Core-S2L2A-249k-OlmoEarth-v1_2-Base

# Core-S2L2A-249k-OlmoEarth-v1_2-Base This dataset contains **248,719 pre-computed image embeddings** for the [Major-TOM Core-S2L2A-249k](https://modelscope.cn/datasets/Major-TOM/Core-S2L2A-249k) Sentinel-2 L2A subset. The embeddings are generated with [OlmoEarth-v1_2-Base](https://modelscope.cn/models/allenai/OlmoEarth-v1_2-Base) and use each image's real acquisition timestamp as the model's temporal input. ## Model and preprocessing - **Model:** OlmoEarth-v1_2-Base - **Encoder parameters:** 114M - **Embedding dimension:** 768 - **Input:** 12 Sentinel-2 L2A bands - **Source chip size:** 384 x 384 pixels - **Encoder input size:** 128 x 128 pixels - **Patch size:** 8 - **Source ground sample distance:** 10 m - **Effective encoder ground sample distance:** 30 m after resizing 384 x 384 to 128 x 128 - **Pooling:** mean over unmasked encoder tokens - **Output normalization:** none; cosine normalization is applied at retrieval time The Major-TOM bands are read and returned in the order expected by OlmoEarth: ```text B02, B03, B04, B08, B05, B06, B07, B8A, B11, B12, B01, B09 ``` Bands stored at 20 m or 60 m resolution are upsampled to the 10 m reference grid with nearest-neighbor interpolation. The 12-band image is normalized with the official OlmoEarth pretraining normalizer and encoded as a single Sentinel-2 timestep. ## Temporal metadata OlmoEarth v1.2 uses 3D rotary positional embeddings (RoPE) with a real calendar time coordinate. For every source image, the acquisition timestamp is converted to: ```text (day, zero-based month, Gregorian year) ``` For example, `20221115T161819` is passed to the model as `(15, 10, 2022)`. The timestamp stored in this dataset is therefore part of the embedding input, not only descriptive metadata. Online query embeddings must use the same product timestamp to reproduce the indexed vector. ## Generation The embeddings were generated with [`generate_embeddings.py`](https://github.com/OpenGeoScope/EarthEmbeddingExplorer/blob/main/generate_embeddings.py) from EarthEmbeddingExplorer using `olmoearth-pretrain-minimal==0.0.7` and the official OlmoEarth-v1_2-Base checkpoint. Each generated record retains the corresponding source `parquet_url` and the row-group index in the 249k subset. These fields can be used with HTTP range requests to download the exact source thumbnail or all 12 GeoTIFF bands. ## Files ```text OlmoEarth_v1_2_Base_crop_384x384.parquet ``` - **Rows:** 248,719 - **File size:** 824,275,960 bytes - **SHA-256:** `978d36cb2b579021ad3f00e22522b0271d19748455e8814979ca24fda2d31f85` ## Columns | Column | Type | Description | | :--- | :--- | :--- | | `unique_id` | string | SHA-256 checksum of geometry, timestamp, product ID, and embedding | | `embedding` | list<float32> | 768-dimensional OlmoEarth-v1_2-Base feature vector | | `timestamp` | string | Sentinel-2 acquisition timestamp used by the encoder | | `product_id` | string | Original Sentinel-2 product identifier | | `grid_cell` | string | Major-TOM grid cell identifier | | `grid_row_u` | int16 | Major-TOM grid row index | | `grid_col_r` | int16 | Major-TOM grid column index | | `geometry` | geometry | EPSG:4326 footprint polygon | | `centre_lat` | float32 | Latitude of the chip center | | `centre_lon` | float32 | Longitude of the chip center | | `utm_footprint` | string | Original UTM footprint as WKT | | `utm_crs` | string | Original UTM CRS | | `pixel_bbox` | list<int> | Pixel bounding box `[x_min, y_min, x_max, y_max]` | | `parquet_url` | string | URL of the source Core-S2L2A-249k Parquet file | | `parquet_row` | int64 | Row-group index in the source subset Parquet file | ## Usage ```python import numpy as np import pandas as pd df = pd.read_parquet("OlmoEarth_v1_2_Base_crop_384x384.parquet") embeddings = np.stack(df["embedding"].to_numpy()) timestamps = df["timestamp"] ``` The accompanying [EarthEmbeddingExplorer](https://github.com/OpenGeoScope/EarthEmbeddingExplorer) application implements timestamp-aware online image retrieval against this file. ## Acknowledgements - [OlmoEarth](https://github.com/allenai/olmoearth_pretrain), Allen Institute for AI - [Major-TOM](https://github.com/ESA-PhiLab/MajorTOM), European Space Agency - [rslearn OlmoEarth embedding guide](https://github.com/allenai/rslearn/blob/master/docs/examples/OlmoEarthEmbeddings.md) ## License and responsible use The Major-TOM-derived dataset is shared under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/). OlmoEarth model weights and derivatives are additionally governed by the [OlmoEarth Artifact License](https://huggingface.co/allenai/OlmoEarth-v1_2-Base/blob/main/LICENSE.txt) and the [Ai2 Responsible Use Guidelines](https://allenai.org/responsible-use).