Antarctic ocean temperatures are rising due to climate change, causing land ice to melt at increasingly higher rates. Ice shelf bottom melt is a key factor responsible for Antarctic ice mass loss and as such understanding melt processes in the Antarctic is therefore key to more accurately predict how the global sea level will respond to climate change in the foreseeable future. Basal melt results in the formation of both basal melt channels underneath an ice shelf and persistent sea ice wakes (named plume-driven polynyas) at the ice shelf shoreline. The goal of this research is to develop a method that can help to automatically infer basal melt locations along the Antarctic shoreline with significantly increased spatio-temporal resolution compared to previously researched basal melt detection methods.
We infer basal melt locations by detecting plume-driven polynyas. We used dual-pol (HH/HV) Sentinel-1 EW SAR data (40x40m resolution) in combination with GLCM textural features as input for a random forest classification that differentiates images as water or ice in four sub-classes: undisturbed ’open’ water, disturbed ’rough’ water, sea ice and (floating) land ice. We assessed what the advantages and limitations of this approach were for plume-driven polynya detection by performing water-ice (sub-class) classifications and examining which GLCM features proved most useful, what GLCM window size is preferred, and how classification can be aided by post-processing classified images.
We computed GLCM textures for window sizes w = [5,11,21] and created a classifier for each choice (GLCM5, GLCM11 and GLC21) and compared results to a classifier based on original dual-pol SAR data (BASE). Via cross validated recursive feature elimination we determined that ’sum average’ (HH and HV polarization) and ’difference variance’ (HV polarization) were most useful for separation of water and ice classes (HH_savg, HV_savg and HV_dvar). Our results have shown that using GLCM texture based dual-pol classifiers improves water-ice classification significantly compared to dual-pol only classifiers, although using HH_savg and HV_savg instead of orignal dual-pol data comes at a cost of reduced spatial resolution. Water-ice classification accuracy of BASE was 92.2% (kappa = 84.4%) was increased to 95.9% (kappa = 91.5%) for GLCM5, 96.3% (kappa = 92.7%) for GLCM11 and to 96.5% (kappa = 93.0%) for GLCM21. From a spatial context, GLCM21 showed an insufficient ability to detect small-scaled bodies of water at a sub-kilometer scale. GLCM5 showed unsatisfactory results in terms of sea ice classification. GLCM11 showed highest robustness in both these performance aspects and proved to be most successful classifier for the application of polynya detection. Using an area filter as a post-processing step proved successful when a classifier is based on GLCM data with a window size no larger than w=11. Noise output (small regions of falsely classified open water pixels) was heavily reduced via this form of post-processing and significantly increased polynya detection performance.
The final classified product however still contained too many incorrectly classified water regions of similar spatial scales as plume-driven polynyas to be able to apply this algorithm as a reliable automated polynya detection method. We urge to build upon this SAR-based detection method, by using additional non-GLCM input features or using extra post-processing steps, such as temporally filtering water body presence, until results are satisfactory for a fully automated plume-driven polynya detection algorithm. The method presented here has the potential to make detection significantly faster, easier and more accessible than the current methods available. Lastly, in its current state, this method can already be used to validate predicted locations of basal melt by ocean-ice sheet models and DEM-based methods.