Preprints
https://doi.org/10.5194/egusphere-2023-295
https://doi.org/10.5194/egusphere-2023-295
03 Mar 2023
 | 03 Mar 2023

Towards a manual-free labelling approach for deep learning-based ice floe instance segmentation in airborne and high-resolution optical satellite images

Qin Zhang and Nick Hughes

Abstract. Floe size distribution (FSD) has become a parameter of great interest in observations of sea ice because of its importance in affecting climate change, marine ecosystems, and human activities in the polar ocean. The sizes of ice floes can range from less than a square metre to hundreds of square kilometres, so the most effective way to monitor FSD in the ice-covered regions is to apply image processing techniques to airborne and satellite remote sensing data. The segmentation of individual ice floes is crucial for obtaining FSD from remotely sensed images, and it is a challenge to separate floes that appear to be connected. Although deep learning (DL) networks have achieved great success in image processing, they still have limitations in this application. A key reason is the lack of sufficient labelled data, which is costly and time-consuming to produce. In order to alleviate this issue, we use classical image processing techniques to achieve a manual-label free ice floe image annotation, which is further used to train DL models for fast and adaptive individual ice floe segmentation, especially for separating visibly connected floes. A post-processing algorithm is also proposed in our work to refine the segmentation. Our approach has been applied to both airborne and high-resolution optical (HRO) satellite images, and successfully derived FSD at local and global scales.

Qin Zhang and Nick Hughes

Status: final response (author comments only)

Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor | : Report abuse
  • RC1: 'Comment on egusphere-2023-295', Anonymous Referee #1, 30 Mar 2023
  • RC2: 'Comment on egusphere-2023-295', Anonymous Referee #2, 14 Apr 2023
  • RC3: 'Comment on egusphere-2023-295', Anonymous Referee #3, 08 May 2023
  • EC1: 'Comment on egusphere-2023-295', Bin Cheng, 15 May 2023
    • AC4: 'Reply on EC1', Qin Zhang, 11 Jun 2023

Qin Zhang and Nick Hughes

Qin Zhang and Nick Hughes

Viewed

Total article views: 586 (including HTML, PDF, and XML)
HTML PDF XML Total BibTeX EndNote
384 176 26 586 11 12
  • HTML: 384
  • PDF: 176
  • XML: 26
  • Total: 586
  • BibTeX: 11
  • EndNote: 12
Views and downloads (calculated since 03 Mar 2023)
Cumulative views and downloads (calculated since 03 Mar 2023)

Viewed (geographical distribution)

Total article views: 584 (including HTML, PDF, and XML) Thereof 584 with geography defined and 0 with unknown origin.
Country # Views %
  • 1
1
 
 
 
 
Latest update: 01 Oct 2023
Download
Short summary
Ice floes image annotation is the key to training deep learning (DL) models to extract individual floes from images. But manually labeling each floe in images is a tedious work. To alleviate this, we first develop an automatic approach to annotate floe images without manual intervention. We then apply DL method for fast and adaptive floe instance segmentation for airborne and high-resolution satellite images, and have successfully derived floe size distributions at local and global scales.