metadata
license: apache-2.0
tags:
- computer-vision
- image-segmentation
- satellite-imagery
- remote-sensing
- pytorch
- uncertainty-estimation
datasets:
- 38-cloud
metrics:
- iou
- dice
- precision
- recall
- f1
🌍 AERIS — Cloud Detection for Landsat 8
AERIS (Automated Environmental Remote Imaging System) is a deep learning model for cloud segmentation in Landsat 8 satellite imagery with built-in uncertainty quantification using Monte Carlo Dropout.
The model performs high-accuracy binary segmentation on 4-channel satellite inputs (RGB + NIR).
🔎 Model Description
- Architecture: U-Net
- Encoder: ResNet34
- Input: 4-channel (Red, Green, Blue, Near-Infrared)
- Input Size: 256 × 256
- Output: Binary cloud segmentation mask
- Framework: PyTorch + segmentation-models-pytorch
- Uncertainty Estimation: MC Dropout (30 stochastic forward passes)
AERIS not only predicts cloud masks but also provides calibrated confidence estimates for more reliable remote sensing workflows.
📊 Performance
Evaluation on the 38-Cloud dataset:
| Metric | Score |
|---|---|
| Validation IoU | 92.20% |
| Dice Coefficient | 94.28% |
| Precision | 92.15% |
| Recall | 96.73% |
| F1 Score | 94.28% |
| Expected Calibration Error (ECE) | 0.70% |
Low ECE indicates strong confidence calibration.
🏋️ Training Details
- Dataset: 38-Cloud (16,800 training patches)
- Loss Function: Combined Dice + Binary Cross-Entropy
- Optimizer: AdamW (learning rate = 1e-4)
- Epochs: 30
- Hardware: NVIDIA RTX 4060
🚀 Usage
Installation
pip install torch torchvision
pip install segmentation-models-pytorch
Load Model
import torch
import segmentation_models_pytorch as smp
model = smp.Unet(
encoder_name="resnet34",
in_channels=4,
classes=1
)
model.load_state_dict(torch.load("Aeris_Model.pth", map_location="cpu"))
model.eval()
Inference
with torch.no_grad():
output = model(input_tensor) # input_tensor shape: [B, 4, 256, 256]
For uncertainty estimation, run multiple stochastic forward passes with dropout enabled.
📌 Intended Use
- Satellite image preprocessing
- Atmospheric correction pipelines
- Cloud cover estimation
- Weather monitoring systems
- Remote sensing research
⚠️ Limitations
- Trained only on Landsat 8 imagery
- Input size fixed at 256×256 patches
- Performance may degrade on unseen satellite domains
- Binary cloud detection (does not classify cloud types)
📜 License
This model is released under the Apache 2.0 License.
🤝 Contributions
Contributions, improvements, and research collaborations are welcome.