DeepStat WP5 Solar Panel Dataset

The DeepSolaris/ DeepGeoStat dataset was developed in two projects with the same names, during the period 2017 - 2022. The dataset contains 20m x 20m cutouts from an aerial image of the Netherlands. Multiple annotators labeled each cutout for the presence or absence of solar panels. The dataset can be used to train machine or deep learning models on a solar panel classification task. The dataset contains cutouts from the 2018 high-resolution (DOP 10) aerial image of the Netherlands. Each cutout is 200x200 pixels which translates to a 20m x 20m grid. Cutouts were sampled from two nearby regions in the Netherlands: 1. The region around the city of <em>Heerlen</em> 2. The region around the city of <em>Valkenburg</em> Between both regions, a 100m strip was left free in which no cutouts were sampled. Both image subsets are therefore independent of each other. In total 26 annotators, annotated 63258 unique pictures with 200795 annotations; 34523 positive annotations indicating that the annotator saw a solar panel in the image, and 166272 negatives indicating that the annotator didn't see a solar panel in the image. More information can be found in the README.md in the archive. Python source code for training models on this dataset can be found here: https://gitlab.com/CBDS/deepgeostat-wp-5. PyTorch weights for a trained model can be found here: https://zenodo.org/record/7547703#.Y8l0phPMLap. This research was conducted under: the ESS action 'Merging Geostatistics and Geospatial Information in Member States' (grant agreement no.: 08143.2017.001-2017.408), under the ESS topic B5674-2020-GEOS (project 101033951 2020-NL-GEOS-DEEP-GEO-STAT), an investment of [Statistics Netherlands](https://www.cbs.nl) for the development of Deep Learning models, practices, and methodology. The researchers want to furthermore thank everyone involved in helping to create and annotate this dataset.DeepSolaris/DeepGeoStat数据集由两个同名项目于2017年至2022年间开发。该数据集包含荷兰航拍图像中截取的20米×20米图像块，多名标注人员对每个图像块是否存在太阳能板进行标注，可用于训练面向太阳能板分类任务的机器学习或深度学习模型。该数据集的图像块源自荷兰2018年的高分辨率（DOP 10）航拍图像，每个图像块为200×200像素，对应20米×20米的实地网格。图像块采样自荷兰境内两处相邻区域：1. 海伦（Heerlen）市周边区域；2. 瓦尔肯堡（Valkenburg）市周边区域。两处区域之间留有100米宽的无采样带，因此两个图像子集相互独立。总计26名标注人员对63258张独特图像完成了200795条标注：其中34523条为正样本标注，表示标注人员在图像中观测到太阳能板；166272条为负样本标注，表示未观测到太阳能板。更多详细信息可参见归档文件中的README.md。用于本数据集模型训练的Python源代码可访问：https://gitlab.com/CBDS/deepgeostat-wp-5。预训练模型的PyTorch权重文件可访问：https://zenodo.org/record/7547703#.Y8l0phPMLap。本研究依托欧洲统计系统（ESS）项目“在成员国中融合地统计学与地理空间信息”（资助协议编号：08143.2017.001-2017.408），以及ESS主题B5674-2020-GEOS项目（项目编号：101033951 2020-NL-GEOS-DEEP-GEO-STAT）开展，同时获得荷兰统计局（Statistics Netherlands）的资助，用于开发深度学习模型、实践方法与技术体系。研究团队谨向所有参与本数据集创建与标注工作的人员致以诚挚谢意。