain shift in few-shot classification needs to be solved;4)the model is expected to have a comprehensive performance on multiple source domains,and the problem of data balance needs to be solved.In Summary,the following four studies are carried out:(1)Multi-source compensation network for remote sensing cross-domain scene classification.The scene classification network trained on different distributed labelled datasets is used to complete the target classification task.Considering the difficulty of finding a labelled datasets(source domain)completely cover all the categories of target dataset(target domain),we propose to learn knowledge from multiple complementary source domains to finish the target domain dataset classification task,and then use adversarial-based domain adaptation method to reduce distribution shift between different domains,finally combining multiple source domain classifiers to finish the target domain classification task.(2)Unknown category separated alignment for cross-domain scene classification.The current domain adaptation methods align the entire target domain distribution to the source domain distribution,but negative transfer is resulted when there is a category shift between the source domain and the target domain.The target domain unknown categories should be processed separately.A separation alignment network is proposed,and a separation mechanism is established according to the knowledge learned from multiple source domains.After the target domain is divided into known and unknown categories,different weights are assigned to known and unknown category data in crossdomain alignment and classification procedures.(3)Cross-domain few-shot classification based on domain mapping network.Fewshot remote sensing scene classification is studied,which just uses a few labelled samples to complete new category recognition.To solve the performance degradation of current methods in deal with few-shot tasks from distribution different from the training dataset,a generalized few-shot classification model with the property of adaptive to different distributions is proposed.A domain mapping network is proposed,which first learns the few-shot classification model on the training set,then maps the training set and test set to the same distribution to reduce the distribution shift,and finally adapts the model to the test data distribution to complete the few-shot classification task.(4)Multi-source generalization with multi-task learning.It aims to reuse the shape and contour information of other related tasks to reduce the label requirements of the current task,by completing multiple tasks at the same time.In order to avoid information interference between different tasks,a multi-branch network is proposed,which uses a shared branch to complete multiple tasks to realize information sharing,and uses multiple task-specific branches to complete specific tasks to reduce interference between tasks.The multiple scene classification task and pixel scene classification task are explored to verify the effectiveness of the proposed method.