Decoding the Partial Pretrained Networks for Sea-Ice Segmentation of 2021 Gaofen Challenge
Sea-ice segmentation is of great importance for environmental research, ship navigation, and ice hazard forecasting. Remote sensing (RS) images have been a unique data source for rapid and large-scale sea-ice monitoring. The 2021 Gaofen Challenge has offered a track of sea-ice segmentation based on...
| Published in: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
|---|---|
| Main Authors: | , , , , , , , |
| Format: | Article in Journal/Newspaper |
| Language: | English |
| Published: |
IEEE
2022
|
| Subjects: | |
| Online Access: | https://doi.org/10.1109/JSTARS.2022.3180558 https://doaj.org/article/141f20c79b9549fc9ac56e01b7a6b865 |
| Summary: | Sea-ice segmentation is of great importance for environmental research, ship navigation, and ice hazard forecasting. Remote sensing (RS) images have been a unique data source for rapid and large-scale sea-ice monitoring. The 2021 Gaofen Challenge has offered a track of sea-ice segmentation based on optical RS images. For the initial competition, our team ranked 3rd place (<monospace>deepjoker</monospace>) in the accuracy leaderboard and the solution has been the most efficient algorithm to achieve a segmentation score above 97.79%. In this article, we briefly introduce our three strategies of the achievement including: 1) decoding the partial pretrained networks which can simultaneously capture the complex boundaries of sea ices and decrease the computational cost without the performance drop; 2) employing the classwise Dice loss for solving the gradient vanishing problem when most ground-truth maps are backgrounds; and 3) replacing the commonly exploited decoder with the one proposed by Silva et al. (2021). The main contributions are twofold: 1) an efficient and effective sea-ice segmentation method is proposed and 2) the gradient vanishing problem of binary Dice loss is investigated under some scenarios and solved by introducing its classwise version. Comparison and ablation experiments demonstrate the effectiveness of the proposed method with respect to other commonly adopted deep segmentation models. |
|---|