GSoC 2026 Final Report

[het-shah04]

GSoC 2026 Final Report

Contributor: Het Shah (@het-shah04)
Organization: Alaska
Mentors: @fwitmer @Ritika-K7
Project: Automated Coastline Extraction for Erosion Modeling in Alaska

Overview

Over the course of GSoC 2026, I built a complete pipeline for automated coastline extraction from Planet Labs satellite imagery. The work spans various phases: Planet Labs data collection, data quality assurance, training dataset construction, deep learning model training and evaluation. Together, these contributions provide the project with a robust, end-to-end workflow from raw satellite rasters to quantitatively validated coastline vectors.

Contributions

1. Dataset extension
   Existing dataset consisted Planet Labs imagery upto 2020. I added Planet Labs imagery for five years 2021-2025. This increased the dataset size significantly, thus improving the model training data.
2. Hybrid Cloud & Haze Filtering Quality Assurance Pipeline
   The first challenge in working with Arctic satellite imagery is cloud contamination. Raw Planet Labs tiles regularly contain clouds, shadows, haze, and snow — all of which corrupt downstream NDWI-based label generation.
   I implemented a hybrid dual-pass quality assurance (QA) pipeline in data_preprocessing/ that:

• Combines the 8-band Planet UDM2 quality data (Snow, Shadow, Haze, and Cloud layers) with a dynamic Otsu brightness thresholding pass to catch fine haze and partially cloudy tiles that UDM2 alone misses.
• Enforces a 15% cloud cover budget — tiles exceeding this threshold are rejected from training data.
• Detects custom shadow and haze patterns using adaptive pixel-level thresholds derived directly from the scene statistics.

3. End-to-end training pipeline
   Created a new Attention U-net architecture in addition to the existing vanilla U-net architecture for coastline extraction. Along with the U-net variants, I also finetuned SAM on our task for coastline extraction.
4. Created visualizations for analyzing the model performance on coastline prediction.

Pull Requests

• #108 — Add hybrid cloud filtering quality assurance pipeline
• #109 — Add multi-year coastal analysis, Configuration 7 QC, and sample/PDF visualization tools
• #111 — Adding end-to-end training and evaluation code
• #112 — Adding a script for model inference
• #113 — Refactoring training pipeline
• #114 — Adding September 9th Planet Labs reference coastline to ground_truth
• #116 — Adding code for SAM fine-tuning and inference
• #118 - Add GSoC 2026 final report