Rooftop Height Estimation from Orthophotos Using ML (Vienna use case)

This dataset contains materials supporting a machine learning pipeline developed for estimating rooftop building heights from orthophoto imagery in Vienna, Austria. The task is part of FAIRiCUBE Use Case 4 (UC4), which aims to assess urban building stock and geometry for energy and infrastructure analysis. The true rooftop heights were computed using LOD2.1 building data obtained from Vienna's open governmental portal. Each building's geometry was analyzed as MULTIPOLYGON Z objects, from which the highest z-value was extracted for each unique (x, y) coordinate. Rooftop heights were then determined by calculating the difference between the maximum and minimum heights within each polygon. The dataset includes: A set of cropped and masked rooftop images (used for training and inference) A CSV file with image filenames and corresponding reference rooftop heights A CSV file with predicted heights using the trained model A trained model file (.pkl) based on a VGG18 → XGBoost architecture A scatter plot showing predicted vs true heights (RMSE = 1.14 m, R² = 0.57) A poster summarizing the methodology and results Feature extraction is performed using VGG18, a convolutional neural network pretrained on ImageNet. These features are fed into an XGBoost regressor to predict continuous rooftop height values. The model was trained and tested on approximately 26,000 rooftop images from Vienna. This dataset enables reproducible research in geospatial ML, urban analytics, and building stock modeling.