debrajsingha/algaeguard-hab-ndci

AlgaeGuard — Harmful Algal Bloom Detection Dataset

A vision-language fine-tuning dataset for detecting and classifying Harmful Algal Blooms (HABs) from Sentinel-2 NDCI colormap imagery. Built to fine-tune edge-deployable VLMs for on-board satellite inference (LFM2.5-VL-450M / Liquid AI), delivering sub-90-minute bloom alerts to water utility operators.

Dataset Summary

Split	Samples	Real (SimSat)	Synthetic
train	350	180	170
eval	19	19	0
holdout	29	19	10
Total	398	218	180

Each sample is a VLM chat turn: the model receives a Sentinel-2 NDCI colormap image plus spectral metadata, and must produce a structured bloom assessment report.

Classes

Label	Bloom Coverage	Action
CLEAR	< 1%	Routine monitoring
LOW	1 – 8%	Increase monitoring, prepare protocol
MEDIUM	8 – 25%	Alert water quality team
HIGH	> 25%	Activate treatment protocol
CRITICAL	severe_pct > 15%	Emergency response

Dataset Structure

Fields

• image — Sentinel-2 NDCI colormap PNG (128×128 px). Color scale: blue = clear water, yellow/red = elevated chlorophyll-a (bloom).
• system_prompt — AlgaeGuard system role instruction
• user_text — Location, date, NDCI Mean/Max, Bloom Coverage, classification thresholds
• assistant_response — Structured bloom assessment: severity, NDCI stats, action recommendation
• severity — Ground-truth class label (CLEAR / LOW / MEDIUM / HIGH / CRITICAL)
• ndci_mean — Mean NDCI value over the scene
• ndci_max — Maximum NDCI value in the scene
• bloom_pct — % pixels above bloom threshold (NDCI > 0.0)
• severe_pct — % pixels above severe threshold (NDCI > 0.2)
• source — simsat_real | synthetic_template_v2 | synthetic_borderline_clear_v2 | synthetic_medium_holdout
• location — Geographic location (Chesapeake Bay, etc.) or "synthetic"
• timestamp — ISO-8601 acquisition timestamp

Sample

{
  "messages": [
    {
      "role": "system",
      "content": "You are AlgaeGuard, an autonomous on-board satellite AI..."
    },
    {
      "role": "user",
      "content": [
        {"type": "image", "image": "train_v2_l0_user_cb3d7802335e.png"},
        {"type": "text", "text": "Location: Chesapeake Bay\nDate: 2024-07-15\nNDCI Mean: -0.080 | NDCI Max: -0.010\nBloom Coverage: 0.00%\n..."}
      ]
    },
    {
      "role": "assistant",
      "content": "BLOOM ASSESSMENT — Status: CLEAR\n\nSeverity: CLEAR\n..."
    }
  ],
  "metadata": {
    "severity": "CLEAR",
    "ndci_mean": -0.08,
    "bloom_pct": 0.0,
    "source": "simsat_real"
  }
}

Source Data

Real observations

Extracted from the DPhi-Space SimSat API — a Sentinel-2 satellite simulator providing multispectral band arrays. NDCI (Normalised Difference Chlorophyll Index) is computed as:

NDCI = (B05 - B04) / (B05 + B04)

Synthetic data

Generated using Perlin-noise bloom maps injected into real NDCI backgrounds. Annotated using Claude API (Anthropic) for high-quality assistant responses. Synthetic samples augment under-represented severity classes.

Usage

from datasets import load_dataset

ds = load_dataset("debrajsingha/algaeguard-hab-ndci")

# Inspect a training sample
sample = ds["train"][0]
print(sample["severity"], sample["bloom_pct"])
sample["image"].show()

Fine-tuning with this dataset

This dataset is designed for SFT (supervised fine-tuning) of vision-language models. The native format (JSONL with image references) is compatible with:

• Liquid AI LEAP — LFM2.5-VL-450M fine-tuning
• LLaMA-Factory
• Hugging Face trl SFTTrainer

Citation

@dataset{algaeguard2026,
  title        = {AlgaeGuard: Harmful Algal Bloom Detection from Sentinel-2 NDCI Imagery},
  author       = {Singha, Debraj},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/debrajsingha/algaeguard-hab-ndci},
  license      = {CC-BY-4.0}
}

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Synthetic data generated with Anthropic Claude API. Real imagery derived from ESA Sentinel-2 data (open access under the Copernicus Open Access Hub terms).