📚 Dataset: SEN12MS-CR (Sentinel-12 Multi-Seasonal Cloud Removal)
🏗️ Architecture: U-Net với skip connections
📊 Input: S2 cloudy (4 bands) + S1 radar (2 bands) = 6 channels
🎯 Output: S2 clean (4 bands)
⏱️ Training time: ~2-3 hours (GPU) / ~20-30 hours (CPU)
🏗️ Architecture: U-Net với skip connections
📊 Input: S2 cloudy (4 bands) + S1 radar (2 bands) = 6 channels
🎯 Output: S2 clean (4 bands)
⏱️ Training time: ~2-3 hours (GPU) / ~20-30 hours (CPU)
⚙️ Cấu hình Training
🤖 Cloud Removal Models
Loading models...
📖 Cloud Removal Deep Learning Architectures
🔹 U-Net
Classic encoder-decoder with skip connections. Fast training, good baseline performance.
Recommended for beginners
🔹 CR-GAN
Cloud Removal GAN - adversarial training cho kết quả chân thực hơn.
Advanced
🔹 SpA-GAN
Spatial Attention GAN - attention mechanism tập trung vào vùng có mây.
Best quality
🔹 GLF-CR
Global-Local Fusion - kết hợp features global và local cho chi tiết tốt hơn.
High accuracy
🔹 SEN12MS-CR
Multi-modal fusion - kết hợp Sentinel-1 radar và Sentinel-2 optical.
Multi-sensor
🔹 RSDehazeNet
Remote Sensing Dehaze Network - chuyên cho ảnh viễn thám.
RS specialized
🔹 Cloud-Net
Encoder-Decoder architecture với residual connections.
Balanced
🔹 DSen2-CR
Deep Sentinel-2 Cloud Removal - tận dụng temporal information.
Temporal fusion