VDS2RAW

The dataset described in the provided information is called VDS2RAW (Vessel Detection from Sentinel-2 Raw). It is a collection of raw granules derived from Sentinel-2 products, specifically focusing on vessels in Danish coastal areas. The dataset is utilized for ship detection and analysis purposes. To construct this, a reference dataset containing Sentinel-2 L1C tiles, including vessel images, was initially identified. This reference dataset was obtained from the work of Ruiloba et al. (2020), consisting of Sentinel-2 L1C tiles acquired in 2019. The dataset contained 1426 ship images in 24x24 pixel dimensions. To download the specific granules of interest (L0 granules), a polygon encompassing the region of interest (ROI) and a date range was defined. The download process retrieved all L0 granules where the reference band (B02) intersected the ROI during the specified time period. By applying a sufficient margin to ensure the inclusion of all relevant bands, a polygon surrounding the ROI and all the ships for each of the eight acquisition days was used. A total of 390 L0 granules were downloaded and decompressed to obtain raw processable data. To label the dataset, bounding boxes were manually created around each ship using bands B02, B03, and B04. The coarse spatial coregistration technique, described in (Meoni et al., 2023), was applied to enable manual labeling. Missing elements due to coregistration procedures were filled using adjacent granule pixels when possible, and in other cases, the areas with missing pixels were cropped. Consequently, the granules had varying pixel areas. After labeling, a subset of 166 raw granules was selected from the initially downloaded 390 granules. The dataset was further divided into training (105), validation (27), and test (34) sets. The number of ship annotations in each set was 483, 119, and 93, respectively. The average pixel area of the granules in the dataset was calculated as 2588.971 pixels by 1669.44 pixels, while the mean annotation area was 13.59 square pixels by 15.67 square pixels.本数据集命名为VDS2RAW（Vessel Detection from Sentinel-2 Raw，即基于Sentinel-2原始数据的船舶检测数据集）。该数据集是一组源自Sentinel-2卫星产品的原始影像块集合，重点聚焦丹麦沿海区域的船舶目标，主要用于船舶检测与相关分析研究。为构建该数据集，研究人员首先从Ruiloba等人（2020）的研究成果中获取参考数据集，该数据集包含Sentinel-2 L1C级影像瓦片及船舶影像样本，由2019年获取的Sentinel-2 L1C级影像瓦片组成，内含1426张24×24像素的船舶图像。为下载目标L0级原始影像块，研究人员划定了涵盖研究区域（ROI）的多边形范围与时间区间。下载流程将提取所有在指定时段内，其参考波段B02与研究区域存在空间交集的L0级影像块。为确保覆盖所有相关波段，研究人员为8个成像日的研究区域及所有船舶添加了足够的缓冲范围，最终下载并解压得到390个L0级影像块，以获取可用于处理的原始数据。为完成数据集标注，研究人员借助B02、B03、B04三个波段，手动为每艘船舶绘制边界框。为保障手动标注的准确性，研究人员应用了Meoni等人（2023）提出的粗空间配准技术。若配准过程中出现像素缺失，将尽可能使用相邻影像块的像素进行填充；其余情况下，则直接裁剪掉存在像素缺失的区域。由此，最终得到的影像块像素尺寸存在差异。标注完成后，研究人员从初始下载的390个影像块中筛选出166个，作为最终可用数据集。该数据集进一步被划分为训练集（105个影像块）、验证集（27个）与测试集（34个），各集合对应的船舶标注数量分别为483、119与93。数据集内影像块的平均像素尺寸为2588.971×1669.44像素，而单个标注框的平均面积为13.59×15.67平方像素。