Arctic Sea Ice Lead Detection From China FY-3D MERSI-II Quality Enhanced Images

The 250-m resolution Chinese Fengyun-3D (FY-3D) Medium Resolution Spectral Imager-II (MERSI-II) thermal infrared (TIR) data can help us understand the rapid variations of Arctic sea ice leads, which are key features within the sea ice. The challenges of utilizing the 250-m FY-3D MERSI-II TIR data are bowtie effect and nonuniform brightness stripe noise.

While previous solutions have addressed these issues separately, this study introduces a more integrated two-step image quality enhancement strategy for MERSI-II TIR images. It considered the interactions between the two issues and overcame the excessive or inadequate destriping in existing models due to the ideal stripe-type assumption. Specifically, for the bowtie effect, a rigorous geometric model suitable for MERSI-II was constructed. For the nonuniform brightness stripe noise, a novel adaptive multiscale frequential (AMSF) algorithm was developed. The proposed strategy was outperforming existing methods in quantitative and qualitative assessments with higher efficiency on both bowtie effect and stripe noise removal. To validate the effectiveness of proposed strategy in sea ice lead detection, the temperature anomaly method was used to extract leads in winter Arctic Baffin Bay from images of varying quality. The results show that the overall accuracy improved from 0.88 to 0.95.

However, the sea ice lead extraction results from the traditional temperature anomaly method on TIR images is affected by the subjective selection of window sizes and is prone to misclassification caused by clouds. To address these issues, a novel deep learning method is applied to quality-enhanced FY-3D MERSI-II TIR images, which adaptively extracts sea ice leads and reduces cloud interference. Several commonly image segmentation networks: PSP Net, U-Net, and Deeplabv3, were compared to identify the most suitable network for extracting sea ice leads from TIR images, with the U-Net architecture providing the best segmentation results.

Nevertheless, the segmentation results of U-Net network still exist some misclassification caused by clouds. Therefore, the network was further optimized by introducing frequency domain filtering modules, which eliminate the interference of low-frequency clouds and enhance the model's focus on the high-frequency regions like linear features of sea ice leads, thereby improving the extraction accuracy of sea ice leads from thermal infrared images. Results show that the novel network effectively reduces misclassification caused by clouds, thereby providing more accurate and reliable sea ice lead data.