Analysis-Ready Multimodal Satellite Feature Datasets for Building Damage Assessment in 20 Ukrainian Conflict Cities (Sentinel-1/Sentinel-2, 2020–2026)

Analysis-ready machine-learning datasets supporting the MSc thesis "Building Damage Assessment with Multimodal Satellite Time Series and Machine Learning" (UNIGIS, Paris Lodron University Salzburg, 2026). The data cover 20 Ukrainian cities affected by the Russia–Ukraine war, with 588,686 Overture building footprints, of which 7,326 carry UNOSAT-confirmed damage labels (binary: severe/destroyed vs. other). The deposit contains the feature-engineered Parquet datasets produced by the thesis pipeline from open Sentinel-1 SAR (backscatter and InSAR coherence) and Sentinel-2 multispectral imagery, aggregated per building footprint (V2) and per UNOSAT point (V3), plus variant datasets V1 and V4–V7 used for sampling-unit and feature-scope ablations. Also included: 64×64 image patches for deep-learning experiments (bda_patches), per-dataset profiling reports (dataset_profiles), and the frozen GroupKFold city-fold assignments, dataset manifest, and product catalog (stack_meta) required to reproduce the published cross-validation exactly. Contents: dataset_V2.zip (per-building datasets, primary), dataset_v1.zip, dataset_v3.zip–dataset_v7.zip (variants), bda_patches.zip, dataset_profiles.zip, stack_meta.zip, CHECKSUMS.txt (SHA256). Usage: unpack the ZIPs into the dataset/ folder of the data stack directory expected by notebooks/global_setup.py in the accompanying code repository (https://github.com/marcoheinzen/bda). With this deposit alone, the full modelling and evaluation chain (notebooks NB06–NB13) is reproducible without any satellite-data downloads or API credentials; see thesis Section 3.4.3.1 (Chapter 3, technical report) for the step-by-step replication procedure. Raw satellite products (~800 GB) are not deposited but can be regenerated from open archives with notebooks NB01–NB03. Attribution: Contains modified Copernicus Sentinel data (2020–2026), processed by the author. Damage reference labels derived from UNOSAT (United Nations Satellite Centre, UNITAR) damage assessments. Building footprints from the Overture Maps Foundation (CDLA-Permissive-2.0 / ODbL sources). Elevation data via the OpenTopography Facility. Comparator method and compiled UNOSAT point data after Dietrich et al. 2025 (doi:10.1038/s43247-025-02183-7).