| Land-cover classification for satellite images,as can provide useful structured semantics,is an important and necessary pre-processing task for many downstream applications.However,the frequent presence of clouds and haze on satellite images poses a big obstacle for this task.To deal with the task of land-cover classification with existence of clouds and haze on satellite images,many solutions have been proposed for the cloud/haze removal task and land-cover classification task,respectively.However,all these solutions can only solve the problem in twostep ways with cloud/haze removal first classification second,while few of them investigate a combination of these two tasks to solve the task in one step.To this end,a multi-task solution has been proposed to solve this problem,which will perform cloud/haze removal and land-cover classification at the same time in a single network and thus get classification results and cloud/haze removal results in one step.In this multi-task solution,we proposed a carefully designed multi-task network HCR&LC-Net(Have and Cloud Removal & Land-cover Classification Network),which is composed of a shared part to get sufficient deep representations,a classification branch and a cloud/haze removal branch.For the weighting of these two task-specific branches,three different weighting strategies have been tried.A large-scale dataset SEN12MS-CR-LC has been utilized for training and testing of this network.Improvement in the land-cover classification performance has been observed through experiments compared with single-task models,demonstrating the effectiveness of this multi-task solution.Furthermore,self-supervised learning strategy has been utilized to generalize the proposed multi-task solution to more general cases without enough training samples.Pretext tasks: S1&S2 matching,S2 rotation and spatial relation prediction have been created and tested,with the spatial relation prediction one proved to be the most ideal pretext task for our final model.Experiments have proved the proposed pretext tasks can further improve the performance of the final task when there are enough training samples and at the same time save the performance of the network from a steep drop when there are only limited training samples,making it possible to generalize our multi-task solution to other cases with limited training samples. |
