Shadow prediction in geospatial data using a deep U-Net model — Maribor DSM and hourly shadow masks

This record contains the data for the paper:

Tjaša Repič, Domen Kavran, Domen Mongus, Niko Lukač. Napovedovanje senčenja v geoprostorskih podatkih z globokim modelom U-Net (Shadow Prediction in Geospatial Data Using a Deep U-Net Model). Proceedings of ERK 2025, Portorož, Slovenia, pp. 382–386.

Code repository:  https://github.com/SAMPA-Project/unet-shadow-prediction-repic

The paper trains a U-Net to predict binary shadow masks from a digital surface model (DSM) and the position of the Sun. The input is a 3-channel raster: elevation, slope and aspect. An auxiliary branch embeds the solar azimuth and elevation. This archive contains the study area used in the paper: a 750 m × 1000 m part of Maribor, Slovenia, at 1 m resolution. It also contains one simulated shadow mask for every hour of the year (8,760 masks) and the matching sun positions.

Download and extract the archive into the repository root. The code expects the files under data/.

Layout

• data/dsm.tif — DSM, surface elevation in metres (251–317 m). 751 × 1001 px, 1 m pixels, float32.
• data/slope.tif — slope β in radians (0 – 1.30), computed from the DSM gradient with central differences. Same grid as the DSM.
• data/aspect.tif — aspect γ = atan2(−∂G/∂x, ∂G/∂y) in radians (−π – π). Same grid as the DSM.
• data/sun_position.txt — 8,760 lines of azimuth elevation, in degrees. Line day × 24 + hour (0-indexed) belongs to mask <day>_<hour>.
• data/shadow_masks/result_shadowing_<day>_<hour>.tiff — 8,760 binary shadow masks: <day> = 0–364, <hour> = 0–23, 1001 × 751 px, uint8.
  • 255 = shadowed, 0 = not shadowed.
  • The masks are on the same grid as the DSM.
  • Hours with the Sun below the horizon are all zeros.

How the data was produced

• DSM — derived from airborne LiDAR data of the Slovenian Environment Agency (ARSO).
• Slope and aspect — computed from the DSM (Eq. 1–4 of the paper).
• Sun positions — computed with the NREL Solar Position Algorithm (SPA; Reda & Andreas, 2004/2007) for each hour of the year.
• Shadow masks — simulated with the GPU-accelerated ray-casting solar model of Lukač et al., Novel GPU-accelerated high-resolution solar potential estimation in urban areas by using a modified diffuse irradiance model, Applied Energy 353 (2024) 122129.

Notes

• The paper splits (tile, hour) samples randomly 60/20/20, using 256 × 256 tiles with stride 128. That is about 210,000 samples. The split is done in code and is not stored in this archive.

Data sources and licences

• LiDAR data — Slovenian Environment Agency (ARSO), used under ARSO's open-data terms (CC BY 4.0).

Acknowledgements

This research was co-funded by the Slovenian Research and Innovation Agency (ARIS) under basic research project J7-50095 and research programme P2-0041.