Evaluating aufeis detection methods using Landsat imagery: Comparative assessment and recommendations

Abstract

In the continuous permafrost environment of the North Slope of Alaska, extensive aufeis fields develop each winter on river floodplains, primarily via discharge from perennial springs. Currently, changes to the spatial and temporal distribution of large aufeis fields are predominantly monitored using optical satellite imagery. However, existing detection methods struggle to distinguish between snow and ice surfaces.

This study compares the accuracy of two techniques for identifying aufeis in a dataset comprising 515 Landsat optical images across four aufeis fields on the North Slope of Alaska. The first method involves empirical thresholding on snow and ice indices (2FT), while the second applies random forest (RF) machine learning methods. We evaluate their performance on multiple training and test datasets with pixel-, image-, and site-based stratification. Additionally, we evaluate the utility of additional bands and indices in aufeis detection using a grid-search for the top features (3FT) and feature importance metrics.

The more complex RF classifier, which relies on an extensive training dataset, outperforms both feature thresholding methods across all validation datasets with an average F1 score of 0.967±0.029. Feature importance metrics indicate that the near-infrared is effective for distinguishing between snow and ice surfaces. These findings demonstrate that machine learning approaches significantly enhance aufeis detection capabilities in snow-affected scenes and improve the retrieval of the annual maximum aufeis extent. While scaling challenges remain for these techniques, the results provide a foundation for improving our ability to monitor regional aufeis dynamics and their role in hydrologic and permafrost systems.