Research On Semi-supervised Sea Fog Detection Technology Based On Temporal Images Of Geostationary Meteorological Satellite

Sea fog is a natural phenomenon in which water vapor in the atmosphere condenses,causing horizontal visibility to drop below 1 kilometer,thereby affecting various activities in coastal and offshore areas.Therefore,accurate and reliable sea fog forecasting can help coastal people to take measures in advance to reduce or even avoid the harm caused by sea fog.In recent years,with the rapid development of remote sensing technology,the time and spatial resolution of remote sensing satellite data have become higher,and the amount of data has also become larger.Therefore,more and more researchers are using machine learning,deep learning,and other methods that can handle large-scale data to solve sea fog detection problems.Although the semantic segmentation neural network commonly used in deep learning can be used for sea fog detection,it cannot solve the specific problems of sea fog detection in a targeted manner.For example,various meteorological phenomena in sea fog detection have different shapes and sizes,single scale and style feature extractor cannot extract all kinds of features;adjacent remote sensing satellite data have correlations in the time dimension,which cannot be used by static semantic segmentation model;and there is only a small amount of clear and recognizable remote sensing data containing sea fog,while there is still a large amount of data mixed with clouds and fog that is difficult for the human eye to distinguish.This paper focuses on these three problems and optimizes the neural network structure and overall training framework.The main work includes:1.Based on the classical semantic segmentation network,this paper introduces three modules,multi-scale module,deformable convolution,and texture extraction,to enhance the model’s ability to extract features of different types such as multi-scale feature,etc.The optimized network has improved mIoU on the sea fog detection dataset.2.Drawing on the algorithm framework of video segmentation,the Swin Transformer structure is introduced,and the data feature maps of adjacent moments are fused to enable the network to utilize the temporal information in the data to filter out more clouds and include more sea fog parts.3.By introducing the Mean Teacher training structure in semisupervised learning,the student model can simultaneously learn the knowledge of both the real label and the teacher model,and thus learn the information in both labeled and unlabeled data.The overall optimized model has significantly improved performance compared to the original classical semantic segmentation neural network in terms of mIoU.Finally,based on the semi-supervised sea fog detection technology of stationary meteorological satellites,this paper constructs a sea fog detection system.The system can detect sea fog in coastal areas of China.The algorithm results can be more intuitively evaluated using CALIPSO satellite data.