Dataset Viewer
Auto-converted to Parquet Duplicate
patch_id
int64
24.2k
64.3k
center_x
float64
606k
780k
center_y
float64
5.31M
5.49M
row_start
int64
11k
29.3k
row_end
int64
11.1k
29.4k
col_start
int64
10.8k
28.2k
col_end
int64
10.9k
28.3k
valid_pixel_pct
float64
1
1
tree_pixel_pct
float64
0.02
0.6
mean_tree_count
float64
0.03
0.93
mean_tree_count_variance
float64
0.03
0.42
split
stringclasses
1 value
block_col
int64
1
4
block_row
int64
1
4
block_id
int64
9
27
distance_to_nearest_test_km
float64
38.1
138
buffered
bool
1 class
in_bavaria
bool
1 class
zarr_idx
int64
1
9
sample_key
stringclasses
9 values
shard_relpath
stringclasses
1 value
64,262
680,230
5,308,770
29,312
29,440
18,176
18,304
1
0.3922
0.5365
0.3174
train
3
1
9
38.106
false
true
1
000001
data/shards/train-000000.tar
62,631
742,950
5,316,450
28,544
28,672
24,448
24,576
1
0.3873
0.5726
0.2973
train
4
1
10
96.615
false
true
2
000002
data/shards/train-000000.tar
50,822
680,230
5,370,210
23,168
23,296
18,176
18,304
1
0.2089
0.3094
0.1684
train
3
2
15
54.279
false
true
3
000003
data/shards/train-000000.tar
52,580
780,070
5,362,530
23,936
24,064
28,160
28,288
1
0.1335
0.1667
0.126
train
4
2
16
137.695
false
true
4
000004
data/shards/train-000000.tar
32,022
629,030
5,455,970
14,592
14,720
13,056
13,184
1
0.4612
0.7325
0.3254
train
2
3
20
79.034
false
true
5
000005
data/shards/train-000000.tar
35,475
748,070
5,440,610
16,128
16,256
24,960
25,088
1
0.3473
0.4949
0.2877
train
4
3
22
132.506
false
true
6
000006
data/shards/train-000000.tar
25,564
605,990
5,485,410
11,648
11,776
10,752
10,880
1
0.022
0.0269
0.0259
train
1
4
25
98.406
false
true
7
000007
data/shards/train-000000.tar
24,175
620,070
5,491,810
11,008
11,136
12,160
12,288
1
0.5992
0.9334
0.3933
train
2
4
26
108.086
false
true
8
000008
data/shards/train-000000.tar
24,222
680,230
5,491,810
11,008
11,136
18,176
18,304
1
0.569
0.806
0.4161
train
3
4
27
117.75
false
true
9
000009
data/shards/train-000000.tar

Bavaria EO Benchmark (12-patch subset with DOP20)

A development subset of the Bavaria EO Benchmark: 12 patches (9 train, 2 validation, 1 test) sampled from the full grid, including DOP20 true-colour aerial orthophotos at 20 cm ground sampling distance alongside the usual 10 m Sentinel-1/2 stacks and tree-structure labels.

This release is not intended as a training corpus at scale.

Distribution format

data/
  schema.json              <- tensor member definitions (dtype, shape, band names)
  index/
    train.parquet          <- 9 patches   (metadata + locators only)
    validation.parquet     <- 2 patches
    test.parquet           <- 1 patch
  shards/
    train-000000.tar       <- ~1.1 GB (all train samples + DOP20)
    validation-000000.tar  <- ~242 MB
    test-000000.tar        <- ~121 MB
metadata.json              <- Croissant 1.0 JSON-LD

Per-sample members inside each .tar

Each patch uses a sample_key (6-digit zero-padded zarr_idx). Members share that prefix:

Member Shape dtype Description
{key}.s2_*.f32 128×128×10 float32 Sentinel-2 seasonal composites (4 seasons)
{key}.s1_*.f32 128×128×2 float32 Sentinel-1 seasonal composites (VV, VH)
{key}.tree_species.u8 128×128 uint8 Tree species class IDs
{key}.*_height*.f32, {key}.tree_*.f32 128×128 float32 Label rasters (see schema.json)
{key}.dop20_rgb.u8 6400×6400×3 uint8 DOP20 RGB orthophoto (20 cm)
{key}.json JSON Patch id, centre coordinates, split

Decode raw buffers with numpy.frombuffer(..., dtype=...).reshape(...) as in data/schema.json.

Croissant

metadata.json conforms to Croissant 1.0 and references data/schema.json, the index Parquets, and the shard FileSet globs.

Downloads last month
62