高分辨率遥感图像细粒度目标识别数据集（FAIR1M_V2.0）

该数据集含15000余幅分辨率优于1米、尺寸从上千到上万像素不等的图像，具有100多万精细化标注、多角度分布的目标，场景覆盖全球上百个典型城市、乡镇，以及常用机场、港口等，其中来自我国自主产权高分系列卫星的数据占比超过80%，标注结果均经权威判读专家确认。
　　根据遥感应用的实际需求，该数据集将目标（包含飞机、船舶、车辆、球场和道路等）进行类别的细分，包括37个细粒度类别。例如，依据型号将飞机细分为波音式（波音737、747、777和787等）、空客式（空客220、321、330和350等）以及国产式飞机（C919和ARJ21等）；将船舶细分为液货船、干货船、渔船、邮轮、拖船和工程船等；对于车辆、球场和道路等要素也分别按照功能、尺寸等细分至多个类型。
　　同时，考虑到不同时间拍摄的遥感图像中蕴含更为丰富的目标关联信息，该数据集还提供了针对同一区域、不同时相的数据，是一套多时相、多分辨、多要素的遥感图像标准化样本集。相较于1.0数据新增了验证集，拓展了测试集。

This dataset contains over 15,000 images with resolution better than 1 meter and pixel sizes ranging from thousands to tens of thousands. It has more than 1 million finely annotated, multi-angle distributed objects. The scene coverage includes hundreds of typical cities, towns around the world, as well as common airports, ports, etc. Among them, data from China's self-owned Gaofen series satellites accounts for more than 80%, and all annotation results have been verified by authoritative interpretation experts.

Based on the actual needs of remote sensing applications, this dataset subdivides objects (including aircraft, ships, vehicles, stadiums, roads, etc.) into categories, including 37 fine-grained categories. For example, aircraft are subdivided into Boeing-type (Boeing 737, 747, 777, 787, etc.), Airbus-type (Airbus 220, 321, 330, 350, etc.) and domestic aircraft (C919, ARJ21, etc.) by model; ships are subdivided into liquid cargo ships, dry cargo ships, fishing boats, cruise ships, tugboats, engineering ships, etc.; vehicles, stadiums, roads and other elements are also subdivided into multiple types according to function, size, etc.

In addition, considering that remote sensing images taken at different times contain more abundant object correlation information, this dataset also provides data for the same area at different time phases. It is a standardized sample set of remote sensing images with multi-temporal, multi-resolution and multi-element features. Compared with version 1.0 data, a validation set has been added and the test set has been expanded.

搜集汇总

背景与挑战

背景概述

FAIR1M_V2.0是一个高分辨率遥感图像细粒度目标识别数据集，包含超过15000张分辨率优于1米的图像，标注了超过100万个目标，细分37个类别（如飞机、船舶、车辆等），并基于型号或功能进一步分类。该数据集覆盖全球典型区域，超过80%数据源自中国自主卫星，标注经专家确认，且提供多时期数据以增强目标关联分析，适用于深度学习遥感应用。