|
Precise mapping of Blue Ice Regions (BIRs) in Antarctica can facilitate resource management during expeditions and help plan field investigations. Satellite-based BIR mapping was traditionally performed with large-swath, coarse to medium-resolution imagery. This study presents a comprehensive evaluation of four classification methods for mapping blue ice regions: Neural Network Classifier (NNC), Support Vector Machine (SVM), Maximum Likelihood (MXL), and Mahalanobis Distance (MD). Each method was trained and tested using a labeled dataset, and their performance was assessed based on Overall Accuracy (OA), Bias, and Average Root Mean Squared Error (RMSe). Among the methods, SVM stood out with exceptional performance, achieving an impressive OA of 92% and demonstrating a low bias of 1.27%. The SVM method also exhibited superior spatial accuracy, as indicated by an Average RMSe of 1978 m2, making it well-suited for precise delineation of blue ice regions. On the other hand, NNC and MXL demonstrated reasonable overall accuracy but exhibited higher spatial errors. MD showed potential for improvement in spatial accuracy, recording an Average RMSe of 3031 m2. Based on the evaluation, SVM emerged as the most effective technique for mapping blue ice regions, providing valuable insights for glaciological research and climate change studies. MODIS, Landsat, Sentinel, and Very High Resolution (VHR) data, such as WorldView-2 (WV-2), are recommended for mapping large, moderate-sized, and small-sized BIR patches, respectively. The inclusion of WV-2 data allows for accurate estimation of the maximum summer extent of BIRs across Antarctica, facilitating comprehensive assessments of their spatial distribution and dynamics. This research contributes to advancing the understanding of BIRs and their significance in climate and glaciological studies. The refined VHR blue ice map from the current study will be useful for future studies of mass balance, radiation budget, and regional climate changes in the study region.
|
