First Update of metadata YAML
Browse files
README.md
CHANGED
@@ -1,42 +1,105 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
## Geolayers-Data
|
2 |
-
This dataset card contains usage instructions and metadata for all data-products released with our paper: Using Multiple Input Modalities can Improve Data-Efficiency and O.O.D. Generalization for ML with Satellite Imagery. We release 3 modified versions of 3 benchmark datasets spanning land-cover segmentation, tree-cover regression, and multi-label land-cover classification tasks. These datasets are augmented with auxiliary, geographic inputs. A full list of contributed data products is shown in the table below.
|
3 |
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
|
8 |
-
|
9 |
-
| [
|
|
|
|
|
|
|
10 |
|
11 |
## 📦 Datasets & Georeferenced Auxiliary Layers
|
12 |
|
13 |
-
### SustainBench – Farmland Boundary Delineation
|
14 |
-
* **Optical input:** Sentinel-2 RGB patches (224
|
15 |
* **Auxiliary layers (all geo-aligned):**
|
16 |
* 19-channel OpenStreetMap (OSM) raster stack (roads, waterways, buildings, biome classes, …)
|
17 |
* EU-DEM (20 m GSD, down-sampled to 10 m)
|
18 |
-
* **Why:** OSM + DEM give an 8 % Dice boost when labels are scarce; gains appear once the training set drops below ≈ 700 images.
|
19 |
|
20 |
---
|
21 |
|
22 |
-
### EnviroAtlas – Land-Cover Segmentation
|
23 |
* **Optical input:** NAIP 4-band RGB-NIR aerial imagery at 1 m resolution.
|
24 |
* **Auxiliary layers:**
|
25 |
* OSM rasters (roads, waterbodies, waterways)
|
26 |
* **Prior** raster – a hand-crafted fusion of NLCD land-cover and OSM layers (PROC-STACK)
|
27 |
-
* **Splits:** Train = Pittsburgh; OOD validation/test = Austin & Durham. Auxiliary layers raise OOD overall accuracy by ~4 pp without extra fine-tuning.
|
28 |
|
29 |
---
|
30 |
|
31 |
-
### BigEarthNet v2.0 – Multi-Label Land-Cover Classification
|
32 |
* **Optical input:** 10-band Sentinel-2 tile pairs; ≈ 550 k patch/label pairs over 19 classes.
|
33 |
* **Auxiliary layer:**
|
34 |
-
* **SatCLIP location embedding
|
35 |
-
* **Splits:** Grid-based; val/test tiles lie outside the training footprint (spatial OOD by design). SatCLIP token lifts macro-F1 by ~3 pp across *all* subset sizes.
|
36 |
|
37 |
---
|
38 |
|
39 |
-
### USAVars – Tree-Cover Regression
|
40 |
* **Optical input:** NAIP RGB-NIR images (1 km² tiles); ≈ 100 k samples with tree-cover % labels.
|
41 |
* **Auxiliary layers:**
|
42 |
* Extended OSM raster stack (roads, buildings, land-use, biome classes, …)
|
@@ -44,10 +107,3 @@ This dataset card contains usage instructions and metadata for all data-products
|
|
44 |
* **Notes:** Stacking the OSM raster boosts R² by 0.16 in the low-data regime (< 250 images); DEM is provided raw for flexibility.
|
45 |
|
46 |
---
|
47 |
-
license: mit
|
48 |
-
task_categories:
|
49 |
-
- image-classification
|
50 |
-
- image-segmentation
|
51 |
-
tags:
|
52 |
-
- climate
|
53 |
-
---
|
|
|
1 |
+
---
|
2 |
+
# ======= 1) Basic info =======
|
3 |
+
pretty_name: "Geolayers"
|
4 |
+
language: en
|
5 |
+
language_creators:
|
6 |
+
- "found"
|
7 |
+
license: mit
|
8 |
+
multilinguality: monolingual
|
9 |
+
size_categories:
|
10 |
+
- 1K<n<100K
|
11 |
+
task_categories:
|
12 |
+
- image-classification
|
13 |
+
- image-segmentation
|
14 |
+
|
15 |
+
# ======= 2) How to cite =======
|
16 |
+
citation: |
|
17 |
+
@inproceedings{rao2025,
|
18 |
+
title={Using Multiple Input Modalities can Improve Data‐Efficiency and O.O.D. Generalization for ML with Satellite Imagery},
|
19 |
+
author={Arjun Rao and Esther Rolf},
|
20 |
+
year={2025},
|
21 |
+
booktitle={Under Review},
|
22 |
+
}
|
23 |
+
|
24 |
+
# ======= 3) Dataset structure =======
|
25 |
+
source_datasets:
|
26 |
+
- "SustainBench"
|
27 |
+
- "USAVars"
|
28 |
+
- "BigEarthNetv2.0"
|
29 |
+
- "EnviroAtlas"
|
30 |
+
|
31 |
+
# features:
|
32 |
+
# image:
|
33 |
+
# dtype: "uint8"
|
34 |
+
# shape: [3, 256, 256]
|
35 |
+
# osm_layers:
|
36 |
+
# dtype: "float32"
|
37 |
+
# shape: [4, 256, 256]
|
38 |
+
# label:
|
39 |
+
# ClassLabel:
|
40 |
+
# names: ["urban", "agriculture", "forest", "water", "bareground"]
|
41 |
+
|
42 |
+
# splits:
|
43 |
+
# train:
|
44 |
+
# name: "train"
|
45 |
+
# num_examples: 8000
|
46 |
+
# validation:
|
47 |
+
# name: "validation"
|
48 |
+
# num_examples: 1000
|
49 |
+
# test:
|
50 |
+
# name: "test"
|
51 |
+
# num_examples: 1000
|
52 |
+
|
53 |
+
# ======= 4) Other metadata =======
|
54 |
+
homepage: "https://huggingface.co/datasets/arjunrao2000/geolayers"
|
55 |
+
repository: "https://huggingface.co/datasets/arjunrao2000/geolayers"
|
56 |
+
download_size: 2.557e+10
|
57 |
+
tags:
|
58 |
+
- climate
|
59 |
+
---
|
60 |
+
|
61 |
+
|
62 |
## Geolayers-Data
|
|
|
63 |
|
64 |
+
This dataset card contains usage instructions and metadata for all data-products released with our paper:
|
65 |
+
*Using Multiple Input Modalities can Improve Data-Efficiency and O.O.D. Generalization for ML with Satellite Imagery.* We release 3 modified versions of 3 benchmark datasets spanning land-cover segmentation, tree-cover regression, and multi-label land-cover classification tasks. These datasets are augmented with auxiliary, geographic inputs. A full list of contributed data products is shown in the table below.
|
66 |
+
|
67 |
+
| **Dataset** | **Task Description** | **Multispectral Input** | **Model** | **Additional Data Layers** | **OOD Test Set Present?** |
|
68 |
+
|--------------------------------------|------------------------------------|-----------------------------|------------|-------------------------------------------------------|---------------------------|
|
69 |
+
| [SustainBench](https://arxiv.org/abs/2111.04724) | Farmland boundary delineation | Sentinel-2 RGB | U-Net | OSM rasters, EU-DEM | ✗ |
|
70 |
+
| [EnviroAtlas](https://arxiv.org/abs/2202.14000) | Land-cover segmentation | NAIP RGB + NIR | FCN | [Prior](https://arxiv.org/abs/2202.14000), OSM rasters | ✓ |
|
71 |
+
| [BigEarthNet v2.0](https://bigearth.net/static/documents/Description_BigEarthNet_v2.pdf) | Land-cover classification | Sentinel-2 (10 bands) | ViT | [SatCLIP](https://arxiv.org/abs/2311.17179) embeddings | ✓ |
|
72 |
+
| [USAVars](https://arxiv.org/abs/2010.08168) | Tree-cover regression | NAIP RGB + NIR | ResNet-50 | OSM rasters, DEM | ✗ |
|
73 |
|
74 |
## 📦 Datasets & Georeferenced Auxiliary Layers
|
75 |
|
76 |
+
### SustainBench – Farmland Boundary Delineation
|
77 |
+
* **Optical input:** Sentinel-2 RGB patches (224×224 px, 10 m GSD) covering French cropland in 2017; ≈ 1.6 k training images.
|
78 |
* **Auxiliary layers (all geo-aligned):**
|
79 |
* 19-channel OpenStreetMap (OSM) raster stack (roads, waterways, buildings, biome classes, …)
|
80 |
* EU-DEM (20 m GSD, down-sampled to 10 m)
|
81 |
+
* **Why:** OSM + DEM give an 8 % Dice boost when labels are scarce; gains appear once the training set drops below ≈ 700 images.
|
82 |
|
83 |
---
|
84 |
|
85 |
+
### EnviroAtlas – Land-Cover Segmentation
|
86 |
* **Optical input:** NAIP 4-band RGB-NIR aerial imagery at 1 m resolution.
|
87 |
* **Auxiliary layers:**
|
88 |
* OSM rasters (roads, waterbodies, waterways)
|
89 |
* **Prior** raster – a hand-crafted fusion of NLCD land-cover and OSM layers (PROC-STACK)
|
90 |
+
* **Splits:** Train = Pittsburgh; OOD validation/test = Austin & Durham. Auxiliary layers raise OOD overall accuracy by ~4 pp without extra fine-tuning.
|
91 |
|
92 |
---
|
93 |
|
94 |
+
### BigEarthNet v2.0 – Multi-Label Land-Cover Classification
|
95 |
* **Optical input:** 10-band Sentinel-2 tile pairs; ≈ 550 k patch/label pairs over 19 classes.
|
96 |
* **Auxiliary layer:**
|
97 |
+
* **SatCLIP** location embedding (256-D), one per image center, injected as an extra ViT token (TOKEN-FUSE).
|
98 |
+
* **Splits:** Grid-based; val/test tiles lie outside the training footprint (spatial OOD by design). SatCLIP token lifts macro-F1 by ~3 pp across *all* subset sizes.
|
99 |
|
100 |
---
|
101 |
|
102 |
+
### USAVars – Tree-Cover Regression
|
103 |
* **Optical input:** NAIP RGB-NIR images (1 km² tiles); ≈ 100 k samples with tree-cover % labels.
|
104 |
* **Auxiliary layers:**
|
105 |
* Extended OSM raster stack (roads, buildings, land-use, biome classes, …)
|
|
|
107 |
* **Notes:** Stacking the OSM raster boosts R² by 0.16 in the low-data regime (< 250 images); DEM is provided raw for flexibility.
|
108 |
|
109 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|