Moriae/ukraine-forest-change-detection

Satellite Forest Change Detection Dataset

This dataset contains paired multispectral satellite image tiles and binary change masks for forest change detection. Each valid sample is a triplet:

• A: image at the first time point
• B: image at the second time point
• label: change mask for the same tile and period

The data is organized into train, val, and test splits. Each split has the same directory structure:

train/
  A/
  B/
  label/
val/
  A/
  B/
  label/
test/
  A/
  B/
  label/

Data Sources

The image tiles were derived from Sentinel-2 imagery. Forest change labels were derived from Hansen Global Forest Change data and converted into binary change masks for the selected temporal windows.

Each sample is provided as a paired before/after observation:

• A: start image for the temporal pair
• B: end image for the temporal pair
• label: binary forest change mask

Temporal Pairs

The dataset contains two temporal windows:

• 2017-2020
• 2020-2024

For each temporal pair, the A image represents the start period and the B image represents the end period. The label mask marks forest cover loss detected within the corresponding temporal window.

Files

All raster files are GeoTIFFs.

• A images: 10 bands, float32, 256 x 256 pixels, EPSG:3857
• B images: 10 bands, float32, 256 x 256 pixels, EPSG:3857
• label masks: 1 band, float32, 256 x 256 pixels, EPSG:3857

Bands

The Sentinel-2 image tiles contain 10 channels. The first six channels are spectral bands, and the remaining four are vegetation/burn/moisture indices computed from Sentinel-2 bands.

Band index	Name	Description	Native Sentinel-2 resolution
1	B2	Blue	10 m
2	B3	Green	10 m
3	B4	Red	10 m
4	B8	Near infrared	10 m
5	B11	Shortwave infrared 1	20 m
6	B12	Shortwave infrared 2	20 m
7	NDVI	Normalized Difference Vegetation Index	derived
8	NBR	Normalized Burn Ratio	derived
9	NDMI	Normalized Difference Moisture Index	derived
10	NBR2	Normalized Burn Ratio 2	derived

The released GeoTIFF tiles are stored as aligned 256 x 256 pixel arrays on a common grid. Indices are computed from the Sentinel-2 bands and provided as additional channels for convenience.

Users can work with all 10 channels or select a subset for a specific model. For example, the tiles can be converted to RGB using bands B4, B3, and B2, or used as multispectral inputs with vegetation indices included.

Filenames follow this pattern:

<tile_id>_A.tif
<tile_id>_B.tif
<tile_id>_label.tif

Example:

train/A/UA_FOREST_CHANGE_TRAIN_00000_2017-2020_A.tif
train/B/UA_FOREST_CHANGE_TRAIN_00000_2017-2020_B.tif
train/label/UA_FOREST_CHANGE_TRAIN_00000_2017-2020_label.tif

Statistics

The dataset contains 58,215 GeoTIFF files, about 93.33 GB in decimal units or 86.92 GiB.

Split	A files	B files	Label files	Valid triplets
train	13,783	13,783	13,783	13,783
val	2,454	2,454	2,454	2,454
test	3,168	3,168	3,168	3,168

Every remaining file belongs to a complete A/B/label triplet within the same split.

Metadata

The file-level metadata is available in metadata.csv. If present, it should be treated as the authoritative metadata for the released GeoTIFF files. It contains one row per final triplet, with paths to the corresponding A, B, and label files.

Final Metadata Summary

The final metadata contains 19,405 valid triplets. The class statistics below were recomputed from metadata.csv using the change_frac column, where 1.0 means 100% of pixels changed.

Samples Per Split

Split	Samples
train	13,783
val	2,454
test	3,168

Change Class Thresholds

Class label	change_frac rule	Percent range
no_change	< 0.01	< 1%
hard_small	>= 0.01 and < 0.04	1% to < 4%
medium	>= 0.04 and < 0.10	4% to < 10%
easy_large	>= 0.10	10% to 100%

Samples Per Class

Class label	Samples
easy_large	1,038
hard_small	3,446
medium	3,135
no_change	11,786

Mean Change Fraction Per Split

Split	Mean change fraction
train	0.026317
val	0.047822
test	0.023582

Note: The following dataset statistics are based on the pre-cleaning metadata. During final dataset preparation, edge tiles, incomplete tiles, and samples with invalid or missing raster data were removed. Therefore, these values describe the original sampled dataset distribution and should be treated as approximate for the final released version.

Pre-cleaning Split Statistics

Temporal pair	Split	Number of records
2017-2020	Train	7,393
2017-2020	Validation	1,305
2020-2024	Train	7,393
2020-2024	Validation	1,305
2017-2020	Test	1,674
2020-2024	Test	1,736
Total	Train/Validation/Test	20,806

Pre-cleaning Area Coverage

Dataset part	Metadata records	Unique spatial tiles	Approx. coverage area, km²	Approx. forest area, km²
Train/Validation	17,396	8,698	23,526.00	14,105.19
Test	3,410	2,984	8,272.21	4,578.22
Total	20,806	11,682	31,798.21	18,683.40

Pre-cleaning Change Category Distribution

Category	Technical label	Description	Train/Val, %	Test, %
No change	no_change	Tiles without confirmed forest cover loss	50.00	50.00
Medium change	medium	Tiles with a medium fraction of changed pixels	28.74	29.33
Hard small change	hard_small	Tiles with a small fraction of changed pixels that are difficult to detect	11.50	14.66
Easy large change	easy_large	Tiles with large or clearly expressed forest cover loss areas	9.76	6.01

Pre-cleaning Regional Forest Coverage Statistics

Region	Number of tiles	Mean forest fraction	Standard deviation	Median	Minimum	Maximum
Carpathians	3,432	0.7500	0.2548	0.8579	0.2005	1.0000
Polissia	4,512	0.6852	0.2607	0.7475	0.2003	1.0000
Sumy	3,316	0.5417	0.2565	0.4872	0.2000	1.0000
Kharkiv	3,318	0.5272	0.2511	0.4733	0.2000	1.0000
Donbas	2,818	0.4352	0.2165	0.3596	0.2001	1.0000
Central-Western	3,410	0.5702	0.2722	0.5246	0.2000	1.0000

These pre-cleaning statistics describe the original sampling design, regional diversity, class balance, and approximate spatial coverage. The final metadata.csv, if present, should be considered the authoritative file-level metadata for the released GeoTIFF files.

Pre-cleaning Statistics Figures

All figures in this section were generated from the pre-cleaning metadata. They show the original sampled dataset distribution before edge tiles, incomplete tiles, and invalid raster samples were removed. They should be treated as approximate context for the released dataset, not as exact statistics for the final GeoTIFF files.

Change Fraction Distribution

These pre-cleaning plots show the distribution of changed-pixel fractions in the originally sampled tiles.

Forest Coverage Distribution

These pre-cleaning plots show the distribution of forest coverage fractions used during sampling.

Regional Distribution

These pre-cleaning plots summarize the geographic distribution of sampled tiles by region.

Samples By Region

This pre-cleaning plot shows the train/validation sample count by region.

Labels

The training code treats the masks as binary change labels:

binary_mask = label > 0

Values equal to 0 are treated as no change. Values greater than 0 are treated as change.

Label Limitations

The labels are derived from Hansen forest change data and inherit the limitations of that source and the preprocessing pipeline. They should be interpreted as binary forest cover loss masks rather than perfect ground-truth field annotations. Small changes, mixed pixels, cloud/shadow artifacts, seasonal effects, edge tiles, and temporal uncertainty can affect label quality.

These labels are suitable for training and evaluating forest change detection models, but users should consider additional validation for high-stakes or site-specific applications.

Loading Example

from pathlib import Path

import rasterio

root = Path("dataset_builder/dataset_cleaned")
split = "train"
tile_id = "UA_FOREST_CHANGE_TRAIN_00000_2017-2020"

with rasterio.open(root / split / "A" / f"{tile_id}_A.tif") as src:
    image_a = src.read()

with rasterio.open(root / split / "B" / f"{tile_id}_B.tif") as src:
    image_b = src.read()

with rasterio.open(root / split / "label" / f"{tile_id}_label.tif") as src:
    label = src.read(1)

binary_mask = label > 0

print(image_a.shape, image_b.shape, binary_mask.shape)

Intended Use

This dataset is intended for research and development of satellite image change detection models, especially binary forest change segmentation from paired multispectral image tiles.