DeepForest P_2: Recovering Historical Image Data Using Automated OrthoRegistration and ImageMatching

The U.S. Forest Service's Aerial Detection Survey (ADS) program has produced tens of thousands of hand-drawn polygons marking forest health damage from aircraft. Due to GPS imprecision and flight geometry, these polygons have systematic spatial offsets of 50–500 meters from their true locations in NAIP satellite imagery. This project builds an automated pipeline to align ADS polygons with NAIP data, recovering historical annotations as a training dataset for forest health computer vision models. The GSoC work focuses on three objectives: (1) improve alignment accuracy through a learned refinement layer using self-supervised displacement estimation techniques from the mentor-provided references, (2) scale the pipeline to all ~48,000 Oregon ADS polygons to generate a weak training dataset, and (3) fine-tune a DeepForest model on the aligned annotations for NAIP-based forest health outbreak detection. Deliverables: an optimized alignment pipeline with quantitative evaluation, aligned polygon GeoPackage for Oregon ADS data, a computer vision model for forest health detection on NAIP imagery, and a blog post documenting the methodology for others to apply to their own datasets.