Aryan Prakhar - Interest, Ideas, and Project Discussion

[AryanPrakhar]

Hey @fwitmer, @rawann31 and @Ritika-K7,
I was exploring UAA GSoC projects and this one immediately caught my interest. I especially liked it because beyond being an engineering problem, it's a research-heavy problem, which I dearly enjoy. Also, there is limited literature out there on this very specific problem linked to Alaska, adding to the thrill.

Quick intro: My name is Aryan Prakhar, an undergrad at the Indian Institute of Technology (BHU) Varanasi. I have co-authored allenai/DiscoveryBench at ICLR 2025, Singapore, in data driven discovery domain. I have also contributed to ICML 2024-Vienna paper DataVoyager, which was again on data-driven scientific discovery. I have also contributed to open source earlier via the Code For GovTech DMP'24 program. I have been very interested in computer vision, and have made several projects, one of which was an amateur solution for a ML4Sci problem statement last year. I couldn’t apply for GSoC then due to summer internship commitments as an AI Engineer Intern.

I have already set up the codebase on my system, and exploring different ideas for improving segmentation in steep or cliff-like coastal areas and handling water shadows. Since the contributors in last years have already spent a lot of time on this problem, I have also been going through Prof. Arun K. Saraf’s Remote Sensing and Digital Image Processing of Satellite Data - IIT Roorkee lecture series to build a stronger fundamentals-first understanding, to meaningfully further their work. I am currently experimenting with a few ideas and will keep sharing updates in this thread.

[AryanPrakhar]

Hey @fwitmer @Ritika-K7 @rawann31 - quick conceptual question. How is coastline being defined in this project?

While going through the pipeline, I noticed we eventually enforce a binary land vs water boundary. Since coastlines can act as transition zones and shift due to tides, waves, wet sand, and shallow intertidal areas, I noticed the decision-making seems locally adaptive, but the final output is still a crisp boundary. Modelling a fuzzy coastline would likely require observations across multiple tidal conditions (for example, captures at highest and lowest tide).

I was curious what assumptions we’re choosing to follow when enforcing this hard boundary. If this approach has already been validated or is working well in practice, I’m completely fine proceeding with it and focusing on other parts of the project.

Cheers!

ps: now I am trying to incorporate DEM

[rawann31]

Hello @AryanPrakhar, thanks for your interest. You are right that coastlines are inherently dynamic and can vary due to tides, waves, and intertidal processes. However, given the data currently available to us, the coastline in this project is defined as a crisp binary boundary separating land and water.

At present, we have three reference coastlines, corresponding to the years 2005, 2016, and 2019. These coastlines represent a fixed land–water separation and do not explicitly account for tidal or wave effects, mainly due to the lack of ground-truth data captured under different tidal conditions (They need manual annotation in QIS apps and we want to find a solution for this or search for data).

One of our key planned enhancements is to expand the temporal coverage of ground-truth coastlines. High-resolution PlanetScope imagery that we use in training is available for many additional years, we currently lack reliable coastline ground truth for those periods. Identifying or generating accurate coastline references for all available Planet imagery would allow us to better study temporal variability.

Waiting for your updates !

[AryanPrakhar]

Thanks for the insights @rawann31. This makes things a lot more clearer for me. And yeah, data, and specifically temporal data will remain a bottleneck, so tacking the problem incrementally will be quite a nice strategy.

Among the features to implement this year, I found that cliff segmentation has also been listed. Since it could be a really rich feature to signify where the water is present and where it is not, I tried incorporating it in my pipeline. My hypothesis is that this should almost solve the coastline extraction problem where the water starts just after a highland/cliff. I planned to use both Elevation and Slope as features.

I tried implementing it using 2015 DTM collected by USACE for Deering. I found out that there is very limited coverage of the terrain in the Deering Airport region. Specifically, there is less than 1% coverage for the Deering airport region. Could you please direct me to a richer DEM/ DTM survey in your knowledge?

Attaching the result of elevation and slope mapping of this region.

[AryanPrakhar]

Also the RGB composite makes the damage done by the shadows really clear. Slope calculation using DTM will almost solve that problem too, since a high slope is almost always correlated with shadows, given we are able to find sufficient data.

[rawann31]

@AryanPrakhar Yes, cliff areas have the least accuracy. When we calculate RMSE for the generated coastlines and the ground truth, we divide the coastline into regions to get a better understanding of which areas need more improvement. The cliff areas show the lowest accuracy.

Regarding the DEM, can you please provide the source from which you obtained it? Maybe you filtered by year or selected a specific region, so you only got a small part.

I found these DEM sources:

• https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_DEM_GLO30
• https://developers.google.com/earth-engine/datasets/catalog/JAXA_ALOS_AW3D30_V4_1

I believe you can find more if you search in the Google Earth Engine datasets. These are global datasets with a resolution of 30 m, so when you merge them with your dataframe, make sure to downscale appropriately. I am not sure if there are other free datasets that provide elevation.

Does this answer your question?

[AryanPrakhar]

Yeah got it @rawann31.
I obtained the dataset from this source: https://elevation.alaska.gov/#66.08657:-162.74240:12 . Specifically, for the above visualization, I used the Deering_2015 survey.
I will explore the sources that you highlighted, and keep the scaling factor in mind.

Thank you for your prompt response :)

[rawann31]

Your welcome and thanks for your interest , if you have updates share with us !