Feedback on my work done so far.

[Shashank-248]

Progress Update: Automated Coastline Extraction for Erosion Modeling in Alaska

I have been actively contributing to the Automated Coastline Extraction for Erosion Modeling in Alaska project and wanted to share my progress so far. My focus has been on improving the accuracy of coastline extraction using high-resolution PlanetLabs and Earth Explorer imagery. Below is a summary of my work, challenges faced, and planned next steps.

I would appreciate feedback on whether my approach aligns with the project’s goals and how I can further refine my contributions.

---

1. Work Done So Far

NDWI & MNDWI Calculation

• Implemented Otsu’s method for adaptive thresholding
• Combined NDWI and MNDWI to improve water detection
• Applied sliding window techniques for localized thresholding

Image: NDWI vs. MNDWI Comparison

Water Mask Refinement

• Integrated SWIR data for improved segmentation
• Generated pseudo-DEMs using SWIR bands
• Applied Gaussian Blur and Canny Edge Detection to refine coastlines

Image: Water Mask Refinement (Before & After)

Visualization & Validation

• Created NDWI, MNDWI, and water mask visualizations for validation
• Compared with high-resolution imagery for qualitative analysis

Image: Extracted Coastline Result

Handling Challenging Conditions

• Improved segmentation in shadowed, urban, and artifact-heavy regions
• Used histogram matching and adaptive thresholding to address data quality issues

---

2. Research & Problem Understanding

Project Analysis & Literature Review

Studied the existing NDWI-based segmentation pipeline and explored ways to improve its accuracy. Key research areas included:

• Adaptive Thresholding: Otsu’s method, sliding windows
• Multi-Index Approaches: NDWI and MNDWI fusion
• Machine Learning Techniques: Supervised learning (SVM, RF) and adversarial training (GANs)
• SWIR & DEM Usage: Elevation-based water mask refinement

---

3. Future Plans

Data Expansion

• Extend dataset to 2020–2023 and beyond
• Apply the pipeline to other Alaskan communities

Improved Cliff Area Segmentation

• Use DEM and edge detection to refine segmentation
• Validate with high-resolution imagery

Machine Learning Approaches

• Train SVM/Random Forest using combined spectral and spatial features
• Develop GAN-based adversarial training for coastline refinement

Integration with Existing Pipeline

• Merge my work into the main project pipeline
• Optimize for efficiency and scalability

---

4. Feedback Needed

• Does my segmentation approach align with the project’s objectives?
• Should my method be a pre-processing step or a separate module?
• What additional improvements should I prioritize?

@rawann31 , @fwitmer ,Looking forward to your thoughts!

[rawann31]

Hi @Shashank-248

Thanks for your time and effort! I have a couple of questions regarding your approach:

• What satellite data are you using for generating pseudo-DEMs from SWIR bands? How do you derive elevation information from SWIR data?
• Have you tested whether your approach works effectively with Planet imagery (specially SWIR data integration) ? Since Deering, Alaska, is our primary area of interest, have you tried applying your method there?

[Shashank-248]

Hi @rawann31,

I have used Landsat 8 satellite data for generating pseudo-DEMs from SWIR bands. Specifically, I use the following bands:

SWIR1 (Band 6): 1.57–1.65 µm

SWIR2 (Band 7): 2.11–2.29 µm

How I Derived Elevation Information from SWIR Data
Pseudo-DEM Generation:
I calculate the ratio of SWIR2 to SWIR1 to approximate elevation information. This ratio is based on the principle that different land surfaces (e.g., water, vegetation, soil) reflect SWIR bands differently, and this can be used to infer relative elevation.

The formula for pseudo-DEM generation is:

Pseudo-DEM = SWIR2 / ( SWIR1 + ϵ )
where ϵ is a small constant to avoid division by zero.

Normalization:
I normalize the pseudo-DEM values to a range of 0 to 1 for easier integration with other data:

Normalized Pseudo-DEM = Pseudo-DEM − min(Pseudo-DEM) / (max(Pseudo-DEM) − min(Pseudo-DEM))

Integration with Water Mask:
I use the pseudo-DEM to refine the water mask by excluding high-elevation areas (e.g., cliffs, mountains) that are unlikely to be water bodies.

Testing with Planet Imagery:
Yes, I have tested my approach with Planet imagery ( just a single imagery, as I have mostly worked with USGS-provided imagery), and it works effectively for water body detection and coastline extraction. Here’s how I integrated SWIR data:

SWIR Data Integration: I used the SWIR band from USGS imagery to generate pseudo-DEMs and refine water masks.

Method: I applied the same pseudo-DEM generation method (SWIR2/SWIR1 ratio) to USGS imagery.

Results: The high resolution of Planet imagery allowed for more detailed elevation information, which improved segmentation accuracy in complex terrains (e.g., cliffs, shadows). The integration of SWIR data improved water detection accuracy, especially in areas with mixed pixels and shadows. The refined water masks were used as input for coastline extraction, resulting in more accurate and reliable results.

Application to Deering, Alaska
Yes, I have applied my method to Deering, Alaska(just once as I am facing difficulty with this section), just as mentioned in the images attached. However, I have applied my method to other complex parts of Alaska, even more extensively than Deering. However, I have successfully applied the method to Deering as well.

Difficulty Accessing Planet Imagery
I am currently facing difficulty in accessing Planet imagery. For now, I have been using USGS-provided satellite imagery (which is similar to Planet imagery, but unable to get Deering, Alaska imagery from it). Could you kindly guide me on how to access Planet imagery more effectively? Your assistance would be greatly appreciated!

[Shashank-248]

Hi @rawann31 ,
I would like to know your thoughts on my approach and the way I researched and thought. I have seen your previous year GSOC works and really would like to hear it out from you. I have made sure to keep my approach as adaptable to your work and the current status of the project. Not only are you a mentor , but also a previous contributor to the project so your experience and suggestions are really valuable to me. I am deeply interested into this project and would like your mentorship to really make my contributions meaningful.

Hoping to hear from you soon ! 😃
Thank you!!

[rawann31]

Hello @Shashank-248

I am really sorry for replying late and thank you so much for your effort; I really appreciate it.

First, I will explain how to access Planet imagery. You can create an account using your educational email, which will grant you a sufficient quota for a long time. Visit this link to find details on how to get free access. Click "Apply," (under Free or Basic section) fill in your details, and they will send you an email to activate your account within a week.

Once your account is activated, you have two options for downloading imagery:

1. Using the Planet platform – After activation, you will find maps and details directly on their platform.
2. Using their API – Planet provides a repository with detailed documentation on how to use their API for downloading imagery.

Additionally, in my previous work, I created a Python script that explains step-by-step how to download images using the API. You can find it here: DeeringAutoDownloadCode.py. This script explains everything in detail, including which notebook contains my code.

If you have questions related to this send me any time. This part is easy but will take you some time to be familiar with

[rawann31]

@Shashank-248

I am confused about something. Do you integrate Planet data with elevation data (derived from Landsat) from Landsat, or do you integrate elevation data (derived from Landsat) with Landsat bands? Please correct me if I have misunderstood.

[Shashank-248]

Hi @rawann31,

To clarify, I integrate elevation data derived from Landsat SWIR bands (specifically, the ratio of SWIR2 to SWIR1) with Landsat bands (e.g., Green, NIR, SWIR) to refine water masks. I do not directly integrate Planet data with elevation data from Landsat. My primary focus has been on using Landsat for both elevation approximation and water mask refinement. (I am planning to improve this further).

Actually , I have used USGS imagery majorly, I did not work with Planet labs imagery , sorry for my miscommunication, but thanks for clarifying it. I will now use Planet imagery. Thankyou for taking time and explaining it to me. 😀

Correct me if I have gone wrong with my approach.

[rawann31]

Yes, our focus in GSoC2025 will be using planet data.

[Shashank-248]

@rawann31 , correct my approach if it's wrong.
I have been working with Planet data 3m resolution along with Sentinel-2's 20m resolution SWIR data for integrating with SWIR bands , this approach would enable me to use Modified NDWI (MNDWI) which would result us in better segmentation accuracy. Thus directly improving the project and making sure of the project goals.

[Shashank-248]

@rawann31 ,

I have been looking actively into the codebase to find any bugs or issues to tackle with and understand the code better , I have seen that in rastertools.py script converts radiance , classifies and contour extracts, also calculating NDWI . But it may struggle with large datasets ( satellite imagery) and also lacks resampling and reprojection . I would like to improve this by adding chunk based processing , resampling , reprojection and logging . This is to handle large datasets efficiently , and better debug and track.

What do you think about this?

[rawann31]

Hi @Shashank-248,
I am interested to see your improvements. Could you share more details on how chunk-based processing, resampling, and reprojection would address the limitations? For example, have you encountered specific cases where the current implementation struggles with large satellite imagery (please share this with us) ?

[Shashank-248]

Hi @rawann31,

• Chunk based processing makes sure the entire imagery is made into smaller tiles(e.g., 1024*1024 pixels) . This tiles are processed independently reducing the memory usage and would eliminate out-of-memory crashes ( I have experienced this on my machine , thus leading to me going about this way) , even if it doesn't crash it significantly slows down the processing.

• Dynamic Resampling eases out multi-sensor analysis ( I have done this as trying to integrate SWIR data with PlanetLabs data), It reduces any inconvenience and improves alignment accuracy ( This is much needed further down while integrating SWIR data from Sentinel-2, which will increase the accuracy of the project and would really be an improvement over last year).

• Reprojection ensures accurate geospatial alignment and automates workflows.

I have faced issues such as OOM crashes , alignment accuracy issues while integrating SWIR data . I have been looking forward to improving the project and increasing the accuracy in segmentation from the start , thus faced these related issues.

Additionally, these changes would directly support the project main goal of improving DeepWaterMap's accuracy , as these changes would help in adapting the algorithm to work with high resolution data(PlanetLab's 3m resolution and Sentinel-2's 20m resolution) unlike original Landsat's 30m resolution data.

What do you think of these improvements and issues also about the direction of my contributions regarding the project improvements. If the improvements look good, then I would like to send a PR.

I would like to know your thoughts on these.

[Shashank-248]

hey @rawann31 ,

Can I send my proposal to your email once? I would like you to once check it so that I could submit it

[rawann31]

@Shashank-248 Yes, sure

[Shashank-248]

hi @rawann31 , I have sent the proposal to your mail . Do check it out

[Shashank-248]

Hey @fwitmer , @pradeeban

Wanted to ask you if could you review my proposal through email. The deadline for GSOC contributor proposal application is nearing deadline and want to submit as soon as possible, also seek feedback from mentors . Could you check it out?

@pradeeban , Nice to talk to you again . I have joined discussion in another project(Creating shareable "albums" from locally stored DICOM images) but soon found this project to suit my interests and strengths.

Could you review my proposal if I could mail it to you, showcasing my understanding of the project's bigger picture and my work done towards it.

Thanks, hoping to collaborate with you soon.

[rawann31]

Hi @Shashank-248

I am sorry. I will check soon.

[Shashank-248]

Hey @rawann31 ,

No problem . Just wanted to interact and share my proposal to both @fwitmer and @pradeeban.

[pradeeban]

You should submit your proposal on the GSoC Portal and then you can continue to edit and re-upload as needed until the deadline.

[Shashank-248]

Yeah @pradeeban ,

I have submitted my proposal on GSoC portal , if possible could you check the proposal , sighting any improvements ?

[Shashank-248]

Hey @rawann31 ,

Could you check the proposal I have sent

Thank you

[rawann31]

Sure

[Shashank-248]

Hey @rawann31 , @fwitmer

I have come across other indexes- AWEI and WNDWI . I did not understand it clearly , could you explain me about it and about any improvements it might bring to our project over proposed MNDWI index. Does it really bring a significant improvement over MNDWI ( Which I have planned to implement ) , over place such as Deering , Alaska (Our primary area of interest).

Thank you