Annotated time-series of lake ice C-band synthetic aperture radar backscatter created using Sentinel-1, ERS-1/2, and RADARSAT-1 imagery of Old Crow Flats, Yukon, Canada

The lake ice backscatter time-series dataset was created for the purpose of developing an automated temporal deep learning method of lake ice regime classification and study of lake ice dynamics in the Old Crow Flats (OCF), Yukon, Canada. The dataset consists of approximately 129,000 labeled backscatter time-series collected using imagery from four C-band synthetic aperture radar (SAR) spaceborne platforms: Sentinel-1 A (VV polarization), ERS-1 and 2 (VV polarization), and RADARSAT-1 (HH polarization), which cover the time period between 1992 to 2021. Labeling was done in Sentinel Application Platform (SNAP) by manually placing pins at locations identified as either floating ice, bedfast ice, or land through visual assessment of the ice regime/land on the last day of the time-series for a given season. Due to variable temporal coverage, the dates of labeling ranged from March 4 to March 22. The labeling date was selected as close as possible to mid-March, and care was taken to ensure that the air temperature was below 0°C. Then, the backscatter values at the locations marked by each pin were extracted for each of the scenes in a SAR stack, creating time-series of labeled backscatter values for each year covering the October to mid-March period. Labels were assigned based on three factors: 1) backscatter values, 2) value of the projected incidence angle of the SAR pulse, and 3) location of the pixel within the scene. Resampling to a daily frequency and linear interpolation were applied to compensate for the temporal irregularity of the data gearing it for the deep learning classification. The final labeled time-series consist of 161 time steps (i.e., one time step per day) covering the time period between October 4 and March 13. In addition, lake ice maps (containing three classes: bedfast ice, floating ice, and land) created using the novel temporal deep learning approach developed based on the time-series dataset are provided in PNG and GeoTIFF formats.

本湖泊冰后向散射时间序列数据集，旨在开发加拿大育空地区旧克洛弗拉茨（Old Crow Flats, OCF）的湖泊冰情自动时序深度学习分类方法，并用于该区域湖泊冰动力学研究，由此构建而成。该数据集包含约12.9万条带标注的后向散射时间序列，所用影像源自4个C波段合成孔径雷达（synthetic aperture radar, SAR）星载平台：哨兵-1A（Sentinel-1 A，VV极化）、欧洲遥感卫星1号与2号（ERS-1和2，VV极化）以及雷达卫星1号（RADARSAT-1，HH极化），时间覆盖范围为1992年至2021年。标注工作依托哨兵应用平台（Sentinel Application Platform, SNAP）完成：通过目视评估对应季节时间序列最后一日的冰情与地表状态，手动在浮冰、固定底冰、陆地三类点位插入标记。由于各季时间覆盖范围存在差异，标注日期介于3月4日至3月22日之间。标注日期尽可能贴近3月中旬，且严格确保当日气温低于0℃。随后，针对SAR影像栈中的每景影像，提取各标记点位的后向散射值，生成覆盖每年10月至次年3月中旬的带标注后向散射时间序列。标注结果基于三项因素确定：1）后向散射数值；2）SAR脉冲的投影入射角；3）像素在影像中的空间位置。为弥补数据时序不规则性以适配深度学习分类任务，研究团队对数据进行了日频率重采样与线性插值处理。最终的带标注时间序列包含161个时间步（即每日对应一个时间步），覆盖时段为每年10月4日至次年3月13日。此外，本数据集还提供了基于该时间序列数据集开发的新型时序深度学习方法所生成的湖泊冰情图，包含固定底冰、浮冰与陆地三类要素，格式为PNG与GeoTIFF。