Datasets:
Commit
·
687a219
1
Parent(s):
4469ce0
Update README
Browse files
README.md
CHANGED
|
@@ -21,25 +21,16 @@ Samples from the TerraMesh dataset with seven spatiotemporal aligned modalities.
|
|
| 21 |
|
| 22 |
---
|
| 23 |
|
| 24 |
-
## Key features
|
| 25 |
-
|
| 26 |
-
| Aspect | Value |
|
| 27 |
-
| ---------- |--------------------------------------------------------------------------------------------------------------------------------|
|
| 28 |
-
| Modalities | S‑2 L1C (13 bands), S‑2 L2A (12 bands), S‑2 RGB (3 bands), S‑1 GRD (VV, VH), S‑1 RTC (VV, VH), NDVI, Copernicus DEM, ESRI LULC |
|
| 29 |
-
| Samples | **9 089 536** train · **89 088** val |
|
| 30 |
-
| Patch size | 264 × 264 pixel (10 m grid) |
|
| 31 |
-
| Years | 2017 – 2024 (peak 2019 ‑ 2023) |
|
| 32 |
-
| License | CC‑BY‑SA‑4.0 |
|
| 33 |
-
|
| 34 |
-
---
|
| 35 |
-
|
| 36 |
## Dataset organisation
|
| 37 |
|
| 38 |
The archive ships two top‑level splits `train/` and `val/`, each holding one folder per modality. More details with the dataset release end of June.
|
| 39 |
|
| 40 |
---
|
| 41 |
|
| 42 |
-
##
|
|
|
|
|
|
|
|
|
|
| 43 |
|
| 44 |
Heat map of the sample count in a one-degree grid. | Monthly distribution of all S-2 timestamps.
|
| 45 |
:-------------------------:|:-------------------------:
|
|
@@ -54,7 +45,7 @@ Heat map of the sample count in a one-degree grid. | Monthly distribution of al
|
|
| 54 |
TerraMesh was used to pre-train [TerraMind-B](https://huggingface.co/ibm-esa-geospatial/TerraMind-1.0-base).
|
| 55 |
On the six evaluated segmentation tasks from PANGAEA bench, TerraMind‑B reaches an average mIoU of 66.6%, the best overall score with an average rank of 2.33. This amounts to roughly a 3pp improvement over the next‑best open model (CROMA), underscoring the benefits of pre‑training on TerraMesh.
|
| 56 |
Compared to an ablation model pre-trained only on SSL4EO-S12 locations TerraMind-B performs overall 1pp better with better global generalization on more remote tasks like CTM-SS.
|
| 57 |
-
|
| 58 |
---
|
| 59 |
|
| 60 |
## Citation
|
|
|
|
| 21 |
|
| 22 |
---
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
## Dataset organisation
|
| 25 |
|
| 26 |
The archive ships two top‑level splits `train/` and `val/`, each holding one folder per modality. More details with the dataset release end of June.
|
| 27 |
|
| 28 |
---
|
| 29 |
|
| 30 |
+
## Description
|
| 31 |
+
|
| 32 |
+
TerraMesh fuses complementary optical, radar, topographic and thematic layers into pixel‑aligned 10 m cubes, allowing models to learn joint representations of land cover, vegetation dynamics and surface structure at planetary scale.
|
| 33 |
+
The dataset is globally distributed and covers multiple years.
|
| 34 |
|
| 35 |
Heat map of the sample count in a one-degree grid. | Monthly distribution of all S-2 timestamps.
|
| 36 |
:-------------------------:|:-------------------------:
|
|
|
|
| 45 |
TerraMesh was used to pre-train [TerraMind-B](https://huggingface.co/ibm-esa-geospatial/TerraMind-1.0-base).
|
| 46 |
On the six evaluated segmentation tasks from PANGAEA bench, TerraMind‑B reaches an average mIoU of 66.6%, the best overall score with an average rank of 2.33. This amounts to roughly a 3pp improvement over the next‑best open model (CROMA), underscoring the benefits of pre‑training on TerraMesh.
|
| 47 |
Compared to an ablation model pre-trained only on SSL4EO-S12 locations TerraMind-B performs overall 1pp better with better global generalization on more remote tasks like CTM-SS.
|
| 48 |
+
More details in our [paper](https://arxiv.org/abs/2504.11172).
|
| 49 |
---
|
| 50 |
|
| 51 |
## Citation
|