Research On The Automatic Extraction Method Of Green Tide From Dual-polarization SAR Remote Sensing Images Of Gaofen-3 Satellit

Green tide is one of the main marine disasters affecting the coast of the Yellow Sea and the East China Sea in summer.Efficient and accurate green tide monitoring is of great significance to marine disasters prevention and oceanic environmental protection.Automatic extraction of green tide is one of the key technologies to understand the source,geographical scope,and drifting path of green tide,which can provide strong support for the prevention and control of green tide.The outbreak time of green tide disasters is mostly in June and July every year,often accompanied by cloudy and rainy weather.The traditional optical images are not suitable for the all-weather observation of green tide due to the fact that they are greatly affected by clouds.Synthetic Aperture Radar(SAR)on the other hand,can realize all-day,all-weather,and all-around Earth observation due to a lower sensitivity to clouds,rain,and fog.SAR images are therefore effective supplementary data for the monitoring of green tide.Therefore,this paper studies the green tide extraction method based on the dual-polarization SAR remote sensing images of the Chinese Gaofen-3(GF-3)satellite.The main work is as follows:(1)Considering the differences in brightness and noise across the different areas in a SAR image,it is difficult to use the fixed global threshold to extract all the green tide information from the image.Combined with iterative threshold algorithm and histogram bimodal algorithm,an automatic green tide detection method based on adaptive threshold is proposed.The results show that our detection method not only improves the extraction accuracy of green tide,but also realizes the automation of green tide extraction.Furthermore,it is also found that cross-polarization images may be more suitable for extracting green tide than co-polarization images due to their lower noise level.(2)Considering that the optimal selection of high-dimensional features can effectively reduce the redundancy between features and improve the classification efficiency,this paper designs a feature optimization method for SAR images combined with Bhattacharyya distance and the Separability index.In order to lighten the burden of model training and further improve the prediction efficiency,a lightweight semantic segmentation network Mobile-Seg Net is designed based on Mobile Nets and Seg Net.Based on the two,an automatic green tide extraction method of GF-3 SAR images based on feature optimization and semantic segmentation is proposed.Experimental results show that this method can not only effectively reduce the feature dimension of SAR image required for green tide extraction,but also has good automatic recognition and extraction ability for both large and small patches of green tide.(3)Multiple different semantic segmentation algorithms based on deep learning are used to extract green tide information from GF-3 dual-polarization SAR remote sensing images.The experimental results show that compared with the classical semantic segmentation algorithms such as FCN,Seg Net,PSPNet,and Deep Labv3+,Mobile-Seg Net not only effectively improves the accuracy and speed of green tide detection,but also has significant advantages in space occupation.And it is found that the semantic relationship contained in single scenes is limited.The deep learning semantic segmentation algorithm with complex structure or deeper network level,such as PSPNet and Deep Labv3+,may not be suitable for the simple binary classification problem such as green tide extraction from SAR remote sensing images,which not only consumes excessive time and space but also is difficult to ensure the accuracy of green tide detection.The dual-polarization SAR image has a broader strip width,which can meet the actual needs of large-scale green tide monitoring.In this paper,the automatic green tide detection and extraction of GF-3 dual-polarization SAR remote sensing images are realized from the two directions of traditional digital image processing and mainstream deep learning,so as to provide technical support for the monitoring of green tide across Chinese coastal areas,and to a certain extent,promote the application and development of microwave remote sensing data of domestic GF-3 satellite.