NDWI Label Generation Workflow and Possible Improvement Directions

[AnvayKharb]

Hi @fwitmer and @Ritika-K7 ,

I’ve been studying the CoastlineExtraction project and would like to contribute. I’m currently an active contributor to the DREAMS
https://github.com/KathiraveluLab/DREAMS project and have been working on AI-related features there, so I’m comfortable working with the codebase structure and ML pipelines, and I’d like to extend that experience to this project.

After going through the README and pipeline, my understanding is:

• PlanetLabs imagery → NDWI + sliding window Otsu (ndwi_labels.py) → binary masks
• NDWI masks used as training labels
• U-Net model trained for water/land segmentation → coastline extraction

From the listed improvement areas, I’m particularly interested in working on:

• UDM integration to filter unusable pixels before training
• Improving NDWI label quality for better training masks
• Dataset expansion beyond the 2017–2019 Deering data

As an initial step, I’m planning to start with smaller contributions to better understand the preprocessing pipeline before moving to larger improvements like UDM integration.

Before starting, I wanted to confirm:

1. Does focusing first on NDWI label quality and preprocessing improvements align with the intended direction of the project?
2. For UDM integration, is the expected approach to filter unusable pixels before NDWI computation or after mask generation?

I’d appreciate any suggestions on where to start.

Thanks!

[AnvayKharb]

Hi @fwitmer @Ritika-K7 a reminder, incase you missed this.

[Ritika-K7]

Hi Anvay, Thank you for your interest in our project . Yes, focusing on improving the NDWI labels is a good starting point. The masks generated from NDWI are currently used as training labels.... image-mask pairs for the U-Net model. So improving their quality can directly improve the model performance. For extending the dataset.... we need to download additional imagery from PlanetLabs. You can create a student account there to access more data. For now, the raw_data folder in the repository contains some images that you can use to experiment and understand the preprocessing pipeline. Regarding UDM integration, the general idea would be to use the UDM files to filter out unusable pixels.. like clouds, shadows, etc. Ideally, this filtering should happen before generating the final training masks, so that noisy regions do not affect NDWI computation and label quality. But you can explore both approaches and see what works better within the current pipeline structure. Looking forward to your updates. regards, Ritika

…

On Sun, Mar 1, 2026 at 9:17 PM Anvay kharb ***@***.***> wrote: Hi @fwitmer <https://github.com/fwitmer> @Ritika-K7 <https://github.com/Ritika-K7> a reminder, incase you missed this. — Reply to this email directly, view it on GitHub <#84 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BD5MT6OC3SIOGQOTQ3CGRF34ORLRPAVCNFSM6AAAAACWCAHXB6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKOJWGI4DINA> . You are receiving this because you were mentioned.Message ID: ***@***.*** com>

[AnvayKharb]

Hi Frank and Ritika,

Thank you for the clarification.

I'll start by running the existing pipeline on the raw_data folder to get familiar with the preprocessing steps, then move toward NDWI visualization and UDM filtering experiments.

I'll post an update once I've successfully run the pipeline.

[AnvayKharb]

Hi @Ritika-K7 and @fwitmer I have openend a PR #86 please tell me am I going in the right direction ?

[AnvayKharb]

Hi @fwitmer and @Ritika-K7 ,

I have been studying the codebase and both previous GSoC reports in depth and have developed a proposed approach for GSoC 2026. Before finalizing my proposal I wanted to confirm my direction with you.

I am currently deciding between two possible focuses:

Option A Prioritize algorithmic accuracy improvements to the existing NDWI water detection and training pipeline, including data quality filtering using UDM2 masks, precision improvements, and a rigorous evaluation framework with independent test data.

Option B Prioritize dataset expansion and model generalization by focusing primarily on acquiring new imagery beyond 2017-2019, retraining the model on a larger and more diverse dataset, and testing performance across different years and conditions.

My instinct is that Option A lays a more solid foundation since the pipeline has never been rigorously validated, but I wanted to confirm whether you feel the dataset expansion in Option B is the more pressing need for the project right now.

Would either of these directions be more valuable to the project at this stage?