A lightweight machine learning-based melt pond extraction scheme for oblique shipborne imagery

High-resolution shipborne optical images can accurately capture the melt ponds on Arctic sea ice surfaces, serving as an important supplement to satellite remote sensing. However, a synthetic scheme ensuring both accuracy and efficiency for extracting ice, melt ponds and open water from a shipborne sea ice imagery set is still lacking. Thus, the Oblique Ice Surface Feature Extraction algorithm (O-ISFEA) was used in this study to extract surface types from shipborne imagery via a lightweight machine learning method. In this study, a shipborne dataset captured during the Chinese National Arctic Research Expedition of 2018 (CHINARE-2018) was used as an example. Watershed transformation was employed as a segmentation method, a random forest model was used to classify segmented regions of interest into three surface types, and orthographic correction was applied for geometrical postprocessing. We found that the overall accuracy and κ coefficient of the O-ISFEA scheme for the CHINARE-2018 shipborne imagery set were 96.50% and 0.90, respectively. It has the advantage of lightweight on training and processing with a classification performance comparable to other deep learning methods. This scheme has the potential to be applied for the intelligent navigation of icebreakers in the summer Arctic in the future.