| --- |
| license: mit |
| task_categories: |
| - image-classification |
| - visual-question-answering |
| language: |
| - en |
| size_categories: |
| - n<1K |
| tags: |
| - satellite-imagery |
| - earth-observation |
| - orbital-triage |
| - vlm |
| - lfm2 |
| - liquid-ai |
| pretty_name: ORION Satellite Triage Dataset |
| --- |
| |
| # Dataset |
|
|
| The ORION dataset is a curated collection of satellite imagery and triage labels used to fine-tune the VLM for orbital image classification. Images are fetched from SimSat's Mapbox API and paired with classification prompts and ground-truth labels. |
|
|
| ## Dataset Structure |
|
|
| ``` |
| images/ |
| low_ocean_pacific_nemo.png |
| med_city_chicago.png |
| high_port_rotterdam.png |
| ... |
| train_dataset.jsonl |
| val_dataset.jsonl |
| test_dataset.jsonl |
| ``` |
|
|
| - **images/**: 512x512 RGB satellite images fetched from SimSat |
| - **train_dataset.jsonl**: Training samples (240 targets x 2 coordinate dropout = 480 records) |
| - **val_dataset.jsonl**: Validation samples (60 targets, always with coordinates; used for eval_loss + best checkpoint selection) |
| - **test_dataset.jsonl**: Test samples (60 targets, always with coordinates; held out for ablation and evaluation) |
| |
| ## Target Definitions |
| |
| 360 target locations organized by triage priority and visual morphology: |
| |
| | Class | Count | Visual Morphology | |
| | ------ | ----- | -------------------------------------------------------------------------------------------------- | |
| | LOW | 120 | Featureless natural terrain: oceans, deserts, ice sheets, dense canopy, geological formations | |
| | MEDIUM | 120 | Standard human civilization: urban grids, suburban sprawl, agriculture, towns, infrastructure | |
| | HIGH | 120 | Strategic anomalies: mega-ports, mega-airports, energy/dams, mega-mines, military/space facilities | |
| |
| ### LOW Morphologies |
| |
| 1. Standard voids: oceans and water bodies |
| 2. Standard voids: deserts, ice sheets, dense canopy |
| 3. Hard LOW: coastlines and boundaries that resemble artificial structures |
| 4. Hard LOW: geological anomalies (craters, calderas) that mimic mines |
| 5. Hard LOW: fractals and textures (deltas, salt flats, reefs) that mimic city streets |
| |
| ### MEDIUM Morphologies |
| |
| 1. Urban grids: dense city centers worldwide |
| 2. Suburban sprawl: low-density residential areas |
| 3. Agriculture: crop fields, pivot irrigation, terracing |
| 4. Standard infrastructure: regional airports, rail yards, commercial zones |
| 5. Towns and settlements: isolated clusters in varied terrain |
| |
| ### HIGH Morphologies |
| |
| 1. Mega-ports: Rotterdam, Singapore, LA, etc. |
| 2. Mega-airports: Atlanta, Denver, Dubai, Daxing, etc. |
| 3. Energy and dams: nuclear plants, solar farms, hydroelectric dams |
| 4. Mega-mines and extreme industrial: open-pit mines, refineries |
| 5. Space, military, and chokepoints: launch pads, naval bases, canals |
| |
| ## Generation Process |
| |
| [`data_gen.py`](https://Saransh-cpp.github.io/ORION/guides/data-gen/) generates the dataset by: |
| |
| 1. **Proximity filter**: removes targets closer than 2 km to each other (Haversine distance) to avoid duplicate imagery |
| 2. **Shuffle and split**: deterministic 3-way IID split (`random.seed(42)`): 240 train / 60 val / 60 test |
| 3. **Image fetch**: for each target, fetches a 512x512 satellite image from SimSat's Mapbox static image API at `GET http://localhost:9005/data/image/mapbox` |
| 4. **JSONL generation**: creates conversation-format records for fine-tuning |
| |
| ### JSONL Record Format |
| |
| Each line in the JSONL files is a JSON object: |
| |
| ```json |
| { |
| "image": "orion_dataset/images/high_port_rotterdam.png", |
| "conversations": [ |
| { |
| "role": "user", |
| "content": "<image>\nYou are an autonomous orbital triage assistant..." |
| }, |
| { |
| "role": "assistant", |
| "content": "{\"reason\": \"Extreme-density geometric cargo terminals...\", \"category\": \"HIGH\"}" |
| } |
| ] |
| } |
| ``` |
| |
| ## Coordinate Dropout Augmentation |
|
|
| For training samples, each target produces two JSONL records: one with GPS coordinates in the prompt and one without. This 50% coordinate dropout teaches the model to classify based on visual features alone, making it robust when GPS data is noisy or unavailable. |
|
|
| ## Prompt Template |
|
|
| The prompt matches the ChatML format used by the fine-tuned model and the on-board VlmInferenceEngine: |
|
|
| ``` |
| You are an autonomous orbital triage assistant. Analyze this high-resolution RGB |
| satellite image captured at Longitude: X, Latitude: Y. |
| Strictly use one of these categories based on visual morphology: |
| - HIGH: ... |
| - MEDIUM: ... |
| - LOW: ... |
| You MUST output your response as a valid JSON object. To ensure accurate visual |
| reasoning, you must output the "reason" key FIRST, followed by the "category" key. |
| ``` |
|
|
| ## Expected Model Output |
|
|
| ```json |
| { |
| "reason": "Extreme-density geometric cargo terminals and massive vessel berthing.", |
| "category": "HIGH" |
| } |
| ``` |
|
|
| ## Related |
|
|
| - [Fine-tuned model](https://huggingface.co/Saransh-cpp/orion-qlora-lfm2.5-vl-1.6b) |
| - [Training guide](https://Saransh-cpp.github.io/ORION/guides/training/) |
| - [Training pipeline](https://Saransh-cpp.github.io/ORION/architecture/ground_segment/training/) |
| - [Data flow architecture](https://Saransh-cpp.github.io/ORION/architecture/flight_segment/data-flow/) |
|
|
