Enhancing and Interpreting Deep Learning for Sea Ice Charting using the AutoICE Benchmark

Accurate mapping of sea ice is crucial for marine navigation and monitoring climate change. Automating sea ice mapping remains challenging due to remotely-sensed signal ambiguity, the dynamic nature of sea ice, and limited field measurements. The AutoICE challenge recently introduced a benchmark to advance deep learning for sea ice mapping. Top-performing solutions used the U-Net architecture with extra pre/post-processing steps and incorporated location features to obtain higher metrics. However, model interpretation and diagnostics remain limited. In this paper, we develop a customized multi-task DeepLabV3 model and achieve state-of-the-art performance of 87.3% combined score without extra pre- and post-processing, and outperform the current state of the art in predicting Stage of Development (SoD). Our approach enhances generalizability and outperforms current cross-scene sea ice retrieval methods. We further use interpretability methods including Gradient SHAP (Gradient Shapley Additive Explanations) and Gradient-Weighted Class Activation Mapping (Grad-CAM) to shed light on the contribution of individual input features and pixels on model decisions, including the role of geospatial (i.e., location) encoding. While incorporating geospatial encoding seemingly improves inference on the benchmark-designated test set, our additional model interpretation and spatial cross-validation reveal over-reliance on geolocation and overfitting to the test set. This suggests that the top-performing solutions on the AutoICE challenge are likely to lack geographic generalizability, a common issue in remote sensing, made worse by the use of location in the models. As such, we urge caution and recommend the use of spatial cross-validation and interpretability methods when using location information as input for remote sensing applications and further development of sea ice mapping algorithms.