| --- |
| license: cc-by-4.0 |
| pretty_name: GeoPathfinder |
| task_categories: |
| - image-classification |
| tags: |
| - remote-sensing |
| - earth-observation |
| - landsat |
| - long-range-reasoning |
| - benchmark |
| - geospatial |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # GeoPathfinder |
|
|
| A Pathfinder-style long-range spatial reasoning benchmark on real satellite |
| imagery. Each sample is a 256×256 Landsat patch (native 30 m resolution, a |
| 7.7×7.7 km footprint) with two red dots on land; the task is to predict whether |
| the dots are connected by land (`connect = 1`) or separated by water |
| (`disconnect = 0`). Labels derive from Overture Maps water polygons rasterized |
| onto the exact pixel grid of each patch and cross-checked against Landsat's own |
| QA water flag. |
|
|
| - **5,273 patches → 10,546 images** across **42 river basins on six |
| continents**, exactly class-balanced (each patch yields one connect and one |
| disconnect sample, distance-matched within 25% so dot separation carries no |
| label signal) |
| - **Adversarial dot placement**: each released pair is the most deceptive of a |
| large candidate pool (96 for train/val, 256 for test), deterministic given |
| the masks — randomly placed dots admit a sight-line shortcut |
| - Fixed 80/10/10 patch-level splits (paired images never straddle a split). |
| A basin-disjoint protocol |
| (`splits_basin.json`, repo root) holds out four basins on four continents |
| for geographic-generalization tests |
| - `eval_variants/` ships a fixed easy/medium/hard placement suite (1,051 |
| images per tier): report per-tier accuracy and the macro average |
| (placement-general score) alongside test accuracy |
| - Code, construction pipeline, baselines: |
| https://github.com/isaaccorley/geo-long-range-arena |
|
|
| ## Download |
|
|
| The full dataset ships as a single archive that extracts to a `geopathfinder/` |
| folder: |
|
|
| ```bash |
| hf download isaaccorley/GeoPathfinder geopathfinder.tar.gz --repo-type dataset --local-dir . |
| tar -xzf geopathfinder.tar.gz |
| ``` |
|
|
| `metadata.parquet` and `water_polygons.parquet` are also available uncompressed |
| at the repo root for querying without downloading the archive. |
|
|
| ## Files (inside geopathfinder.tar.gz) |
|
|
| ``` |
| pathfinder/connect/ patch_XXXXX.jpg benchmark image, label = 1 |
| pathfinder/disconnect/ patch_XXXXX.jpg benchmark image, label = 0 |
| images/ patch_XXXXX.jpg true color without dots (pristine) |
| raw/ patch_XXXXX.tif uint16 RGBN surface-reflectance DNs, |
| scale/offset in band metadata, scene |
| provenance in file tags (ZSTD) |
| masks/ patch_XXXXX.png 0=land 1=water 2=invalid 3=suspect |
| overlays/ patch_XXXXX.jpg water mask blended over true color (QC) |
| metadata.parquet one row per patch: split, region, scene |
| id, bounds, CRS/transform, QC stats, |
| dot pixel + lon/lat coordinates |
| water_polygons.parquet 281k Overture water polygons clipped to |
| patch bounds (GeoParquet, WGS84), keyed |
| by sample_id; geometry_patch_crs holds |
| the same shapes as WKB in each patch's |
| UTM CRS (patch_crs column) |
| splits.json train/val/test patch ids |
| plan.json build configuration snapshot |
| ``` |
|
|
| Rasterizing a patch's rows from `water_polygons.parquet` onto its |
| `transform`/`crs` (in `metadata.parquet`) reproduces its stored mask |
| (IoU ≥ 0.999). |
|
|
| ## Quick start |
|
|
| ```python |
| import pandas as pd |
| |
| meta = pd.read_parquet("metadata.parquet") |
| train = meta[meta.split == "train"] |
| # each labeled patch contributes two images: |
| # pathfinder/connect/{sample_id}.jpg label 1 |
| # pathfinder/disconnect/{sample_id}.jpg label 0 |
| ``` |
|
|
| Raw reflectance (for geospatial foundation models — dots are not burned into |
| these; use the dot coordinates from `metadata.parquet`): |
|
|
| ```python |
| import rasterio |
| |
| with rasterio.open("raw/patch_00122.tif") as src: |
| reflectance = src.read().astype("float32") * src.scales[0] + src.offsets[0] |
| ``` |
|
|
| ## Baselines |
|
|
| Two numbers per model: accuracy on the released test split, and the |
| **placement-general score** (macro accuracy over the easy/medium/hard suite, |
| seed-42 checkpoints). Accuracies are mean ± std over three seeds. |
|
|
| Reference pipelines and probes: |
|
|
| | approach | test acc | macro | |
| |---|---|---| |
| | Water index (blue-red on RGB), no learning | 52.8% | 54.1 | |
| | Water index (NDWI on green/NIR), no learning | 69.9% | 76.2 | |
| | U-Net masks -> connected components | 69.2 ± 0.6% | 78.2 ± 1.1 | |
| | Best frozen linear probe (13 pretrainings) | 62.7% | — | |
|
|
| Architectures (fine-tuned on the released placements, three-seed means): |
|
|
| | model | test acc | macro | |
| |---|---|---| |
| | MambaVision-B (ext. recipe) | **88.0 ± 0.6%** | **62.7 ± 0.7** | |
| | OverLoCK-B | **88.0 ± 0.8%** | 62.1 ± 1.2 | |
| | FocalNet-B LRF | 87.8 ± 0.4% | 62.2 ± 1.1 | |
| | MaxViT-B | 87.5 ± 0.9% | 61.0 ± 4.4 | |
| | ConvNeXt-B | 87.2 ± 0.9% | 61.0 ± 0.5 | |
| | FocalNet-B SRF | 86.8 ± 0.6% | 61.0 ± 0.9 | |
| | UniRepLKNet-T | 86.6 ± 1.0% | 60.7 ± 0.9 | |
| | RepLKNet-31B (ext. recipe) | 86.4 ± 0.8% | 60.7 ± 0.6 | |
| | MambaOut-B | 86.4 ± 0.3% | 62.0 ± 1.5 | |
| | DaViT-B (2/3 seeds converge) | 86.4 ± 0.5% | 59.9 ± 1.3 | |
| | ResNet-50 | 85.3 ± 0.6% | 59.0 ± 1.0 | |
| | Swin-B | 83.7 ± 1.3% | 58.5 ± 2.2 | |
| | ViT-B/16 (ext. recipe) | 82.0 ± 1.0% | 56.4 ± 1.2 | |
| | ResNet-18 (scratch) | 78.0 ± 0.4% | 53.2 ± 1.1 | |
|
|
| All fourteen architectures land between 53.2 and 62.7 macro (a further 22 |
| size/pretraining/weight variants stay inside 58-65). Every model trained on |
| the released placements collapses on the macro score (models learn "the |
| obvious answer is wrong" and fail easy placements); retraining with dots |
| resampled every epoch recovers up to 85.1 ± 0.3 macro (MaxViT) while the |
| hard tier stays around 75% — the benchmark is far from solved. Marker shape |
| and color barely matter (±2 points). Full protocol and training recipes: |
| https://github.com/isaaccorley/geo-long-range-arena |
|
|
| ## Provenance and licensing |
|
|
| - Imagery: Landsat Collection 2 Level-2 (USGS, public domain), accessed via the |
| Microsoft Planetary Computer. Scene ids, WRS path/row, and acquisition times |
| are embedded in each GeoTIFF and in `metadata.parquet`. |
| - Water geometry: Overture Maps `base/water` (release 2026-06-17.0), which |
| includes OSM-derived data © OpenStreetMap contributors. The |
| `water_polygons.parquet` and `masks/` layers are therefore available under |
| ODbL; annotations and imagery composites are CC-BY-4.0. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{corley2026geopathfinder, |
| title = {GeoPathfinder: Long-Range Spatial Reasoning in Satellite Imagery}, |
| author = {Corley, Isaac}, |
| year = {2026}, |
| url = {https://huggingface.co/datasets/isaaccorley/GeoPathfinder} |
| } |
| ``` |
|
|
