Civimatica/blue-marble-flood

Blue Marble Flood

91.9 mIoU on Sen1Floods11 from Sentinel-2 alone, tying the best published result (an unreleased Sentinel-1+2 TerraMind-Tiny variant), and 89.7 mIoU on the held-out Bolivia split, above the best published result (Prithvi-CAFE, 88.9).

Introduction

Blue Marble Flood is a family of five flood and surface-water 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

• 91.9 mIoU on Sen1Floods11 test with Blue Marble nano (0.19M parameters, released checkpoint); best published model under 15M: 91.9, TerraMind-Tiny, patch 4 (unreleased; S1+S2).
• 89.7 mIoU on Sen1Floods11 Bolivia (held-out) with Blue Marble nano (0.19M parameters, released checkpoint); best published model under 15M: 83.0, Segformer-B0 (3.7M); best published of any size: 88.9, Prithvi-CAFE (frozen Prithvi 2.0 600M backbone).
• 3.4 ms per 512x512 tile (nano) on the Arduino VENTUNO Q NPU, int8, every layer on the NPU.

Model Downloads

Model	Params	int8 size	VENTUNO Q latency	Ops on NPU	Sen1Floods11 test: fp32 / int8	Download
Blue-Marble-Flood-pico	0.032M	0.14 MB	2.8 ms	61/61	91.0 / 90.9	pico/
Blue-Marble-Flood-nano	0.19M	0.36 MB	3.4 ms	75/75	91.9 / 91.8	nano/
Blue-Marble-Flood-tiny	1.26M	1.62 MB	4.5 ms	116/116	90.9 / 90.7	tiny/
Blue-Marble-Flood-small	2.4M	2.86 MB	5.2 ms	129/129	90.2 / 90.1	small/
Blue-Marble-Flood-medium	13.5M	14.58 MB	10.1 ms	223/223	91.0 / 90.8	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 full official Sen1Floods11 test split (90 chips) and the Bolivia hold-out (15 chips). 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	Sen1Floods11 test (mIoU)	Sen1Floods11 Bolivia (held-out) (mIoU)	Water-2026 (mIoU)	STURM-Flood (mIoU)
Blue Marble pico	0.032M	91.0	87.7	86.0 ± 0.8	76.7 ± 0.3
Blue Marble nano	0.19M	91.9	89.7	85.2 ± 0.9	77.2 ± 0.3
Blue Marble tiny	1.26M	90.9	89.7	88.0 ± 0.7	77.6 ± 0.3
Blue Marble small	2.4M	90.2	88.1	87.7 ± 0.3	76.4 ± 0.1
Blue Marble medium	13.5M	91.0	89.2	87.3 ± 0.2	77.0 ± 0.2
PANGAEA UNet (S1+S2)	14.8M	91.4	-	-	-
Segformer-B0	3.7M	89.4	83.0	-	-
EfficientViT-B0	0.7M	87.7	-	-	-
TerraMind-Tiny, patch 4 (unreleased; S1+S2)	Tiny (count not published)	91.9	-	-	-
TerraMind-Tiny (S1+S2)	Tiny (count not published)	90.2	-	-	-
THOR-Tiny (S1+S2)	7.6M	90.2	-	-	-
TerraMind-v1-tiny*	6.1M	88.4	88.4	88.2	77.2
MobileNetV4-S*	0.41M	90.3	81.9	69.2	75.9

Reference: larger models

Model	Params	Sen1Floods11 test (mIoU)	Sen1Floods11 Bolivia (held-out) (mIoU)	Water-2026 (mIoU)	STURM-Flood (mIoU)
TerraMind TiM	not disclosed	91.1	-	-	-
SpectralEarth-FM-a	not disclosed	91.1	-	-	-
Prithvi-CAFE (frozen Prithvi 2.0 600M backbone)	45.5M trainable	90.5	88.9	-	-
Prithvi 2.0 300M + LoRA (frozen backbone)	5.5M trainable	90.0	87.6	-	-
U-Prithvi	not disclosed	89.7	87.7	-	-
DeCUR, full fine-tuning	25M	86.9	85.8	-	-
TerraMind-v1-small*	22.9M	88.0	88.7	89.5	77.3

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. Water-2026: our held-out Sentinel-2 scenes from 2025-2026 with JRC water labels. STURM-Flood: public Sentinel-2 flood dataset (CC BY 4.0), scored zero-shot (no STURM data in training); it is surface reflectance with 9 bands, the others set to 0. Water-2026 and STURM-Flood were also used to choose between training recipes that tied on validation, so they are not untouched test sets and are 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 land, 1 water.

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 nano/model_fp32.onnx      # or nano/model_int8.onnx

Expected mIoU for nano/: test 91.92 (fp32) / 91.79 (int8); Bolivia 89.70 (fp32) / 89.72 (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}
}