kreyesp/beavercube-cloud-segmentation

BeaverCube Cloud Segmentation — SmallCloudNet

U-Net trained to detect clouds in imagery simulated from the Matrix Vision mvBlueFOX-IGC-200w camera on the BeaverCube 2 CubeSat (MIT), using the CloudSEN12-L1C Sentinel-2 dataset.

Model details

Property	Value
Architecture	U-Net (SmallCloudNet)
Parameters	1.86M
Input size (training)	33×33 px
Input size (inference)	any (fully convolutional)
Classes	clear, thick cloud, thin cloud, shadow

Performance

Metric	Value
Mean IoU	0.38
Mean F1	0.54
Accuracy	64%
Clear IoU	0.58
Thick cloud IoU	0.45
Shadow IoU	0.27
Thin cloud IoU	0.22

Usage

import torch
from model import SmallCloudNet

model = SmallCloudNet(in_ch=3, num_classes=4)
checkpoint = torch.load("best_model.pth", map_location="cpu")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# img: float32 tensor (1, 3, H, W) normalised to [0, 1]
with torch.no_grad():
    logits = model(img)               # (1, 4, H, W)
    mask   = logits.argmax(dim=1)     # (1, H, W)

Training data

CloudSEN12-L1C (Sentinel-2 L1C), preprocessed to simulate BlueFOX GSD (153.75 m) via:

1. Gaussian PSF blur (σ = 1.6 px, derived from Kowa 16mm lens Airy disk)
2. 15.65× INTER_AREA downsample (512×512 → 33×33)
3. Read + shot noise augmentation applied each epoch