danvinci/mosaic_sen2_cc_change_detection

Multi-image change-detection reformat of the MOSAIC-SEN2-CC dataset (Tuzlupinar et al., IEEE JSTARS 2025), prepared for vision-language fine-tuning of LiquidAI/LFM2.5-VL-1.6B.

Splits

split	rows	source
train	29281	full MOSAIC TRAIN partition
validation	700	balanced 33% positive sample

Schema

Each row is a (before_img, after_img, prompt, answer, schema_id) tuple:

field	type	description
before_img	PIL Image	Sentinel-2 RGB tile before the change window
after_img	PIL Image	Sentinel-2 RGB tile after the change window
prompt	string	Per-schema predicate prompt
answer	string	JSON-shaped object or literal "null"
schema_id	string	One of 7 change-event schemas (see below)

Change-event schemas (7)

Each schema lives at schemas/<schema_id>.json and defines the positive-event field spec. Predicate prompts in the dataset reference these contracts.

• ConstructionProgress — building / road footprint expansion
• FloodExtentChange — water surface expansion (excludes stable water)
• IceExtentChange — sea / lake / glacier ice retreat or growth
• ThermalHotspot — wildfire / volcanic / industrial heat anomaly
• UrbanChangeDelta — generic urban land-cover delta
• VegetationStress — drought / disease / clear-cut signal
• WetlandChange — wetland surface delta

Usage

from datasets import load_dataset
ds = load_dataset("danvinci/mosaic_sen2_cc_change_detection")
print(ds)
# DatasetDict({
#     train: Dataset(...29281 rows),
#     validation: Dataset(...700 rows),
# })

row = ds["train"][0]
row["before_img"]  # PIL.Image (RGB Sentinel-2 tile)
row["after_img"]   # PIL.Image
row["prompt"]      # "Did urban-change-delta occur..."
row["answer"]      # JSON object or "null"
row["schema_id"]   # "UrbanChangeDelta"

Auxiliary files at repo root

• schemas/<schema_id>.json — 7 field-spec contracts
• mosaic_reformat_v0.jsonl — original row-metadata (before/after tile paths, split assignment, expected_output) used to build the prepared Arrow

Provenance + license notes

This is a derivative reformat of the MOSAIC-SEN2-CC dataset, not the canonical release. The reformat collapses MOSAIC's change-detection annotations into a uniform image-text-to-text supervised contract suitable for VLM fine-tuning, but the underlying tile pairs and labels originate from the upstream MOSAIC project.

License posture (the upstream repo does not ship an explicit LICENSE file — typical academic-dataset gray zone):

• Sentinel-2 imagery is redistributable: ESA Copernicus Sentinel data is released under the Copernicus "free, full and open access" license.
• MOSAIC change-detection annotations: research/academic use under citation is the established norm; non-research / commercial redistribution may require explicit permission from the original authors. Contact the upstream maintainers via the GitHub repo for any non-research use.
• The reformat itself (this repo's prompt structure, schema definitions, Arrow layout) is shared under the same research-use-cite-source posture as the upstream MOSAIC.

Built during the DPhi Space x Liquid AI hackathon (May 2026) as part of the liquid_space project.

Citation

Always cite the upstream MOSAIC-SEN2-CC paper when using this reformat:

@ARTICLE{karaca2025robust,
  author={Busra Tuzlupinar and Enes Ozelbas and Mehmet Fatih Amasyali and Ali Can Karaca},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  title={Introducing MOSAIC-SEN2-CC: A Multispectral Dataset and Adaptation Framework for Remote Sensing Change Captioning},
  year={2025},
  volume={18},
  pages={25410-25426},
  doi={10.1109/JSTARS.2025.3615113}
}