| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "examples": { |
| "@id": "cr:examples", |
| "@type": "@json" |
| }, |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isArray": "cr:isArray", |
| "arrayShape": "cr:arrayShape", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform", |
| "wd": "https://www.wikidata.org/wiki/", |
| "prov": "http://www.w3.org/ns/prov#" |
| }, |
| "@type": "sc:Dataset", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "name": "Bavaria EO Benchmark (12-patch subset with DOP20)", |
| "description": "A tiny stratified subset (12 patches) of the Bavaria EO Benchmark grid, exported for development and smoke tests. Each patch includes Sentinel-2 seasonal composites (10 bands, 4 seasons), Sentinel-1 GRD seasonal composites (VV+VH, 4 seasons), tree-species raster, tree-structure labels at 10 m, and co-registered DOP20 true-colour aerial orthophotos at 20 cm GSD (6400×6400×3 uint8 per patch). Splits: 9 train, 2 validation, 1 test. Data is distributed as WebDataset .tar shards plus slim Parquet index files; tensor layouts are defined in data/schema.json.", |
| "url": "https://github.com/your-org/Bavaria-EO-Benchmark", |
| "version": "0.2.0-subset-dop20", |
| "license": "https://creativecommons.org/licenses/by/4.0/", |
| "keywords": [ |
| "earth-observation", |
| "remote-sensing", |
| "sentinel-2", |
| "sentinel-1", |
| "aerial-imagery", |
| "DOP20", |
| "orthophoto", |
| "tree-inventory", |
| "Bavaria", |
| "geospatial", |
| "webdataset" |
| ], |
| "creator": [ |
| { |
| "@type": "Organization", |
| "name": "Bavaria EO Benchmark Authors" |
| } |
| ], |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "schema-json", |
| "name": "schema.json", |
| "description": "Machine-readable tensor schema: maps each .tar member suffix to dtype, shape, band names, and description.", |
| "contentUrl": "data/schema.json", |
| "encodingFormat": "application/json", |
| "sha256": "34c8f651e74511606de93ff844ade1595faaf590aea1d9bcf3bf52c38a50b3b9" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "index-train", |
| "name": "train.parquet", |
| "description": "Slim patch index for the train split (9 patches). Columns: patch metadata + sample_key + shard_relpath.", |
| "contentUrl": "data/index/train.parquet", |
| "encodingFormat": "application/x-parquet", |
| "sha256": "2f851ae9e22b694302a38087ec1da3effe30b80690d62667146059a3e0147df9" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "index-validation", |
| "name": "validation.parquet", |
| "description": "Slim patch index for the validation split (2 patches).", |
| "contentUrl": "data/index/validation.parquet", |
| "encodingFormat": "application/x-parquet", |
| "sha256": "fa5b0ae02c966e73a94a4b2eb67ef3a968f12d15ccbcdc484fc86d58906a2cbb" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "index-test", |
| "name": "test.parquet", |
| "description": "Slim patch index for the test split (1 patch).", |
| "contentUrl": "data/index/test.parquet", |
| "encodingFormat": "application/x-parquet", |
| "sha256": "786e089f76c3a4ba90973a46b5bc802c1457842b7444c50aa841e78a3b448145" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "train-shards", |
| "name": "train shards", |
| "description": "WebDataset .tar shard for the train split (1 shard, ~1.1 GB). Each sample key is a 6-digit zero-padded zarr_idx. Members documented in schema.json.", |
| "encodingFormat": "application/x-tar", |
| "includes": "data/shards/train-*.tar" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "validation-shards", |
| "name": "validation shards", |
| "description": "WebDataset .tar shard for the validation split (1 shard, ~242 MB).", |
| "encodingFormat": "application/x-tar", |
| "includes": "data/shards/validation-*.tar" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "test-shards", |
| "name": "test shards", |
| "description": "WebDataset .tar shard for the test split (1 shard, ~121 MB).", |
| "encodingFormat": "application/x-tar", |
| "includes": "data/shards/test-*.tar" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "patches", |
| "name": "patches", |
| "description": "One record per patch. Columns are patch metadata and locators (sample_key, shard_relpath) pointing to WebDataset samples in the tar shards. Tensor data is not stored in the Parquet rows; decode binary members per data/schema.json.", |
| "key": { |
| "@id": "patches/patch_id" |
| }, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "patches/patch_id", |
| "name": "patch_id", |
| "description": "Unique integer patch identifier from the full-Bavaria grid.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "patch_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/center_x", |
| "name": "center_x", |
| "description": "Patch centre easting in EPSG:25832 (metres).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "center_x" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/center_y", |
| "name": "center_y", |
| "description": "Patch centre northing in EPSG:25832 (metres).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "center_y" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/row_start", |
| "name": "row_start", |
| "description": "Top row index of the patch in the master grid (pixels).", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "row_start" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/row_end", |
| "name": "row_end", |
| "description": "Exclusive bottom row index of the patch in the master grid.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "row_end" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/col_start", |
| "name": "col_start", |
| "description": "Left column index of the patch in the master grid (pixels).", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "col_start" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/col_end", |
| "name": "col_end", |
| "description": "Exclusive right column index of the patch in the master grid.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "col_end" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/valid_pixel_pct", |
| "name": "valid_pixel_pct", |
| "description": "Fraction of non-NaN pixels in Sentinel-2 bands for this patch.", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "valid_pixel_pct" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/tree_pixel_pct", |
| "name": "tree_pixel_pct", |
| "description": "Fraction of pixels in this patch with at least one inventory tree (0-1).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "tree_pixel_pct" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/mean_tree_count", |
| "name": "mean_tree_count", |
| "description": "Mean number of inventory trees per non-zero pixel in this patch.", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "mean_tree_count" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/mean_tree_count_variance", |
| "name": "mean_tree_count_variance", |
| "description": "Mean tree-count variance statistic for this patch.", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "mean_tree_count_variance" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/split", |
| "name": "split", |
| "description": "Dataset split: train, val, or test.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "split" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/block_col", |
| "name": "block_col", |
| "description": "Block column index for spatial train/val/test assignment.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "block_col" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/block_row", |
| "name": "block_row", |
| "description": "Block row index for spatial train/val/test assignment.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "block_row" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/block_id", |
| "name": "block_id", |
| "description": "Geographic block ID used for spatial train/val/test assignment.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "block_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/distance_to_nearest_test_km", |
| "name": "distance_to_nearest_test_km", |
| "description": "Distance from patch centre to nearest test block (km).", |
| "dataType": "sc:Float", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "distance_to_nearest_test_km" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/buffered", |
| "name": "buffered", |
| "description": "Whether the patch lies in a buffer zone around test blocks.", |
| "dataType": "sc:Boolean", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "buffered" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/in_bavaria", |
| "name": "in_bavaria", |
| "description": "True if the patch centre falls within the official Bavaria boundary.", |
| "dataType": "sc:Boolean", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "in_bavaria" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/zarr_idx", |
| "name": "zarr_idx", |
| "description": "Row index in the source Zarr store (0-based); equals the integer value of sample_key.", |
| "dataType": "sc:Integer", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "zarr_idx" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/sample_key", |
| "name": "sample_key", |
| "description": "WebDataset sample key: 6-digit zero-padded decimal equal to zarr_idx.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "sample_key" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/shard_relpath", |
| "name": "shard_relpath", |
| "description": "Relative path from the dataset root to the .tar shard containing this sample.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "index-train" |
| }, |
| "extract": { |
| "column": "shard_relpath" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s2_spring", |
| "name": "s2_spring", |
| "description": "Sentinel-2 L2A spring composite. 10 bands. Stored as {sample_key}.s2_spring.f32 (raw float32 LE, shape 128×128×10).", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,10" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s2_summer", |
| "name": "s2_summer", |
| "description": "Sentinel-2 L2A summer composite. Stored as {sample_key}.s2_summer.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,10" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s2_autumn", |
| "name": "s2_autumn", |
| "description": "Sentinel-2 L2A autumn composite. Stored as {sample_key}.s2_autumn.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,10" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s2_winter", |
| "name": "s2_winter", |
| "description": "Sentinel-2 L2A winter composite. Stored as {sample_key}.s2_winter.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,10" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s1_spring", |
| "name": "s1_spring", |
| "description": "Sentinel-1 GRD spring composite (VV, VH). Stored as {sample_key}.s1_spring.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,2" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s1_summer", |
| "name": "s1_summer", |
| "description": "Sentinel-1 GRD summer composite. Stored as {sample_key}.s1_summer.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,2" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s1_autumn", |
| "name": "s1_autumn", |
| "description": "Sentinel-1 GRD autumn composite. Stored as {sample_key}.s1_autumn.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,2" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/s1_winter", |
| "name": "s1_winter", |
| "description": "Sentinel-1 GRD winter composite. Stored as {sample_key}.s1_winter.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128,2" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/tree_species", |
| "name": "tree_species", |
| "description": "Tree species classification. Stored as {sample_key}.tree_species.u8 (uint8, shape 128×128).", |
| "dataType": "sc:Integer", |
| "isArray": true, |
| "arrayShape": "128,128" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/mean_height_arr", |
| "name": "mean_height_arr", |
| "description": "Mean tree height per 10 m pixel (m). Stored as {sample_key}.mean_height.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/median_height_arr", |
| "name": "median_height_arr", |
| "description": "Median tree height per 10 m pixel (m). Stored as {sample_key}.median_height.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/height_variance_arr", |
| "name": "height_variance_arr", |
| "description": "Height variance per pixel. Stored as {sample_key}.height_variance.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/tree_count_arr", |
| "name": "tree_count_arr", |
| "description": "Inventory tree count per 10 m pixel. Stored as {sample_key}.tree_count.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/tree_density_arr", |
| "name": "tree_density_arr", |
| "description": "Tree density (3×3 neighbourhood). Stored as {sample_key}.tree_density.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/tree_count_variance_arr", |
| "name": "tree_count_variance_arr", |
| "description": "Tree-count variance per pixel. Stored as {sample_key}.tree_count_variance.f32.", |
| "dataType": "sc:Float", |
| "isArray": true, |
| "arrayShape": "128,128" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "patches/dop20_rgb", |
| "name": "dop20_rgb", |
| "description": "DOP20 true-colour aerial orthophoto at 20 cm GSD, co-registered with the 10 m stack. Stored as {sample_key}.dop20_rgb.u8 (raw uint8 LE, shape 6400×6400×3, channels R,G,B).", |
| "dataType": "sc:Integer", |
| "isArray": true, |
| "arrayShape": "6400,6400,3" |
| } |
| ] |
| } |
| ], |
| "rai:hasSyntheticData": false, |
| "prov:wasDerivedFrom": [ |
| { |
| "@id": "https://developers.google.com/earth-engine/datasets/catalog/sentinel-2", |
| "prov:label": "Sentinel-2" |
| }, |
| { |
| "@id": "https://docs.sentinel-hub.com/api/latest/data/sentinel-1-grd/", |
| "prov:label": "Sentinel-1" |
| }, |
| { |
| "@id": "https://geodaten.bayern.de/opengeodata/OpenDataDetail.html?pn=dop20rgb", |
| "prov:label": "DOP20" |
| }, |
| { |
| "@id": "https://geodaten.bayern.de/opengeodata/OpenDataDetail.html?pn=einzelbaeume", |
| "prov:label": "Einzelbaume" |
| } |
| ], |
| "prov:wasGeneratedBy": [ |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q5227332" |
| }, |
| "sc:description": "Individual s2 patches were made into mosiac to remove clouds, 20m bands from sentinel 2 were increased to 10m. For each 10mx10m pixel height mean was calculated and tree count was aggregated" |
| } |
| ], |
| "rai:dataLimitations": "The dataset has several known constraints and domain restrictions. Geographically, the dataset is restricted to the federal state of Bavaria, meaning models trained on this data may face distributional gaps and may not generalize to other global regions without domain adaptation. The targets are provided at a 10 m resolution, meaning individual trees cannot be directly observed and their properties must be inferred indirectly from medium-resolution radiometric and structural signals. Regarding data quality, our ground truth labels are extremely sparse and unevenly distributed. Furthermore, our use of seasonal composites removes short-term dynamics, which may obscure fine-grained temporal variability.", |
| "rai:dataBiases": "We acknowledge significant selection and label bias in our dataset due to extreme label sparsity. As noted in Section 3, only a small fraction of the grid cells contain tree observations (a median valid label coverage of about 26% per patch). This highly imbalanced supervision may bias learning, requiring models to operate under heavy imbalance and potentially skewing model behavior to perform better on managed landscapes compared to dense, wild forests.", |
| "rai:personalSensitiveInformation": "The dataset consists of satellite imagery and forest structure data. Regarding the categories provided, the dataset contains Geography information, as all data is spatially aligned to a 10 m reference grid (EPSG:25832) covering the federal state of Bavaria, and we utilize a geographic block-splitting strategy.\nThe dataset does not contain any personal or sensitive human information.", |
| "rai:dataUseCases": "The dataset is intended to measure fundamental, discrete ecological quantities of forest structure—specifically tree density (the number of trees per unit area) and average tree height—along with their associated variances for uncertainty quantification.\nWe have validated the dataset for use cases such as multi-task learning (joint prediction of continuous height and discrete count from shared inputs), cross-modal knowledge distillation (using 20cm high-resolution RGB with 10m Sentinel data), spatial generalization, and the benchmarking of Earth Observation (EO) models. We provide supporting evidence for this by establishing baseline evaluations using four distinct modeling approaches: XGBoost, U-Net, SegFormer-B3, and the Clay Foundation Model.", |
| "rai:dataSocialImpact": "Positive Effects: By facilitating the accurate, large-scale characterization of forest structures and integrating uncertainty quantification, our dataset can positively impact downstream decision-making in crucial climate science domains. Better modeling of tree density and height supports understanding of the global carbon cycle, ecosystem resilience, carbon accounting, biodiversity monitoring, and forest management risk assessment.\nNegative Effects & Fairness: A potential risk is the misuse of models trained on our benchmark for automated environmental policy-making or carbon accounting in regions unlike Bavaria or in unmanaged natural forests. Because our ground truth over-represents managed/urban trees and is strictly localized, applying these models elsewhere could yield unfair, biased, or highly inaccurate ecological assessments.\nMitigations: To mitigate overconfident or globally generalized deployments, we strictly define our dataset's intended use as methodological research rather than operational deployment. To mitigate geographic data leakage and ensure robust, honest evaluation of model generalization, we implemented a rigorous deterministic spatial block splitting strategy that strictly separates training, validation, and test sets, enforced by a 5 km buffer zone. We also incorporate a Masked Smooth L1 loss in our evaluation protocols to ensure models are not penalized for predicting structures in unannotated regions, thereby handling the label sparsity fairly during training." |
| } |
