Civimatica/blue-marble-burn

Blue Marble Burn

Introduction

Blue Marble Burn is a family of five burn-scar segmentation models for Sentinel-2 imagery, from 32k to 13.5M parameters. Every model is under 15M parameters and runs fully in int8 on the Qualcomm Dragonwing QCS8275 NPU (Arduino VENTUNO Q).

Highlights

• 89.5 mIoU on HLS Burn Scars (validation split) with Blue Marble small (2.4M parameters, released checkpoint).
• 88.1 mIoU at 191k parameters.
• 5.2 ms per 512x512 tile (small) on the Arduino VENTUNO Q NPU, int8, every layer on the NPU.

Model Downloads

Model	Params	int8 size	VENTUNO Q latency	Ops on NPU	HLS Burn Scars (validation split): fp32 / int8	Download
Blue-Marble-Burn-pico	0.032M	0.14 MB	2.7 ms	61/61	86.4 / 86.0	pico/
Blue-Marble-Burn-nano	0.19M	0.36 MB	3.4 ms	75/75	88.1 / 87.9	nano/
Blue-Marble-Burn-tiny	1.26M	1.62 MB	4.5 ms	116/116	89.0 / 88.9	tiny/
Blue-Marble-Burn-small	2.4M	2.86 MB	5.2 ms	129/129	89.5 / 89.4	small/
Blue-Marble-Burn-medium	13.5M	14.58 MB	10.1 ms	223/223	89.0 / 87.7	medium/

Each size folder contains:

File	Use
model_qcs8275.dlc	Compiled for the Qualcomm QCS8275 NPU (QNN runtime): runs as-is on the Arduino VENTUNO Q
model_int8.onnx	int8 ONNX, the source of the compiled model; ONNX Runtime or your own NPU toolchain
model_fp32.onnx	Full precision ONNX (CPU / GPU); the scores below are for this model
model.safetensors + config.json	PyTorch weights for fine-tuning, with models.py in this repository
ventuno_q_profile.json	On-device profile on the Arduino VENTUNO Q (Qualcomm AI Hub): the compute unit of every op, total latency, memory, load times

Scores in this table are for the exact files in this repository, re-scored on the official HLS Burn Scars validation split (264 scenes). Latency and ops on NPU: on-device profile of model_qcs8275.dlc on an Arduino VENTUNO Q (Qualcomm AI Hub).

Evaluation Results

Models under 15M parameters

Model	Params	HLS Burn Scars (validation split) (mIoU)
Blue Marble pico	0.032M	86.4
Blue Marble nano	0.19M	88.1
Blue Marble tiny	1.26M	89.0
Blue Marble small	2.4M	89.5
Blue Marble medium	13.5M	89.0
TerraMind-v1-tiny*	6.1M	88.6
MobileNetV4-S*	0.41M	85.6

Blue Marble, public benchmarks: score of the released checkpoint in this repository (re-scored from these files). Blue Marble, our own benchmarks: mean ± std over 3 training runs. Other models: published numbers, with inputs noted where they differ from ours (our models use Sentinel-2 only); * = trained and evaluated by us on the same data and split (parameter count includes our decoder). Best per column among models under 15M in bold. HLS Burn Scars has no separate test split: scores are on the official validation split, which was also used to select each checkpoint, so that column is not bolded.

Quickstart

import numpy as np, onnxruntime as ort

MEAN = np.array([2357.089, 2137.385, 2018.788, 2082.986, 2295.651, 2854.537, 3122.849, 3040.560, 3306.481, 1473.847, 506.070, 2472.825, 1838.929], np.float32)[:, None, None]
STD  = np.array([1624.683, 1675.806, 1557.708, 1833.702, 1823.738, 1733.977, 1732.131, 1679.732, 1727.26, 1024.687, 442.165, 1331.411, 1160.419], np.float32)[:, None, None]

def prepare(dn):                       # dn: Sentinel-2 L1C digital numbers, [13, H, W], H and W multiples of 16
    dn = np.clip(dn - 1000, 0, None)   # only for products from 2022 on (processing baseline 04.00+)
    return ((dn - MEAN) / STD)[None].astype(np.float32)

sess = ort.InferenceSession("tiny/model_int8.onnx")
out = sess.run(None, {"image": prepare(dn)})[0]

PyTorch:

import json, torch
from safetensors.torch import load_file
from models import BlueMarble                     # models.py from this repository

cfg = json.load(open("tiny/config.json"))
model = BlueMarble(cfg["arch"], cfg["num_classes"]).eval()
model.load_state_dict(load_file("tiny/model.safetensors"))

• Input: 13 Sentinel-2 L1C bands, B01 to B12 including B8A and B10. Missing bands (other sensors): set to 0 after normalisation.
• Output: [1, 2, H, W] logits: 0 unburned, 1 burned.

Reproduce

evaluate.py re-scores any model file in this repository on the public benchmark; the data is downloaded on the first run.

pip install numpy onnxruntime rasterio huggingface_hub
python evaluate.py --model small/model_fp32.onnx      # or small/model_int8.onnx

Expected mIoU for small/: validation 89.50 (fp32) / 89.44 (int8).

License

Apache-2.0. Contains modified Copernicus Sentinel data.

Citation

@misc{bluemarble2026,
  title  = {Blue Marble: Edge Earth Observation Models},
  author = {Harsh Kumar},
  year   = {2026},
  url    = {https://huggingface.co/Civimatica}
}