Wilson Cheung - Introduction and (hopefully) Ongoing Project Discussion

[wirrywoo]

Hello @fwitmer and @Ritika-K7,

My name is Wilson and I am interested in being involved with the Automated coastline extraction for erosion modeling in Alaska project for Google Summer of Code this year. I have previously built and led the developments of prototypes to process and transcode various geospatial data sources to improve AI readiness and enable downstream machine learning model developments, resulting in creating new contracting opportunities for an employer I previously worked with. Although I am technically considered a working professional on paper, I wanted to take part in GSoC this year to dive more deeply into ways to contribute towards open source development, with an interest to explore ways to combine my skills in data engineering and machine learning to the geospatial sciences.

Having gone through the accepted orgs list and all of the respective project ideas for GSoC this year, it's inspiring to see how the University of Alaska continues to promote open source and fosters an active community supporting the unique challenges faced in understanding climate change. Both the community and the projects drew me into wanting to leverage my skill sets to be part of this active community and contribute to developing solutions together. I'm excited to spend the next month with this organization trying to explore if there is a potential fit and to see where I can help contribute the most!

Before officially submitting a proposal, I would like to get a better understanding on the overall scope of this project.

• Under Expected Outcomes, I noticed that it states that after successful data ingestion of the new source and model retraining, "the model can be applied to large amounts of imagery to model coastline changes over time." Do we have more insight on some of these applications?
• Has there been any painpoints using the "auto-download script is available to retrieve PlanetLabs imagery" in a time efficient manner?
• Are there other data sources to consider to further improve the performance of the model?

In the meantime, I will:

• Review development progress made in the past GSoC results to better understand context of the overall project.
• Attempt to run the download scripts and replicate the ability to retrain the segmentation model locally
• Research papers that utilizes different segmentation techniques to address varying cases as mentioned in the project description.

I will post my findings in this discussion thread as I'd like this to be a continuing conversation between us. Thank you! :)

[wirrywoo]

Hello @fwitmer and @Ritika-K7,

I wanted to check whether there are possible ways to further reduce download time using concurrency, in preparation for one of this year’s GSoC action items; namely, to extend the temporal range of the dataset for possible accuracy gains in the machine learning model.

So far, I have successfully:

• Set up my codebase and installed the relevant Python dependencies,
• Created a Planet API account, and
• Added my API key to the DeeringAutoDownloadCode.py script to run benchmark tests:

python3 DeeringAutoDownloadCode.py

However, I’ve encountered limitations with my trial account within my organization. Some sample error messages include:

{'field': None, 'general': [{'message': "Unable to accept order: Your organization does not have permission to run the 'clip' tool"}]}

no access to assets: PSScene/20230809_212804_22_2451/[ortho_analytic_4b_sr ortho_analytic_4b_xml ortho_udm2]

Would it be possible for me to become a tentative member of the organization, or are there additional instructions I can follow to enable the necessary features in my trial organization? This would allow me to investigate further and continue testing the codebase efficiently. I understand there may be possible cost implications adding new members in the org, so I wanted to ask to be sure.

Thanks,
Wilson

[wirrywoo]

Looping in @Ritika-K7 for further discussion. Thanks!

[Ritika-K7]

Hi Wilson,

Thanks for your detailed message and for testing the pipeline so thoroughly.

Project Scope
Currently, the project is focused on erosion monitoring for the Deering region of Alaska. The main goal is to model coastline changes over time using PlanetLabs imagery.

Download Script & API Access
For downloading additional imagery (especially after 2018), I would suggest creating a PlanetLabs account under the student quota if possible. That should allow you to access more data for experimentation.
At the moment, optimizing the auto-download script is not our main priority, so I would recommend focusing more on how we can do the segmentation better and retraining the model.

Other Data Sources
For now, it is advisable to stick to PlanetLabs imagery. While combining other sources like Landsat or Sentinel may sound exciting(i also tried doing it ) but it will only introduces additional dependencies and complications.
PlanetLabs images are captured at specific time intervals, and matching them exactly with other satellite imagery can be difficult. Even if taken on the same day, weather conditions -like cloud cover etc ....may differ, which can introduce inconsistencies in training data.

Also, you may have noticed that there are sample images are present in the raw data folder . you can use them for experimenting purpose for now.

Looking forward to your updates.

regards,
Ritika