Annotated High-Resolution Aerial Imagery of the Dutch Landscape for Solar Panel Detection and Segmentation

Overview This dataset consists of true-ortho high-resolution aerial images captured in 2023 by the local government of Emmen, Netherlands for the purpose of solar panel segmentation. The images are taken under similar conditions using a small aircraft, providing a significantly higher spatial resolution (7.5 cm per pixel) compared to satellite imagery (30 cm per pixel). This results in sharper and more detailed images suitable for solar panel detection and related spatial analyses. Data Collection and Processing The aerial images cover an area of 346.26 km², with four selected regions comprising diverse building types and vegetation, totaling 18.55 km². The selected areas are divided into 30x30 meter grid cells, resulting in 20,618 squares of 900 m² each. Within these areas, solar panels were manually annotated with polygons, identifying 4,389 unique solar panel objects. Given that the proportion of solar panel surface relative to the total area is small, the dataset includes only 224x224 pixel RGB images from grid cells that either contain solar panels or are in close proximity to a cell with solar panels. This selection avoids a significant class imbalance. The final dataset consists of 5,327 annotated images, of which 1,743 contain solar panels. For our research purposes, the dataset is enriched with elevation and slope data from the (also publicly available) Actueel Hoogtebestand Nederland (AHN) dataset: Elevation Data (AHN4 DEMs and LiDAR-derived Point Cloud Data): AHN4 provides precise elevation measurements with a minimum of 10 measurements per square meter. The digital terrain model (DTM) was generated using a Squared Inverse Distance Weighting (IDW) method. Slope Calculation: The tilt of surfaces was computed based on the AHN-4 dataset using the Planar Method. Dataset Structure The dataset is organized into the following directories: /input -- Stores the raw aerial images - 000000.png - 000001.png /mask -- Pixel-wise ground truth masks indicating the presence of solar panels - 000000.png (corresponds to input 0000000.png) - 000001.png /height -- Elevation data derived from the AHN4 dataset for each pixel - 000000.tiff - 000001.tiff /slope -- Slope values computed using a 3 × 3 sliding window for each pixel - 000000.tiff - 000001.tiff /annotations -- Contains meta information about the annotations - annotations.shp -- the annotations in polygon form - grids.shp -- the grid cells indicating the whole area that we annotated and contains for each cell whether it was included fold_info.pkl -- Python dictionary containing indices for stratified 5-fold cross-validation Code The code for producing baseline deep learning models on this dataset can be found at the following Gitlab repository. How to cite If you want to cite this work, please cite our underlying paper found here.概述 本数据集包含2023年由荷兰埃门（Emmen）当地政府为太阳能光伏面板分割任务采集的真正射（true-ortho）高分辨率航空影像。该系列影像采用小型航空器在统一采集条件下拍摄，空间分辨率达每像素7.5厘米，远高于卫星影像的每像素30厘米，因此影像更为清晰细腻，适用于光伏面板检测及相关空间分析任务。 数据采集与处理 航空影像覆盖总面积346.26平方千米，其中选取了包含多样建筑类型与植被的4个区域，总面积达18.55平方千米。选取区域被划分为30米×30米的网格单元，共得到20618个面积为900平方米的网格。在此范围内，研究人员采用多边形对光伏面板进行人工标注，共识别出4389个独立的光伏面板对象。 考虑到光伏面板总面积占比极低，本数据集仅选取包含光伏面板或紧邻含光伏面板网格的单元，裁剪为224×224像素的RGB影像，以此避免严重的类别不平衡问题。最终数据集共包含5327张标注影像，其中1743张含有光伏面板。 为满足研究需求，本数据集补充了同样公开可用的荷兰最新高程数据集（Actueel Hoogtebestand Nederland, AHN）中的高程与坡度数据： 高程数据（AHN4数字高程模型（DEMs）与激光雷达（LiDAR）点云数据）：AHN4提供高精度高程测量数据，每平方米至少包含10个测量点。研究人员采用平方反比距离加权（Squared Inverse Distance Weighting, IDW）方法生成了数字地形模型（DTM）。 坡度计算：基于AHN4数据集，采用平面法（Planar Method）计算地表倾斜度。 数据集结构 本数据集的目录结构如下： /input：存储原始航空影像，示例文件：000000.png、000001.png /mask：存储逐像素的真值掩膜，用于标识光伏面板的存在位置，示例文件：000000.png（与input目录下的0000000.png对应）、000001.png /height：存储每个像素对应的AHN4高程数据，示例文件：000000.tiff、000001.tiff /slope：存储采用3×3滑动窗口计算得到的各像素坡度值，示例文件：000000.tiff、000001.tiff /annotations：存储标注元信息，包括： - annotations.shp：多边形格式的标注数据 - grids.shp：标注区域的网格单元矢量文件，用于标识每个网格是否被纳入数据集 fold_info.pkl：存储用于分层5折交叉验证的索引的Python字典对象 代码 可在以下Gitlab仓库中获取用于在本数据集上构建基准深度学习模型的代码。 引用方式 若需引用本数据集相关工作，请引用此处列出的研究论文。