遥感影像地物分割样本数据集(Zurich Summer Dataset)

该数据集是 2002 年 8 月从 QuickBird 采集的苏黎世市（瑞士）图像中提取的 20 个片段（作物）的集合。QuickBird 图像由 4 个通道（近红外-红外-近红外-红外）组成，并经过平差处理，达到约 0.62 厘米 GSD 的 PAN 分辨率。我们人工标注了 8 个不同的城市和近郊区类别： 道路、建筑物、树木、草地、裸土、水域、铁路和游泳池。类别样本的累积数量非常不平衡，以反映真实世界的情况。需要注意的是，注释并不完美，也不是超密度的（并非每个像素都有注释），而且还可能存在一些错误。我们通过联合选择超像素（SLIC）和在我们可以确定对象类别的区域内绘图（徒手绘制）来进行标注。数据集由 20 对 geotiff 格式的图像和地面实况组成。图像以原始 DN 值分布。

This dataset consists of 20 image crops extracted from a QuickBird satellite image of Zurich, Switzerland, acquired in August 2002. The QuickBird image features 4 spectral bands (near-infrared, infrared, near-infrared, infrared) and has undergone adjustment processing to achieve a panchromatic (PAN) resolution with a Ground Sampling Distance (GSD) of approximately 0.62 cm. Eight distinct urban and suburban land cover classes were manually annotated: roads, buildings, trees, grassland, bare soil, water bodies, railways, and swimming pools. The cumulative sample counts across these classes are highly imbalanced to reflect real-world conditions. It should be noted that the annotations are not perfect, nor are they fully dense (not every pixel has an associated annotation), and some annotation errors may exist. Annotations were generated via a combined method of selecting superpixels (using SLIC) and drawing freehand within regions where the object class could be unambiguously identified. The dataset comprises 20 pairs of images and their corresponding ground truth masks, both stored in GeoTIFF format. The pixel values of the images follow their original Digital Number (DN) distribution.

搜集汇总

背景与挑战

背景概述

该数据集是2002年8月瑞士苏黎世市的QuickBird遥感影像地物分割样本集，包含20个图像-地面真值对，分辨率为0.61米，标注了8个城市和城郊类别（如道路、建筑、树木），类别样本不平衡以模拟真实场景。数据用于支持城市土地利用分类和计算机视觉研究，但标注可能存在不完美之处，适用于地理信息系统和模式识别应用。