Weakly Supervised Sea Fog Detection Based On Image-level Annotation

Sea fog is a kind of catastrophic weather that occurs at sea and can have a large impact on maritime activities.Sea fog detection is an essential technology in the meteorological field,which can timely grasp the formation and dissipation of sea fog and improve the safety of marine work.The traditional method in the meteorological field is to use data such as brightness temperature difference(BTD)observed by meteorological satellites to set thresholds for sea fog detection.With the development of deep learning,some scholars have also used cloud images generated by meteorological satellites to conduct research on sea fog detection using fully supervised semantic segmentation methods.However,visual interpretation in remote sensing requires experienced interpreters to use their professional knowledge to identify sea fog regions in satellite images,and fully supervised semantic segmentation also requires accurate pixellevel labels as a training dataset,while the irregularity of sea fog texture and morphology leads to a large amount of manpower and time consumed in obtaining pixel-level sea fog labels.In response to the above issues,this paper has decided to introduce a weakly supervised semantic segmentation technology,using trusted image-level sea fog labels,to perform sea fog detection tasks in satellite images.The following research works have been carried out in this paper:1.A sea fog detection dataset with image-level annotation is proposed.The dataset is labeled at image level for each sea fog image with the help of meteorological summaries of past sea fog occurrences from the Marine Weather Review,which reduces the time cost and improves the reliability of the labels.2.Modeling classification models based on class activation maps(CAM)to obtain more complete target area localization.This paper uses a basic classification network for feature extraction,and then uses global average pooling to calculate feature vectors and generate CAM to locate the sea fog occurrence area.Adding a puzzle branch and the corresponding loss function to optimize CAM makes the generated sea fog area more complete.3.Develop pseudo label generation strategies and use pseudo labels for semantic segmentation training.This paper uses the relationship between pixels to calculate the affinity matrix,and semantic propagation in the sea fog region is completed through iterative training to generate more accurate pseudo labels as supervised information for semantic segmentation training.Based on the above research work,this paper verifies the effectiveness of weakly supervised semantic segmentation methods for sea fog detection tasks.It is the first time to use image-level labels as the supervision information,and verify the accuracy and reliability of the model through CALIPSO data and ICOADS observation station data.