SAR-Based Tropical Deforestation Detection
Sentinel-1 temporal features and U-Net variants for tropical deforestation segmentation.
2026
Five temporal feature maps · three PyTorch segmentation models.
Radar sequences to deforestation masks
SAR-Based Tropical Deforestation Detection uses Sentinel-1 VV/VH observations and temporal features for spatial change segmentation. The recovered PyTorch/Colab notebook works with BraDD-S1TS data and compares three U-Net families. Radar backscatter supplies input for cloud-prone tropical monitoring.
The benchmark shares feature preparation and training structure while varying how models collect spatial context.
Radar sequence to change mask
- VV / VH sequences
- Temporal features
- Normalization
- U-Net variants
- Evaluation & maps
The notebook’s feature extractor combines current backscatter with change, variation and anomaly information.
Segmentation architectures
Scroll horizontally to compare.
| Model | Spatial-context approach |
|---|---|
| ResNet-34 Dilated U-Net | Residual encoder with dilated context and an encoder–decoder segmentation path. |
| ResNet-50 U-Net | A deeper residual encoder feeding the segmentation decoder. |
| ASPP U-Net | Atrous spatial pyramid pooling combines features at different receptive-field scales. |
Engineering decisions
The feature extractor produces five maps: latest VH, latest VV, VH temporal change, VH standard deviation and latest VH deviation from its temporal mean. Recent backscatter, difference, variation and anomaly describe complementary aspects of the same sequence. Dataset/DataLoader preparation, normalization and augmentation supply consistent tensors.
AdamW, mixed precision and checkpoint saving support the comparison. The loss combines weighted binary classification with Tversky overlap, connecting pixel prediction to class-imbalance handling. Saved weights are reloaded for qualitative input/target/predicted-mask inspection. Segmentation localizes change boundaries; regional patch classification is a distinct task in the broader study. Residual encoders, dilation and ASPP offer different spatial-context mechanisms within the shared workflow.