rasmuspjohansson/KDS_befaestelse

KDS befæstelse

Training data for the surface-type (befæstelse) segmentation models of the Danish Agency for Climate Data (Klimadatastyrelsen, KDS). The models classify each pixel as asphalt, tiles, gravel, unpaved, green roof and so on.

The models were trained with ML_sdfi_fastai2.

Contents

Path	What
data/splitted/rgb/part_NN/	RGB image tiles
data/splitted/cir/part_NN/	CIR (color infrared) image tiles
data/splitted/OrtoRGB/part_NN/	RGB orthophoto tiles
data/splitted/OrtoCIR/part_NN/	CIR orthophoto tiles
data/splitted/DSM/part_NN/	Digital surface model tiles
data/splitted/DTM/part_NN/	Digital terrain model tiles
labels/splitted_labels/part_NN/	One label raster per tile, with the same file name
data/all.txt	All 22,997 tiles, used for training and validation
data/train.txt	Training tiles
data/valid_all_minus_train.txt	The 25 validation tiles: all.txt minus train.txt
data/READMKE.txt	Short descriptions of some of the hand-picked validation tiles
codes.txt	Class names, one per label value

Every data type has the same 22,997 tiles, about 280 GB in total. All rasters are 1000 x 1000 pixel GeoTIFFs at 10 cm in EPSG:25832.

Data type	Bands	Type
rgb, cir, OrtoRGB, OrtoCIR	3	uint8
DSM, DTM	1	float32
labels	1	uint8

The tile name prefix is the image year: O2021_* (14,174 tiles), O2023_* (5,150) and O2025_* (3,673).

Some DSM and DTM tiles are all zeros (no elevation data). In a random sample this was about 25% of the O2025_* tiles and 5-7% of the O2021_* tiles. The models were trained on the data as it is here.

The Hub allows at most 10,000 files per folder, so the tiles are split into part_00/ to part_02/ subfolders: the names in all.txt, sorted, in chunks of 10,000. A tile has the same part_NN in every data type and in the labels. Training expects one flat folder per data type; see Download.

Classes

Value	Class
0	unknown (ignore_index, excluded from the loss)
1	asfalt (asphalt)
2	fliser (paving tiles)
3	grus (gravel)
4	ubefestet (unpaved)
5	green_roof
6	drivhus (greenhouse)
7	betonflade (concrete)
8	brosten (cobblestones)
9	unknown2
10	solceller (solar panels)

How the labels were made

The labels were rasterized from the polygons in 20261023_labels.gpkg, which is not part of this dataset.

Training setup

The models read the dataset like this (from the ML_sdfi_fastai2 .ini configs):

[CONFIG]
datatypes = ["rgb","cir","OrtoRGB","OrtoCIR","DSM","DTM"]
channels  = [[0,1,2],[0],[0,1,2],[0],[0],[0]]

[DATASET]
path_to_images    = <dataset>/data/splitted/rgb
path_to_labels    = <dataset>/labels/splitted_labels
path_to_dataset   = <dataset>/
path_to_all_txt   = <dataset>/data/all.txt
path_to_valid_txt = <dataset>/data/valid_all_minus_train.txt
path_to_codes     = <dataset>/codes.txt

The 10 input channels are the three RGB bands, the first CIR band, the three OrtoRGB bands, the first OrtoCIR band, DSM and DTM. Training uses random 992 x 992 crops of the tiles.

Download

Use download_dataset.sh from ML_datasets_on_hugging_face. --flatten-parts moves the tiles out of the part_NN/ folders, which gives the layout the training configs expect:

./download_dataset.sh \
    --repo-id rasmuspjohansson/KDS_befaestelse \
    --output-dir ./KDS_befaestelse \
    --flatten-parts

To skip the elevation models, add --exclude "data/splitted/DSM/*" --exclude "data/splitted/DTM/*".

With only the hf CLI:

hf download rasmuspjohansson/KDS_befaestelse \
    --repo-type dataset \
    --local-dir ./KDS_befaestelse
cd KDS_befaestelse
for d in data/splitted/* labels/splitted_labels; do
    find "$d" -mindepth 2 -type f -exec mv -n -t "$d" {} + && rmdir "$d"/part_*
done

Contact

rajoh@kds.dk