Automated Coastline Extraction for Erosion Modeling in Alaska

Alaska's Arctic coastline is eroding rapidly, and tracking these changes at scale requires automated coastline extraction from satellite imagery. Manual digitization is too slow, and the existing pipeline has several gaps: training data includes cloud-affected and artifact-contaminated pixels, the dataset covers only 2017-2019, the model underperforms near coastal bluffs and in shadowed regions, and the pipeline outputs binary raster masks rather than the GIS-ready vectorized coastlines needed for erosion modeling. To address this, I have already studied the codebase end-to-end, fixed the bugs that were preventing the pipeline from running, and replaced the baseline U-Net with an Attention ResUNet - a pretrained ResNet-34 encoder with attention-gated skip connections and a combined BCE + Dice loss - achieving a Dice score of 0.941 and IoU of 0.921 on the current dataset. During GSoC, I will close the remaining gaps: integrate PlanetLabs Usable Data Masks to exclude bad pixels from training; expand the dataset to cover 2020 onward; improve segmentation in cliff and shadow regions; investigate SWIR and elevation data as additional inputs; and produce vectorized coastline outputs with per-transect erosion rates. The result will be a pipeline that runs from raw PlanetLabs imagery to quantified shoreline change, ready to apply across Alaska.