How karthikram-p/marine-debris-detection works
AquaScan is a marine-debris semantic-segmentation project built around Sentinel-2 imagery and UNet++. The principal user workflow is uploading an image in the Streamlit app, preprocessing it, running the model, and downloading a GeoTIFF mask. Separate training and evaluation workflows use MARIDA patches, augmentation, custom loss, metrics, and saved checkpoints. The graph includes documented capabilities, but some wiring—especially app internals and unsampled training/evaluation code—is represented only where the README or source index supports it.
Main entry points: train.py Training loop · evaluation.py Evaluation runner · dataloader.py MARIDA loader · app.py Streamlit app
Components 14
Web inference
- Streamlit app
semantic_segmentation/unet/app.py - UNet++ model
semantic_segmentation/unet/unet_plus_plus.py - Prediction mask
Training
- Training loop
semantic_segmentation/unet/train.py - Dice loss
semantic_segmentation/unet/loss_functions.py - Model checkpoints
semantic_segmentation/unet/trained_models
Evaluation
- Evaluation runner
semantic_segmentation/unet/evaluation.py - U-Net model
semantic_segmentation/unet/unet.py - Segmentation metrics
utils/metrics.py - Saved predictions
data/predicted_unet
Dataset and domain
- MARIDA loader
semantic_segmentation/unet/dataloader.py - Satellite patches
data/patches - Class mappings
utils/assets.py
Other components
- Image uploader
Connections 15
- Image uploader → Streamlit app: uploads image
README.md - Streamlit app → UNet++ model: runs inference
semantic_segmentation/unet/app.py - Streamlit app → Prediction mask: writes GeoTIFF
semantic_segmentation/unet/app.py - Prediction mask → Image uploader: downloads mask
- Training loop → MARIDA loader: loads batches
semantic_segmentation/unet/train.py - MARIDA loader → Satellite patches: reads imagery
semantic_segmentation/unet/dataloader.py - MARIDA loader → Class mappings: maps classes
- Training loop → UNet++ model: trains model
semantic_segmentation/unet/train.py - Training loop → Dice loss: computes loss
semantic_segmentation/unet/train.py - Training loop → Segmentation metrics: scores model
semantic_segmentation/unet/train.py - Training loop → Model checkpoints: saves weights
semantic_segmentation/unet/train.py - Evaluation runner → MARIDA loader: loads test set
semantic_segmentation/unet/evaluation.py - Evaluation runner → U-Net model: runs evaluation
semantic_segmentation/unet/evaluation.py - Evaluation runner → Segmentation metrics: calculates scores
semantic_segmentation/unet/evaluation.py - Evaluation runner → Saved predictions: writes masks
semantic_segmentation/unet/evaluation.py