Deep Learning Based Automatic Segmentation of Gas Flares in Single Beam Echo Sounder Data

Abstract

This thesis introduces the first study of instance segmentation applied to gas flares in single beam echo sounder data. We develop a comprehensive dataset consisting of 1,414 images, featuring 5,142 segmented objects identified as gas flare. A key contribution is the adaptation of the Brier score specifically for instance segmentation. Further, we show how to adapt the Weighted Box Fusion (WBF) algorithm for instance segmentation. Using the newly developed Brier metric for instance segmentation, as well as the mAP metric, we show that our ensemble models fused with WBF are quantitatively as good as the average human expert. However, our qualitative analysis identifies critical areas where these models fall short, indicating the need for further refinement to reach human-level performance. The thesis concludes by proposing potential improvements and future research directions. We remark that if implemented, these could bridge the gap between human and machine-level performance.