Aerial Building Classification Dataset (ABCD)

The dataset, entitled Aerial Building Classification Dataset (ABCD), was developed to support research on the classification of building types in high-resolution remote sensing imagery using deep learning methods. It consists of RGB aerial images acquired from an airborne platform, characterized by a spatial resolution of 0.25 m Ground Sampling Distance (GSD). The source data were derived from orthophotomaps provided by Geoportal.gov.pl, ensuring high geometric and radiometric quality. All images were standardized in terms of format and input size to ensure compatibility with convolutional neural network architectures.
The ABCD dataset includes six classes of building objects: single-family buildings, multi-family buildings, agricultural (farm) buildings, industrial and commercial buildings, public utility buildings, and greenhouses. In total, the dataset comprises 26,379 image patches, divided into training, validation, and test subsets to support reproducible machine learning experiments. The class distribution is highly imbalanced, with single-family buildings constituting the majority class (67.84%), while minority classes such as public buildings (1.25%) and greenhouses (1.01%) are significantly underrepresented.
This imbalance reflects real-world spatial distributions of building types but introduces additional challenges for model training, including the risk of bias toward dominant classes.
An additional challenge inherent to the dataset is the high visual similarity between certain classes, particularly between residential, agricultural, and industrial buildings, which increases the difficulty of the classification task. As a result, the dataset provides a demanding benchmark for advanced classification models and enables the assessment of their robustness under realistic conditions. Overall, the ABCD dataset constitutes a high-quality resource for the development and evaluation of deep learning approaches in aerial image classification, with particular relevance to class imbalance handling, fine-grained visual discrimination, and real-world geospatial applications.