Interest in CoastlineExtraction Project & Initial Understanding

[deepesh1singh]

Hi @fwitmer and @Ritika-K7,

I’ve been exploring the CoastlineExtraction repository and the overall pipeline, and I’m really interested in contributing to this project for GSoC 2026.

From going through the codebase and structure, here’s my understanding of the current workflow:

1. PlanetLabs imagery is first processed using NDWI-based water detection
2. A sliding window + adaptive thresholding approach is used to generate pseudo-labels (binary water/land masks)
3. These NDWI-generated masks act as training labels
4. A deep learning segmentation model (U-Net style) is then trained on this data
5. The final output is a segmented water mask that can be vectorized into coastline geometries

I find this hybrid approach (remote sensing + weak supervision + deep learning) very interesting, especially the idea of bootstrapping labels from NDWI and then improving them via learning.

From the listed improvement areas, I’d particularly like to work on:

1. Improving training data quality - Enhancing NDWI-based labeling (better thresholding, noise reduction, handling edge cases like shadows and turbid water)
2. UDM integration - Using PlanetLabs Usable Data Mask to filter out unreliable pixels and improve both training and inference robustness
3. Handling challenging regions - Especially cliffs, shadows, and built-up coastal regions where NDWI tends to fail
4. Data expansion - Extending beyond the current Deering (2017–2019) dataset to improve generalization across time

As a starting point, I’m planning to begin with the preprocessing pipeline and label generation (NDWI masks), so I can fully understand the data flow before moving into model-level improvements.

I had a couple of clarifications before I proceed:

1. Would it be a good direction to first focus on improving NDWI-based label quality and preprocessing before moving to model retraining?
2. For UDM integration, should the masking ideally be applied before NDWI computation, or is it expected to be incorporated after label generation / during training?
3. Are there any known failure cases (specific regions or conditions) that you would recommend prioritizing early on?

I’d really appreciate your guidance on how to best align my initial contributions with the project goals.

Thanks!