On Mathews Correlation Coefficient and Improved Distance Map Loss for Automatic Glacier Calving Front Segmentation in SAR Imagery

The vast majority of the outlet glaciers and ice streams polar sheets end in ocean. Ice mass loss via calving into ocean has increased over last few decades. Information on temporal variability front position (CFP) provides fundamental information state glacier stream, which can be exploited as calibration validation data to enhance dynamics modeling. To identify CFP automatically, deep neural network-based semantic segmentation pipelines used delineate acquired synthetic aperture radar (SAR) imagery. However, extreme class imbalance is highly challenging for accurate these images. Therefore, we propose use Mathews correlation coefficient (MCC) an early stopping criterion because its symmetrical properties invariance toward imbalance. Moreover, improvement distance map-based binary cross-entropy (BCE) function. map adds context function about important regions helps account imbalanced data. Using MCC demonstrates average 2.16% dice slightly ice/nonice 17.43% segmenting CFPs, compared commonly BCE. modified further improves performance by another 1.6%. These results are encouraging they support effectiveness proposed methods problems suffering from imbalances.