GeoFormer: Source Code and Prediction Results for Scene-Level Building Height and Footprint Estimation

GeoFormer: A Swin Transformer-Based Framework for Scene-Level Building Height and Footprint Estimation from Sentinel Imagery. This deposit contains: geoformer_code.zip — Full source code including the GeoFormer model, baseline models (ResNet, SENet, UNet, ConvNeXt), training/testing scripts, ablation experiment scripts, and a data preparation script (prepare_h5_data.py) for reproducing H5 training data from SHAFTS rasters. geoformer_predictions.zip — Model prediction results (.npz) and evaluation metrics (.json) for all experiments, including modality ablation (No-SAR, No-Optical), receptive field comparison (3×3, 5×5, 9×9), and baseline model comparisons. Data Source Attribution: The training data used in this study is derived from the SHAFTS dataset (Zenodo DOI: 10.5281/zenodo.6370003), which provides global 100m building height and footprint rasters. Satellite inputs include Sentinel-1 SAR composites, Sentinel-2 optical composites (Copernicus/ESA), and SRTM DEM (NASA). This repository does NOT contain the original training data — users should obtain rasters from SHAFTS and use the provided prepare_h5_data.py script to generate HDF5 training files.