Core-S2RGB-249k-Qwen3-VL-Embedding

# Core-S2RGB-249k-Qwen3-VL-Embedding Geospatial-vision embedding dataset computed from **Core-S2RGB-249k** using the **Qwen3-VL-Embedding-2B** model. ## Overview | Property | Value | |----------|-------| | Source imagery | Core-S2RGB-249k (248,719 patches) | | Model | Qwen3-VL-Embedding-2B ([Qwen/Qwen3-VL-Embedding-2B](https://huggingface.co/Qwen/Qwen3-VL-Embedding-2B)) | | Input bands | RGB [B04, B03, B02] | | Embedding dimension | 2048 | | Output format | GeoParquet | | License | CC-BY-SA-4.0 | ## Computation Pipeline 1. **Pre-processing**: Each Sentinel-2 L2A patch is read from the source parquet files. The RGB bands [B04, B03, B02] are stacked and scaled by `2.5 × DN / 1e4`, clipped to [0, 1], to convert digital numbers to display-range reflectance. 2. **Resize**: The RGB tensor is interpolated to **384 × 384** pixels. 3. **Encoding**: The tensor is fed into Qwen3-VL-Embedding-2B (a 2B multimodal embedding model built on Qwen3-VL, whose vision encoder shares its architecture with the video-capable Qwen3-VL stack) to extract a 2048-dimensional image embedding. No instruction prompt is used for image documents, matching the embedder's default. 4. **Post-processing**: Embeddings are L2-normalised by the model encoder, so a dot product between rows equals cosine similarity. 5. **Geospatial metadata**: The original UTM footprint is reprojected to EPSG:4326 (WGS-84) to obtain the `geometry`, `centre_lat`, and `centre_lon` fields. Additional metadata (`product_id`, `grid_cell`, `timestamp`, `utm_crs`, `pixel_bbox`) is preserved from the source dataset. ## File Layout ``` . ├── Qwen3VL_2B_crop_384x384.parquet # Main embedding GeoParquet (248,719 rows) └── README.md ``` ## Schema | Column | Type | Description | |--------|------|-------------| | `unique_id` | string | SHA-256 hash of geometry + timestamp + product_id + embedding | | `embedding` | list<float> | 2048-dim Qwen3-VL-Embedding feature vector | | `timestamp` | datetime | Acquisition timestamp | | `product_id` | string | Original Sentinel-2 product identifier | | `grid_cell` | string | Major-TOM grid cell identifier | | `grid_row_u` | int16 | Grid row index | | `grid_col_r` | int16 | Grid column index | | `geometry` | geometry | WGS-84 polygon (footprint) | | `centre_lat` | float32 | Latitude of patch centre | | `centre_lon` | float32 | Longitude of patch centre | | `utm_footprint` | string | Original UTM footprint as WKT | | `utm_crs` | string | UTM CRS (e.g. EPSG:32633) | | `pixel_bbox` | list<int> | Pixel bounding box [x_min, y_min, x_max, y_max] | | `parquet_url` | string | Source parquet file path in the image dataset | | `parquet_row` | int64 | Row index within the source parquet file | ## Usage ```python import pandas as pd df = pd.read_parquet("Qwen3VL_2B_crop_384x384.parquet") print(len(df), "embeddings") print(df.iloc[0].embedding.shape) # (2048,) ``` ## Citation If you use this embedding dataset, please cite the original Major-TOM paper and the Qwen3-VL-Embedding paper: ```bibtex @article{qwen3vlembedding, title={Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking}, author={Li, Mingxin and Zhang, Yanzhao and Long, Dingkun and Chen Keqin and Song, Sibo and Bai, Shuai and Yang, Zhibo and Xie, Pengjun and Yang, An and Liu, Dayiheng and Zhou, Jingren and Lin, Junyang}, journal={arXiv preprint arXiv:2601.04720}, year={2026} } ``` ```bibtex @inproceedings{francis2024majortom, title={Major TOM: Expandable Datasets for Earth Observation}, author={Francis, Alistair and Czerkawski, Mikolaj}, year={2024}, booktitle={IGARSS 2024}, eprint={2402.12095}, archivePrefix={arXiv} } ```