HiRISE Image Patches Obscured by Atmospheric Dust
收藏资源简介：
Overview The purpose of this dataset is to train a classifier to detect "dusty" versus "not dusty" patches within browse-resolution HiRISE observations of the Martian surface. Here, "dusty" refers to images in which the view of the surface has been obscured heavily by atmospheric dust. The dataset contains two sets of 20,000 image patches each from EDR (full resolution) and RDR ("browse" resolution) non-map-projected ("nomap") HiRISE images, with balanced classes. The patches have been split into train (n = 10,000), validation (n = 5,000), and test (n = 5,000) sets such that no two patches from the same HiRISE observation appear in more than one of these subsets. There could be some noise in the labels, but a subset of the validation images have been manually vetted so that label noise rates can be estimated. More details on the dataset creation process are described below. Generating Candidate Images and Patches To begin constructing the dataset, the paper "The origin, evolution, and trajectory of large dust storms on Mars during Mars years 24–30 (1999–2011)," by Wang and Richardson (2015), was used to compile a set of time ranges for which global or regional dust storms were known to be occurring on Mars. All HiRISE RDR nomap browse images acquired within these time ranges were then inspected manually to determine sets of images that were (1) almost entirely obscured by dust and (2) almost entirely clear of dust. Then, 10,000 patches from the two subsets of images were extracted to form the "dusty" and "not dusty" classes. The extracted patches are 100-by-100 pixels, which roughly corresponds to the width of one CCD channel within the browse image (the width of the raw EDR data products that are stitched together to form a full RDR image). Some small amount of label noise is introduced in this process, since a patch from a mostly dusty image might happen to contain a clear view of the ground, and a patch from a mostly non-dusty image might contain some dust or regions on the surface that are featureless and appear like dusty patches. A set of "vetting labels" is included, which includes human annotations by the author for a subset of the validation set of patches. These labels can be used to estimate the apparent label noise in the dataset. Corresponding to the RDR patch dataset, a set of patches are extracted from the same set of EDR images for the "dusty" and "not dusty" classes. EDRs are raw images from the instrument that have not been calibrated or stitched together. To provide some form of normalization, EDR patches are only extracted from the lower half of the EDRs, with the upper half being used to perform a basic calibration of the lower half. Basic calibration is done by subtracting the sample (image column) averages from the upper half to remove "striping," then computing the 0.1th and 99.9th percentiles of the remaining values in the upper half and stretching the image patch to 8-bit integer values [0, 255] within that range. The calibration is meant to implement a process that could be performed onboard the spacecraft as the data is being observed (hence, using the top half of the image acquired first to calibrate the lower half of the image which is acquired later). The full resolution EDRs, which are 1024 pixels wide, are resized down to 100-by-100 pixel patches after being extracted so that they roughly match the resolution of the patches from the RDR browse images. Archive Contents The compressed archive file contains two top-level directories with similar contents, "edr_nomap_full_resized" and "rdr_nomap_browse." The first directory contains the dataset constructed from EDR data and the second contains the dataset constructed from RDR data. Within each directory, there are "dusty" and "not_dusty" directories containing the image patches from each class, "manifest.csv," and "vetting_labels.csv." The vetting labels file contains a list of manually labeled examples, along with the original labels to make it easier to compute label noise rates. The "manifest.csv" file contains a list of every example, its label, and whether it belongs to the train, validation, or test set. An example ID encodes information about where the patch was sampled from the original HiRISE image. As an example from the RDR dataset, the ID "003100_PSP_004440_2125_r4805_c512" can be broken into several parts: "003100" is a unique numerical ID "PSP_004440_2125" is the HiRISE observation ID "r4805_c512" means the patch's upper left corner starts at the 4805th row and 512th column of the original observation For the EDR dataset, the ID "200000_PSP_004530_1030_RED7_1_r9153" is broken down as follows: "200000" is a unique numerical ID "PSP_004530_1030" is the HiRISE observation ID "RED7" is the CCD ID "1" is the CCD channel (either 0 or 1) "r9153" means that the patch is extracted starting at the 9153rd row (since all columns of the 1024-pixel EDR are used, no column is specified; it is implicitly always 0) Original Data The original HiRISE EDR and RDR data is available via the Planetary Data System (PDS), hosted at https://hirise-pds.lpl.arizona.edu/PDS/
概述 本数据集旨在训练分类器，用于在火星表面的浏览分辨率HiRISE（High Resolution Imaging Science Experiment，高分辨率成像科学实验）观测图像中，检测“有尘埃覆盖”与“无尘埃覆盖”的图像块。此处“有尘埃覆盖”指地表视野被大气尘埃严重遮挡的图像。 本数据集包含两组各20000个图像块，分别来自非地图投影（nomap）的EDR（全分辨率数据）和RDR（浏览分辨率数据）HiRISE图像，且类别平衡。图像块被划分为训练集（n = 10000）、验证集（n = 5000）与测试集（n = 5000），且同一份HiRISE观测图像生成的任意两个图像块，不会同时出现在多个子集当中。 数据集中的标签可能存在一定噪声，但验证集的一个子集已经过人工审核，因此可用于估算标签噪声率。数据集构建流程的更多细节如下方所述。 生成候选图像与图像块 为构建本数据集，我们采用Wang与Richardson（2015）发表的论文《火星年24–30（1999–2011）期间火星大型沙尘暴的起源、演化与轨迹》，整理出已知发生全球或区域性沙尘暴的火星时间范围。 随后，手动检视所有在此时间范围内获取的非地图投影HiRISE RDR浏览图像，筛选出两类图像：(1) 几乎完全被尘埃遮挡的图像，以及(2) 几乎完全无尘埃覆盖的图像。随后从这两类图像中各提取10000个图像块，分别构成“有尘埃覆盖”与“无尘埃覆盖”两个类别。 提取的图像块尺寸为100×100像素，大致对应浏览图像中单条CCD（电荷耦合器件）通道的宽度，即用于拼接为完整RDR图像的原始EDR数据产品的宽度。 此过程中会引入少量标签噪声：例如，来自几乎完全被尘埃覆盖的图像的图像块，可能恰好拍到了清晰的地表；而来自几乎无尘埃覆盖的图像的图像块，可能包含部分尘埃区域，或是地表无特征、外观类似尘埃覆盖区的区域。 数据集附带一组“审核标签”，包含作者对验证集部分图像块的人工标注结果，可用于估算数据集中的表观标签噪声率。 对应RDR图像块数据集，我们从同一批图像中提取了“有尘埃覆盖”与“无尘埃覆盖”类别的EDR图像块。EDR是仪器采集的原始图像，未经过校准或拼接。 为实现一定程度的归一化，EDR图像块仅从EDR的下半部分提取，上半部分用于对下半部分进行基础校准：具体操作为，减去上半部分的列（图像列）均值以消除“条纹伪影”，随后计算上半部分剩余像素值的0.1%与99.9%分位数，将图像块的像素值拉伸至该范围内的[0, 255] 8位整数值区间。该校准流程模拟了航天器在轨观测时可执行的校准过程，即利用先采集的图像上半部分，对后采集的图像下半部分进行校准。 全分辨率EDR图像宽度为1024像素，提取后会被缩放至100×100像素的图像块，以匹配RDR浏览图像生成的图像块的分辨率。 归档文件内容 压缩归档包含两个顶级目录，内容结构相似，分别为`edr_nomap_full_resized`与`rdr_nomap_browse`。前者包含基于EDR数据构建的数据集，后者包含基于RDR数据构建的数据集。 每个目录下均设有`dusty`与`not_dusty`文件夹，分别存储对应类别的图像块，此外还包含`manifest.csv`与`vetting_labels.csv`两个文件。 `vetting_labels.csv`文件包含一系列人工标注样本，附带原始标签，便于计算标签噪声率。`manifest.csv`文件则列出了所有样本、其标签以及所属的训练、验证或测试集。 样本ID编码了该图像块从原始HiRISE图像中采样的位置信息。以RDR数据集的ID`003100_PSP_004440_2125_r4805_c512`为例，其各部分含义如下： - `003100`：唯一数值ID - `PSP_004440_2125`：HiRISE观测ID - `r4805_c512`：图像块的左上角起始于原始观测图像的第4805行与第512列 对于EDR数据集的ID`200000_PSP_004530_1030_RED7_1_r9153`，其各部分拆解如下： - `200000`：唯一数值ID - `PSP_004530_1030`：HiRISE观测ID - `RED7`：CCD标识 - `1`：CCD通道编号（仅为0或1） - `r9153`：图像块从第9153行起始，由于1024像素宽的EDR图像的所有列均被使用，未指定列号，默认列起始为0 原始数据 原始HiRISE EDR与RDR数据可通过PDS（行星数据系统，Planetary Data System）获取，托管地址为https://hirise-pds.lpl.arizona.edu/PDS/
