Robinson Ridge Manual and Semi-Automated Vegetation Labels with UAV Orthomosaics

This dataset supports the development of machine learning models for vegetation segmentation in Antarctic ecosystems and enables reproducible remote sensing analyses by providing both imagery and associated ground-truth labels in standard GIS-compatible formats. 1. Ground Truth Labels This record contains manually and semi-automatically generated pixelwise shapefiles, used to annotate vegetation types with high confidence. This component contains both manually and semi-automatically generated pixel-wise shapefiles classifying four vegetation classes: Usnea spp., black lichen, moss, and non-vegetation. The purpose of these labels is to provide high-quality reference data for training and validating machine learning models for vegetation segmentation. Manual labelling was conducted using ground truth points gathered during field campaigns. Polygons were drawn around confidently identified vegetation patches, focusing on central areas to minimise edge misclassification. (Files: .shp, .dbf, .prj, .shx, and readme.txt) Semi-automatic labelling was introduced to address the limitations of manual annotation at a GSD of 2.93 cm/pixel. A suite of 28 vegetation indices (VIs) were evaluated, and three (MSAVI and GNDVI) were selected for their spectral separability between classes. (Files: .shp, .dbf, .prj, .shx, and readme.txt) 2. Orthomosaic The RGB and multispectral (MS) images captured by UAVs were processed into high-resolution orthomosaics using Agisoft Metashape 1.6.6 and georeferenced via QGIS 3.2.0. These orthomosaics serve as the base imagery for generating labels and training machine learning models. Output formats include .tif for orthomosaics, .ovr for overviews, and .pdf reports documenting the image processing steps. Data Collection and Analysis Imagery was captured in January 2023 at Robinson Ridge, Antarctica, using a BMR3.9RTK UAV equipped with a MicaSense Altum sensor and Sony Alpha 5100 camera, flown at 70 m altitude (GSD ≈ 2.93 cm/pixel). Over 2,800 images were collected over ~5.15 ha. Usage Notes · Embargoed files require permission for access. · Refer to the included readme.txt files in each record for file structure, formats, and usage instructions. · The dataset is optimised for developing and validating deep learning models for remote sensing classification in polar environments.

本数据集可支撑南极生态系统植被分割机器学习模型的开发，并通过提供标准地理信息系统（GIS, Geographic Information System）兼容格式的影像及配套真值标签（ground-truth labels），实现可复现的遥感分析工作。

1. 真值标签
本数据集包含手动与半自动生成的逐像素矢量形状文件（Shapefile），用于高置信度标注植被类型。该组件涵盖两类逐像素形状文件，可对四类植被类别进行分类：松萝属（Usnea spp.）、黑色地衣、苔藓及非植被区域。此类标签旨在为植被分割机器学习模型的训练与验证提供高质量参考数据。
手动标注基于野外作业采集的真值点开展，研究人员在经确认的植被斑块周边绘制多边形，并优先选取斑块中心区域以降低边缘误分类概率。（配套文件：.shp、.dbf、.prj、.shx及readme.txt）
半自动标注用于解决手动标注在2.93 cm/像素的地面采样距离（Ground Sampling Distance, GSD）下存在的局限性。研究团队评估了28种植被指数（Vegetation Indices, VIs），最终选取3种（含修改型土壤调整植被指数MSAVI与归一化绿波段差值植被指数GNDVI），因其在各植被类别间具备优异的光谱可分性。（配套文件：.shp、.dbf、.prj、.shx及readme.txt）

2. 正射镶嵌影像
研究团队使用Agisoft Metashape 1.6.6将无人机（Unmanned Aerial Vehicle, UAV）采集的RGB及多光谱（MS）影像处理为高分辨率正射镶嵌图，并通过QGIS 3.2.0完成地理配准。此类正射影像可作为生成标签及训练机器学习模型的基础影像源。输出格式包括正射镶嵌图文件.tif、概览文件.ovr，以及记录影像处理流程的.pdf报告。

数据采集与分析
2023年1月，研究团队于南极洲罗宾逊山脊（Robinson Ridge）使用搭载MicaSense Altum传感器与Sony Alpha 5100相机的BMR3.9RTK无人机开展影像采集，飞行高度为70米，地面采样距离约为2.93 cm/像素。本次作业共采集超过2800张影像，覆盖面积约5.15公顷。

使用须知
· 受访问限制的文件需获得授权后方可获取。
· 请查阅各数据记录中附带的readme.txt文件，以了解文件结构、格式及使用说明。
· 本数据集专为极地环境遥感分类深度学习模型的开发与验证优化设计。