Civimatica/blue-marble-landcover

Blue Marble Landcover

Introduction

Blue Marble Landcover is a family of five 9-class land-cover segmentation models for Sentinel-2 imagery, from 33k 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

• 76.9 mIoU on SSL4EO-S12 LULC, held-out shard with Blue Marble medium (13.5M parameters, released checkpoint).
• 3.1 ms per 256x256 tile (medium) on the Arduino VENTUNO Q NPU, int8, every layer on the NPU.

Model Downloads

Model	Params	int8 size	VENTUNO Q latency	Ops on NPU	SSL4EO-S12 LULC, held-out shard: fp32 / int8	Download
Blue-Marble-Landcover-pico	0.033M	0.14 MB	1.0 ms	61/61	63.0 / 61.6	pico/
Blue-Marble-Landcover-nano	0.19M	0.36 MB	1.1 ms	75/75	72.4 / 72.2	nano/
Blue-Marble-Landcover-tiny	1.26M	1.62 MB	1.5 ms	117/117	74.2 / 74.5	tiny/
Blue-Marble-Landcover-small	2.4M	2.86 MB	1.7 ms	130/130	76.1 / 75.9	small/
Blue-Marble-Landcover-medium	13.5M	14.58 MB	3.1 ms	223/223	76.9 / 77.1	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 SSL4EO-S12 v1.1 validation shard 2 (2,048 images, one fixed 256x256 crop each), held out from training; shard 1 selected the checkpoint. SSL4EO-S12 has no test split. 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	SSL4EO-S12 LULC, held-out shard (mIoU)	LULC-2024 (mIoU)
Blue Marble pico	0.033M	63.0	38.8 ± 0.7
Blue Marble nano	0.19M	72.4	47.9 ± 0.2
Blue Marble tiny	1.26M	74.2	50.0 ± 0.4
Blue Marble small	2.4M	76.1	49.9 ± 0.9
Blue Marble medium	13.5M	76.9	51.1 ± 0.2
TerraMind-v1-tiny*	6.1M	77.6	50.2
MobileNetV4-S*	0.41M	61.3	40.0

Reference: larger models

Model	Params	SSL4EO-S12 LULC, held-out shard (mIoU)	LULC-2024 (mIoU)
TerraMind-v1-small*	22.9M	76.9	50.2

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. LULC-2024: held-out 2024 Sentinel-2 scenes with Esri 2024 land-cover maps. LULC-2024 was also used to choose between training recipes that tied on validation, so it is not an untouched test set and 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, 9, H, W] logits: water, trees, flooded vegetation, crops, built-up, bare ground, snow/ice, clouds, rangeland.

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 "zarr>=3" huggingface_hub
python evaluate.py --model medium/model_fp32.onnx      # or medium/model_int8.onnx

Expected mIoU for medium/: held-out shard 76.89 (fp32) / 77.13 (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}
}