Automatic Crater Detection Tool for Moon, Mars, and Mercury

Impact craters are fundamental in planetary science, providing insights into the geological evolution of planetary surfaces and influencing landing site selection. However, many crater extraction algorithms are either not open-sourced or require complex configurations. To address these issues, we present the Automatic Crater Detection (ACD) Tool, a user-friendly software designed to support scientists, even those without a computational background.Key FeaturesThe ACD Tool utilizes the YOLOv8 deep learning algorithm, enhanced by transfer learning and active learning techniques. This integration enables the detection of craters with a minimum diameter of 10 pixels on terrestrial planets. The tool has undergone rigorous testing on datasets from the Moon, Mars, and Mercury, achieving F1 scores of up to 95%, making it reliable for tasks like surface age estimation and landing site analysis.User Interface and OperationThe interface of the ACD tool is divided into two parts: the left side serves as the input and control panel, while the right side functions as the display area. Users can select among three celestial bodies – Mars, Moon, or Mercury – by clicking on the respective options. The software offers two input modes. In "single image" mode, users directly enter the image paths. In "batch" mode, the program processes all TIFF files within a specified folder. Input images should be GeoTIFF files in Mercator projection to ensure circular shapes for impact craters.The automatic crater detection process begins with a click of the “Start” button.The right-hand viewing interface shows a preview image displaying the crater detection results. In batch mode, the “Previous Image” and “Next Image” buttons allow sequential preview of all images and detection results.Data and OutputInput images should be in GeoTIFF format with Mercator projection centered on the equator to ensure accurate detection.Output results are saved in the specified output path in both shapefile and TXT file formats. The attribute table of the shapefile includes the crater central latitude and longitude (‘Lon’ and ‘Lat’) and the geodesic diameter in meters (‘Dia’). The projection distortion caused by projection has been considered and corrected through computation within the tool. As the craters are extracted in Mercator projection, the corresponding information is provided to better correspond with the images: ‘X_Merc’ and ‘Y_Merc’ for the x and y coordinates in Mercator and ‘D_Merc’ for the distorted diameter measured in the Mercator projection. The confidence score (‘Score’), ranging from 0 to 1, is also provided; higher values indicate greater confidence in the crater detection results. Additionally, a TXT file containing information identical to the shapefile attributes is generated. If an error occurs during the detection process, the names of the erroneous images and the reasons are recorded in an Error.txt file.VersionThe ACD Tool runs on Windows and offers both CPU and GPU versions, allowing users to choose the version that best fits their needs. For faster performance, the GPU version is recommended if a graphics card is available. Otherwise, the CPU version is suitable for systems without a dedicated graphics card.DatasetsOur dataset includes annotated craters on the Moon, Mars, and Mercury. The initial annotations were sourced from publicly available online databases and were further supplemented with manual annotations. For the Moon, LROC WAC images (Speyerer, 2011; Wagner, 2015) and SELENE TC images (Kato et al., 2008) were annotated based on the studies of Robbins (2019) and Wang & Wu (2019), respectively, while LROC NAC images were fully manually annotated. For Mars, crater annotations for the THEMIS-IR mosaic (Edwards et al., 2011) followed the work of Robbins & Hynek (2012), and additional manual annotations were applied to CTX (Malin et al., 2007) and HiRISE (McEwen et al., 2007) images. For Mercury, the MDIS global mosaic (Becker et al., 2009; Hawkins III et al., 2009) underwent complete manual annotation and verification.The dataset includes craters of various sizes and resolutions. On the Moon, the LROC WAC dataset has a resolution of 150 m/pixel, covering 186,302 craters with diameters ranging from 1,000 to 46,692 meters. The LROC NAC dataset, with a resolution of 1.1 m/pixel, includes 34,012 craters with diameters between 7 and 307 meters. The SELENE TC dataset has a resolution of 6.2 m/pixel and includes 54,563 craters with diameters between 120 and 1,949 meters. For Mars, the THEMIS-IR dataset has a resolution of 150 m/pixel and includes 121,517 craters with diameters ranging from 994 to 47,322 meters. The CTX dataset has a resolution of 5 m/pixel and includes 6,678 craters with diameters between 21 and 2,467 meters. The HiRISE dataset, with a resolution of 0.5 to 1 m/pixel, includes 5,978 craters with diameters ranging from 2 to 144 meters. For Mercury, the MDIS dataset has a resolution of 166 m/pixel and includes 17,749 manually annotated craters with diameters ranging from 996 to 26,347 meters.The dataset is organized by body, stored in YOLO annotation format, and located in the respective subfolders within the “Datasets” directory. Each body folder contains “images” and “labels” folders, which store crater images and position annotations, respectively. Within each folder, data is further divided into “train” and “val” subfolders for training and validation.

撞击坑是行星科学领域的基础性研究对象，可为行星表面地质演化研究提供关键视角，同时对着陆点选址工作具有重要指导意义。然而，当前多数撞击坑提取算法要么未开源，要么配置流程复杂繁琐。为解决上述问题，我们推出了自动撞击坑检测（Automatic Crater Detection, ACD）工具——一款面向科研人员（即使无计算背景）设计的易用型软件。

### 核心功能
本工具基于YOLOv8深度学习算法开发，并结合迁移学习与主动学习技术进行优化。该方案可在类地行星影像中检测直径最小为10像素的撞击坑。团队已在月球、火星与水星的公开数据集上对本工具开展了严格测试，其F1得分最高可达95%，可可靠应用于表面年龄估算、着陆点分析等科研任务。

### 用户界面与操作流程
ACD工具的界面分为左右两个区域：左侧为输入与控制面板，右侧为结果显示区域。用户可通过点击对应选项，在火星、月球、水星三类天体中选择检测目标。软件支持两种输入模式：在「单张图像」模式下，用户可直接输入影像路径；在「批量处理」模式下，程序将自动处理指定文件夹内的所有TIFF文件。输入影像需为墨卡托投影（Mercator projection）下的GeoTIFF文件，以确保撞击坑的圆形形态保真。
启动自动撞击坑检测仅需点击「开始」按钮。右侧预览界面将实时展示撞击坑检测结果。在批量处理模式下，用户可通过「上一张」「下一张」按钮依次预览所有影像的检测结果。

### 数据与输出要求
输入影像需为以赤道为中心的墨卡托投影GeoTIFF文件，以保障检测精度。
检测结果将以形状文件（shapefile）与文本文件（TXT）两种格式保存至指定输出路径。形状文件的属性表包含撞击坑中心经纬度（字段名分别为`Lon`与`Lat`）、以米为单位的测地直径（字段名`Dia`）。工具内部已通过计算校正了投影带来的畸变：由于撞击坑提取基于墨卡托投影，工具同时提供了与影像匹配的投影坐标信息——墨卡托投影下的x、y坐标（字段名`X_Merc`与`Y_Merc`），以及墨卡托投影下测得的畸变直径（字段名`D_Merc`）。此外还包含置信度得分（字段名`Score`），取值范围为0至1，分值越高代表撞击坑检测结果的可信度越强。
同时将生成一份与形状文件属性表内容完全一致的TXT文件。若检测过程中出现错误，出错影像的文件名与错误原因将被记录至`Error.txt`文件中。

### 版本信息
ACD工具支持Windows操作系统，提供CPU与GPU两个版本，用户可根据自身设备配置选择适配版本。若配备独立显卡，推荐使用GPU版本以获得更快的运行性能；若无独立显卡，则可使用CPU版本适配普通桌面系统。

### 数据集说明
本数据集包含月球、火星与水星上的标注撞击坑数据。初始标注数据来源于公开在线数据库，后续通过人工标注进行了补充完善。
针对月球数据集：分别基于Robbins(2019)与Wang & Wu(2019)的研究，对LROC WAC影像（Speyerer, 2011; Wagner, 2015）与SELENE TC影像（Kato et al., 2008）进行标注；同时对LROC NAC影像开展了全量人工标注。
针对火星数据集：THEMIS-IR镶嵌影像（Edwards et al., 2011）的撞击坑标注参考了Robbins & Hynek(2012)的工作，并针对CTX（Malin et al., 2007）与HiRISE（McEwen et al., 2007）影像补充了人工标注数据。
针对水星数据集：MDIS全球镶嵌影像（Becker et al., 2009; Hawkins III et al., 2009）已完成全量人工标注与校验。
本数据集涵盖多种尺寸与分辨率的撞击坑数据：
- 月球数据集：LROC WAC数据集分辨率为150 m/像素，包含186302个直径介于1000米至46692米的撞击坑；LROC NAC数据集分辨率为1.1 m/像素，包含34012个直径介于7米至307米的撞击坑；SELENE TC数据集分辨率为6.2 m/像素，包含54563个直径介于120米至1949米的撞击坑。
- 火星数据集：THEMIS-IR数据集分辨率为150 m/像素，包含121517个直径介于994米至47322米的撞击坑；CTX数据集分辨率为5 m/像素，包含6678个直径介于21米至2467米的撞击坑；HiRISE数据集分辨率为0.5至1 m/像素，包含5978个直径介于2米至144米的撞击坑。
- 水星数据集：MDIS数据集分辨率为166 m/像素，包含17749个人工标注的撞击坑，直径范围介于996米至26347米。
本数据集按天体分类组织，以YOLO标注格式（YOLO annotation format）存储，存放于「Datasets」目录下的对应子文件夹中。每个天体文件夹包含「images」与「labels」子文件夹，分别存储撞击坑影像与位置标注信息。每个子文件夹内进一步划分为「train」（训练集）与「val」（验证集）子目录。

搜集汇总

背景与挑战

背景概述

该数据集是一个用于月球、火星和水星陨石坑检测的工具，基于YOLOv8深度学习算法，支持多种分辨率和尺寸的陨石坑检测，并提供详细的输出结果和错误记录。数据集包含多个天体的陨石坑图像和标注，适用于表面年龄估计和着陆点分析等任务。