Zurich Summer Dataset

The "Zurich Summer v1.0" dataset is a collection of 20 chips (crops), taken from a QuickBird acquisition of the city of Zurich (Switzerland) in August 2002. QuickBird images are composed by 4 channels (NIR-R-G-B) and were pansharpened to the PAN resolution of about 0.62 cm GSD. We manually annotated 8 different urban and periurban classes : Roads, Buildings, Trees, Grass, Bare Soil, Water, Railways and Swimming pools. The cumulative number of class samples is highly unbalanced, to reflect real world situations. Note that annotations are not perfect, are not ultradense (not every pixel is annotated) and there might be some errors as well. We performed annotations by jointly selecting superpixels (SLIC) and drawing (freehand) over regions which we could confidently assign an object class.

The dataset is composed by 20 image - ground truth pairs, in geotiff format. Images are distributed in raw DN values. We provide a rough and dirty MATLAB script (preprocess.m) to:

i) extract basic statistics from images (min, max, mean and average std) which should be used to globally normalize the data (note that class distribution of the chips is highly uneven, so single-frame normalization would shift distribution of classes).

ii) Visualize raw DN images (with unsaturated values) and a corresponding stretched version (good for illustration purposes). It also saves a raw and adjusted image version in MATLAB format (.mat) in a local subfolder.

iii) Convert RGB annotations to index mask (CLASS \in {1,...,C}) (via rgb2label.m provided).

iv) Convert index mask to georeferenced RGB annotations (via rgb2label.m provided). Useful if you want to see the final maps of the tiles in some GIS software (coordinate system copied from original geotiffs).

Some requests from you

We encourage researchers to report the ID of images used for training / validation / test (e.g. train: zh1 to zh7, validation zh8 to zh12 and test zh13 to zh20). The purpose of distributing datasets is to encourage reproducibility of experiments.

Acknowledgements

We release this data after a kind agreement obtained with DigitalGlobe, co. This data can be redistributed freely, provided that this document and corresponding license are part of the distribution. Ideally, since the dataset could be updated over the time, I suggest to distribute the dataset by the official link from which this archive has been downloaded.

We would like to thank (a lot) Nathan Longbotham @ DigitalGlobe and the whole DG team for his / their help for granting the distribution of the dataset.

We release this dataset hoping that will help researchers working in semantic classification / segmentation of remote sensing data in comparing to other state-of-the-art methods using this dataset as well in testing models on a larger and more complete set of images (with respect to most benchmarks available in our community). As you can imagine, it has been a tedious work in preparing everything. Just for you.

If you are using the data please cite the following work

Volpi, M. & Ferrari, V.; Semantic segmentation of urban scenes by learning local class interactions, In IEEE CVPR 2015 Workshop "Looking from above: when Earth observation meets vision" (EARTHVISION), Boston, USA, 2015.

"苏黎世夏季v1.0（Zurich Summer v1.0）"数据集包含20张裁剪瓦片，均取自2002年8月瑞士苏黎世市的QuickBird遥感影像。该QuickBird影像包含4个波段（近红外-红-绿-蓝，NIR-R-G-B），并经全色锐化处理至全色波段分辨率，地面采样距离约为0.62厘米（Ground Sampling Distance, GSD）。我们手动标注了8类城市及城郊地物类别：道路、建筑、树木、草地、裸土、水体、铁路与游泳池。各类别样本总数存在显著不均衡性，以贴合真实世界的地物分布现状。需注意，本次标注并非完美无缺，亦未进行超密标注（并非每个像素均被标注），且可能存在少量错误。标注过程通过联合选取超像素（SLIC）与在可明确判定类别的区域进行手绘描边的方式完成。

本数据集包含20组图像-真值标签对，格式为地理标记图像文件格式（GeoTIFF）。影像以原始DN值（Digital Number, DN）发布。我们提供了一份简易粗糙的MATLAB脚本（preprocess.m），可实现以下功能：
i) 从影像中提取基础统计量（最小值、最大值、均值与平均标准差），用于对数据进行全局归一化（需注意：裁剪瓦片的类别分布极不均衡，单帧归一化会改变各类别的实际分布）。
ii) 可视化原始DN值影像（未饱和状态下）及其拉伸增强后的版本（便于演示展示），同时会在本地子文件夹中保存MATLAB格式（.mat）的原始与调整后的影像版本。
iii) 将RGB格式的标注转换为索引掩码（类别索引范围为{1,…,C}）（通过提供的rgb2label.m脚本实现）。
iv) 将索引掩码转换为带地理参考的RGB格式标注（同样通过提供的rgb2label.m脚本实现），若需在GIS软件中查看瓦片的最终分类图（坐标系沿用原始GeoTIFF的坐标系），该功能将十分实用。

致使用者
我们鼓励研究者注明训练/验证/测试集所使用的图像ID（例如：训练集：zh1至zh7，验证集：zh8至zh12，测试集：zh13至zh20）。发布本数据集的核心目的在于推动实验的可复现性。

致谢
本数据集是在与DigitalGlobe公司达成友好合作协议后发布的。只要随分发包一并包含本说明文档与对应的授权协议，即可自由再分发本数据集。考虑到数据集未来可能进行更新，我们建议通过本档案的官方下载链接来分发该数据集。

我们衷心感谢DigitalGlobe的Nathan Longbotham及其整个团队，感谢他们为数据集的分发许可提供的大力支持。

我们发布本数据集，希望能够帮助从事遥感数据语义分类/分割研究的学者，通过本数据集与其他前沿方法进行对比实验，同时也可借助本数据集规模更大、样本更完整的影像集（相较于学界现有多数基准数据集）来测试模型。想必诸位也能理解，筹备整套数据集是一项繁琐的工作，谨以此献给各位研究者。

若您使用本数据集，请引用以下文献：
Volpi, M. 与 Ferrari, V.；《基于局部类别交互学习的城市场景语义分割》，发表于2015年IEEE CVPR研讨会"俯瞰万物：地球观测与计算机视觉的融合"（EARTHVISION），美国波士顿，2015年。