You can deploy this model and start earning money today!

debrajsingha/
algaeguard-lfm2-5-vl-450m

Name: algaeguard-lfm2-5-vl-450m
Availability: InStock

Views

No views yet

AlgaeGuard — LFM2.5-VL-450M Fine-tuned for Harmful Algal Bloom Detection

A fine-tuned vision-language model for detecting and classifying Harmful Algal Blooms (HABs) from Sentinel-2 NDCI colormap imagery. Designed for autonomous on-board satellite inference, delivering sub-90-minute bloom alerts to water utility operators.

Model Summary

Property	Value
Base model	LiquidAI/LFM2.5-VL-450M
Fine-tune method	LoRA SFT (merged)
Training hardware	Modal A10G GPU
Epochs	15
Learning rate	2e-5
Eval loss	0.066
Holdout accuracy	76.9% (20/26)
Dataset	debrajsingha/algaeguard-hab-ndci

Classes

Label	Bloom Coverage	Recommended 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

Evaluation Results (Holdout, 26 samples)

Overall accuracy: 76.9% (20/26)

Class	F1	Precision	Recall	Support
CLEAR	0.83	0.83	0.83	6
LOW	0.67	0.67	0.67	3
MEDIUM	0.87	0.83	0.91	11
HIGH	0.57	0.67	0.50	4
CRITICAL	0.50	0.50	0.50	2

Confusion Matrix

True \ Pred	CLEAR	LOW	MEDIUM	HIGH	CRITICAL
CLEAR	5	1	0	0	0
LOW	0	2	0	1	0
MEDIUM	1	0	10	0	0
HIGH	0	0	1	2	1
CRITICAL	0	0	1	0	1

All 6 errors are adjacent-class confusions — no wild misclassifications. The HIGH/CRITICAL boundary is the weakest point (3 of 6 errors).

Usage

python

from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image
import torch

model_id = "debrajsingha/algaeguard-lfm2-5-vl-450m"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Load your Sentinel-2 NDCI colormap image
image = Image.open("ndci_colormap.png").convert("RGB")

messages = [
    {
        "role": "system",
        "content": "You are AlgaeGuard, an autonomous on-board satellite AI for Harmful Algal Bloom (HAB) early-warning using Sentinel-2 NDCI imagery. You analyze NDCI colormap images and produce structured bloom assessments for water utility operators. Your assessments guide 6–12 hour treatment protocol decisions — be precise, actionable, and structured."
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Location: Chesapeake Bay\nDate: 2024-07-15\nNDCI Mean: 0.12 | NDCI Max: 0.41\nBloom Coverage: 18.5%\n\nClassification thresholds (bloom coverage):\n  CLEAR <1% | LOW 1–8% | MEDIUM 8–25% | HIGH >25% | CRITICAL if severe coverage >15%\nNote: Bloom Coverage <1% = CLEAR regardless of NDCI max spikes.\n\nAnalyze this Sentinel-2 NDCI colormap and issue an AlgaeGuard bloom assessment report for the water utility operator."}
        ]
    }
]

text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(text=text, images=[image], return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=256)

response = processor.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

Training Details

Data: 350 training samples from debrajsingha/algaeguard-hab-ndci — 180 real Sentinel-2 observations via SimSat API + 170 Perlin-noise synthetic bloom images annotated with Claude API.

Fine-tuning: LoRA adapters (rank=16) trained via Liquid AI LEAP SDK on Modal A10G, then merged into full weights. This checkpoint contains the fully merged model — no adapter files needed.

Input format: Sentinel-2 NDCI colormap PNG (any resolution; processor tiles to 512×512). The model expects NDCI colormaps where blue = clear water and yellow/red = elevated chlorophyll-a.

Limitations

Trained primarily on Chesapeake Bay observations — may generalise less well to other water bodies
HIGH/CRITICAL boundary is the weakest point (F1=0.50–0.57); use with human review for CRITICAL alerts
Requires pre-computed NDCI colormaps as input — does not process raw Sentinel-2 bands directly

Citation

bibtex

@model{algaeguard_lfm2_2026,
  title        = {AlgaeGuard: Fine-tuned LFM2.5-VL-450M for Harmful Algal Bloom Detection},
  author       = {Singha, Debraj},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/debrajsingha/algaeguard-lfm2-5-vl-450m},
  base_model   = {LiquidAI/LFM2.5-VL-450M},
  license      = {CC-BY-4.0}
}

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Based on LiquidAI/LFM2.5-VL-450M. Training data includes ESA Sentinel-2 imagery (Copernicus Open Access) and Claude API annotations (Anthropic).