This repository contains the dataset prepared for the master’s thesis:
Evaluating_Sentinel-2_Super-Resolution_Algorithms_for_Automated_Building_Delineation
The dataset includes high-resolution reference imagery, low-resolution Sentinel-2 data, multiple super-resolved outputs, and stratification tables for experiments. Additionally, sample subsets and figures are provided for illustration.
.
в”њв”Ђв”Ђ metadata.zip
в”‚ в”њв”Ђв”Ђ image_samples # Small illustrative subset of the dataset
в”‚ в”‚ в”њв”Ђв”Ђ building_delineation_inference # Predicted masks (Nearest Neighbour + Sen2SR_RGBN)
в”‚ в”‚ в”њв”Ђв”Ђ hr_mask # Ground truth masks (TIF)
в”‚ в”‚ в”њв”Ђв”Ђ lr_s2 # Sentinel-2 input (TIF)
в”‚ в”‚ в”њв”Ђв”Ђ orthophoto # Reference HR orthophotos
в”‚ в”‚ в””в”Ђв”Ђ super_resolved # Super-resolved images from models
в”‚ в”‚ в”њв”Ђв”Ђ deepsent
в”‚ в”‚ в”њв”Ђв”Ђ evoland
в”‚ в”‚ в”њв”Ђв”Ђ ldsr_s2
в”‚ в”‚ в”њв”Ђв”Ђ sen2sr_lite
в”‚ в”‚ в”њв”Ђв”Ђ sen2sr_rgbn
в”‚ в”‚ в”њв”Ђв”Ђ sr4rs
в”‚ в”‚ в””в”Ђв”Ђ swin2_mose
в”‚ в”‚
в”‚ в”њв”Ђв”Ђ stratification_tables # Train/val/test splits
в”‚ в”‚ в”њв”Ђв”Ђ filtered # Filtered subsets (CSV/GPKG)
в”‚ в”‚ в””в”Ђв”Ђ full # Full splits (CSV)
в”‚ в”‚
в”‚ в”њв”Ђв”Ђ thesis_figures # Figures used in the thesis
в”‚ в”‚
в”‚ в””в”Ђв”Ђ metric_results # All metric scores for SR models & interpolations
в”‚ в”њв”Ђв”Ђ all_sr_model_results.csv # Metrics for each image and model
в”‚ в”њв”Ђв”Ђ model_experiment.csv # Metrics used during model experimentation
в”‚ в”њв”Ђв”Ђ sr_mean_object_metrics.csv # Mean object detection metrics (all images)
в”‚ в”њв”Ђв”Ђ sr_mean_object_metrics_per_size.csv # Metrics grouped by building size
в”‚ в””в”Ђв”Ђ sr_model_results.csv.zip # Metrics aggregated per model
в”‚
в”њв”Ђв”Ђ building_delineation_inference.zip # Inferred building masks for all models & interpolation methods
в”‚
в”њв”Ђв”Ђ hr_mask.zip # Ground truth cadastral masks (2.5m)
в”њв”Ђв”Ђ hr_orthophoto.zip # High-res (2.5m) orthophotos
в”њв”Ђв”Ђ lr_s2.zip # Low-res (10m) Sentinel-2 inputs
в”‚
в”њв”Ђв”Ђ sr_deepsent.zip # DeepSent SR outputs (2.5m)
в”њв”Ђв”Ђ sr_evoland.zip # EvoLands SR outputs (2.5m)
в”њв”Ђв”Ђ sr_ldsr_s2.zip # LDSR SR outputs (2.5m)
в”њв”Ђв”Ђ sr_sen2sr_Lite.zip # Sen2SR Lite outputs (2.5m)
в”њв”Ђв”Ђ sr_sen2sr_RGBN.zip # Sen2SR RGBN outputs (2.5m)
в”њв”Ђв”Ђ sr_sr4rs.zip # SR4RS outputs (2.5m)
в”њв”Ђв”Ђ swin2_mose # Swin2-MoSE SR outputs (2.5m)
в”‚
в”њв”Ђв”Ђ tracasa.zip # Tracasa dataset for model evaluation
в”‚ в”њв”Ђв”Ђ processed_tracasa_images
в”‚ в””в”Ђв”Ђ tracasa_S2_SR.tif
в”‚
в””в”Ђв”Ђ README.md
High-Resolution Data STAC-compliant data for all 51,185 Orthophotos (hr_orthophoto.zip) and building masks (hr_mask.zip), which serve as ground truth for evaluation.
Low-Resolution Data STAC-compliant data for all 51,185 Sentinel-2 images stored in lr_s2.zip.
Super-Resolved Outputs Results from the used super-resolution algorithms: sr_deepsent.zip, sr_evoland.zip, sr_ldsr_s2.zip, sr_sen2sr_lite.zip, sr_sen2sr_rgbn.zip, sr_sr4rs.zip, sr_swin2_mose.zip. The data is STAC-compliant and includes all 22,665 images of the stratified dataset Dataset Filtered.
Inference Results Building delineation outputs are provided in building_delineation_inference.zip.
Metric Results Results for each SR model and interpolation method across several metrics are provided in metadata.zip/metric_results.
Sample Data The metadata.zip/image_samples directory contains a small, ready-to-use subset demonstrating the dataset’s structure.
Stratification Tables Training, validation, and test splits are organized in CSV and GeoPackage (gpkg) format.
Evaluation Additional data for evaluating proprietory models (Tracasa) is available in tracasa.zip.
Figures Thesis figures are included in metadata.zip/thesis_figures.
All .tif files are georeferenced. When using them, please be advised to RETAIN their geospatial attributes troughout your processing steps.
All geospatial files adhere to the STAC format. Images are zipped in folders to adhere to the required data structure of this hosting service. No band specific metadata is included, as spectral shifts from the SR might change each images spectral signature.
If you use this dataset, please refer to the corresponding git-repository for the current citation recommendations: https://github.com/Zerhigh/Evaluating_Sentinel-2_Super-Resolution_Algorithms_for_Automated_Building_Delineation