Author(s): Fei Zhang, Rochester Institute of Technology (First Author, Presenting Author) Robert Chancia, Rochester Institute of Technology Amirhossein Hassanzadeh, Rochester Institute of Technology Jan van Aardt, Rochester Institute of Technology
LiDAR scans are used to map forests but models trained in open source urban datasets often fail under complex canopies. Annotating new forest data is costly, and fine-tuning big models can exceed typical computing power. We propose a simple method that combines city and forest scans, includes a short phase to learn from unlabeled forest data, and adapts the model in just a few steps—without needing new labels. We expect to improve forest mapping accuracy by more than 15% and cut computing needs by ~30%, making large-scale ecological surveys faster and cheaper.
