Preliminary Experiment on Complex Sea Ice Motion Monitoring Based on Deep Learning Architectures

The dataset used in this study is organized into three main folders, each serving a distinct purpose in the analysis of ice floe dynamics and the evaluation of matching algorithms under rotational motion.

The first folder, Ice Floe Data for the Experiment, contains the ice floe imagery selected for our experiments. It includes three subcategories: the original raw images, the processed images after preliminary enhancement, and the binarized masks generated through ice–water classification, which are essential for subsequent feature extraction and matching tasks.

The second folder, Extraction and Matching Results of Traditional Methods for Ice Floe Monitoring, stores the outcomes of feature point extraction and matching using traditional computer vision techniques, specifically SIFT and A‑KAZE, applied to monitor rotational motion of ice floes. Within this folder, subfolders B1 through B5 correspond to the five individual ice floes examined in our study.

The third folder, Influence of Ice Floe Rotation on Deep-Model Matching, provides data to investigate how ice floe rotation impacts the matching performance of deep learning models. To generate this data, each selected ice floe image was rotated incrementally by 5° over a full 360° range, resulting in 72 rotated versions per floe. These images were then input into a deep matching model, and all matching outcomes were recorded. An Excel file in this folder summarizes the precision variation for each of the five ice floes (B1–B5) across the 72 rotation angles. Additionally, subfolders B1 to B5 contain the complete set of matching results for every rotation trial, enabling detailed analysis of rotation effects on deep model performance.