Skip to main navigation menu Skip to main content Skip to site footer

Region-level Ground Truth can produce Pixel-level Segmentation for Sea Ice Concentration


Accurate high-detail maps depicting the concentration of sea ice are vital for a range of human activities and the precise monitoring of polar sea ice changes. However, the current method of creating sea ice concentration (sic) maps through manual expert annotations is both imprecise in terms of spatial resolution and time-intensive to prepare. Although various convolutional neural network (cnn)-based techniques have emerged for automating the process of generating sea ice maps from synthetic aperture radar (sar) images, a major challenge lies in the absence of finely-grained pixel-level labels for training these models. This lack of detailed training data limits the ability of these methods to yield reliable high-resolution mapping outcomes. To address this issue, this research generate predictions for pixel-level sic by training a model to learn from region-level sic ground truth data. Specifically, a novel regional loss function is developed, which allows direct integration of the regional sic values in ice charts to train a u-net-based sic estimation model. By doing so, this method avoids inaccuracies associated with translating region-level sic values into pixel-level ground truth values. Consequently, it enables the creation of estimations for pixel-level sic and the classification of areas as either ice or water. The effectiveness of this proposed approach is evaluated using the recently released ai4arctic sea ice challenge dataset, consisting of 533 sentinel-1 sar scenes and ancillary data along with their corresponding ice charts. The results of these tests showcase the capability of the proposed model in generating high-resolution sic maps at the pixel level that align well with existing ice charts and visual interpretations.