High-Resolution Remote Sensing Building Dataset

WHU Aerial Building Dataset : Collected and created by the Photogrammetry and Computer Vision Research Group of Wuhan University, this dataset is sourced from regions including New Zealand, covering an area of over 450 km². It has a spatial resolution of 0.3 m and contains 8,188 optical remote sensing images cropped to 512 × 512 pixels, along with their corresponding annotated images. The entire dataset is divided into a training set, a validation set, and a test set, with 4,736, 1,036, and 2,416 images respectively. In this dataset, the annotations for building and background categories are clear; however, it presents several typical challenges: for instance, dense building distribution leads to mutual occlusion and blurred boundaries, impairing recognition accuracy; shadow areas caused by illumination make it difficult to extract building features; small-scale buildings have a low pixel ratio and weak features, making them easily overlooked by the model.

Massachusetts Building Dataset : Collected and organized by the Massachusetts Institute of Technology (MIT), this dataset comprises 151 aerial images from the Boston area. Each image has a size of 1500 × 1500 pixels and a spatial resolution of 1 m, with the entire dataset covering approximately 340 km². The data is divided into a training set (137 images), a test set (10 images), and a validation set (4 images). During the experiment, data augmentation was performed on the original images through operations such as cropping, horizontal flipping, and vertical flipping. Typical challenges in this dataset include: annotation errors due to blurred building boundaries, occlusions by nearby trees, and variations in roof texture and color. These factors collectively increase the difficulty of accurate building boundary segmentation.

GF-7 Building Dataset : This dataset is derived from imagery of the Gaofen-7 Satellite, covering 6 typical cities in China (including Tianjin, Chongqing, Guangzhou, among others). It has a total coverage area of 573.17 km², consisting of 5,175 images with each image sized 512 × 512 pixels and a spatial resolution of 0.65 m. The annotations include 170,015 buildings, of which 84.8% are urban buildings and 15.2% are rural buildings. The dataset is divided into three subsets at a ratio of 6:2:2, with 3,106 images used for training, 1,034 for validation, and 1,035 for testing. In this dataset, the annotations of building outlines and background features have undergone manual correction combined with automated algorithm optimization, resulting in high boundary accuracy. Its typical challenges include: the superposition of rain and cloud shadows and building shadows, which interferes with feature extraction; rural small-scale buildings have a low pixel ratio and are difficult to distinguish from vegetation.