[Proposal Discussion] Enhancing Coastline Extraction using 4-Band U-Net + EfficientNet-B0 with Geometric Refinement

[Devashish-Mishra-2003]

Hi mentors,

I’ve been exploring the coastline extraction pipeline and looking through ndwi_labels.py as well as the discussions about “false water” detections caused by terrain shadows.

While experimenting with Planet 4-band imagery (RGB + NIR), I tried a segmentation-based approach using a U-Net with an EfficientNet-B0 encoder and a small spectral stem that directly ingests the NIR band. In preliminary experiments this helped reduce shadow misclassification in cliff-heavy regions.

I also noticed that even when the segmentation mask is accurate, the extracted coastline can look quite “jagged” because of the pixel grid (especially with 256×256 tiles). This made me wonder whether a geometric refinement stage might be useful.

One idea I tested was applying Active Contours (Snakes) followed by Douglas-Peucker simplification to smooth the predicted coastline before exporting GIS vectors.

A few questions I had for the project:

• Are the current shoreline labels generated from vectors or manually annotated masks?
• Has a segmentation-based approach been explored before in the project?
• Would improving geometric smoothness of the final coastline vectors be valuable for downstream erosion modeling?

I’ve been experimenting with a prototype pipeline and documented the experiments in a Kaggle notebook:

https://www.kaggle.com/code/devashishmishra2003/coastline-notebook

I’d really appreciate any feedback or pointers on whether this direction aligns with the project goals.

Best regards,
Devashish

[Ritika-K7]

hi Devashish,
Thanks for exploring the project and showing interest.
Hopefully these answers will help clear your doubts................

1.The current labels are mostly created from vectors, not manually drawn masks. The pipeline uses coastline shapefiles and NDWI (with sliding windows and Otsu thresholding) to generate masks automatically.
These masks are then converted back into coastline vectors. There are some ground-truth files available -ground_truth\2016_HiRes_Final_Coastline.shp

2.Segmentation is still in the exploration phase. However, you can find a U-Net based training and prediction pipeline in the repository. The project direction also mentions using models like DeepWaterMap with Planet imagery.

3.Yes, it would likely help. Even if the segmentation is good, the coastline can look jagged due to pixel resolution. This can affect measurements like shoreline position and erosion distance. Adding a smoothing step .....like active contours or simplification can improve results, as long as we make sure it doesn’t shift the coastline too much and affect accuracy.

regards,
Ritika