Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Phragmites Mapping topic
No spam. Unsubscribe anytime.
Refuge mapping compares manual and machine-learning approaches to track phragmites
Summary
Researchers compared manual photo-interpretation maps and random-forest models across Bear River Refuge units and found models perform well for dense Phragmites but less reliably for treated stands and small fragments; high-resolution imagery and robust training data are critical for scaling.
Forestry, Fire and State Lands staff and collaborators presented a mapping and modeling pilot for Phragmites on lower Bear River Refuge units (6–10) that compared manual photo-interpretation to automated random-forest models using multiple imagery sources.
Pete Goodwin described a multi-tiered approach using high-resolution WorldView imagery (collected 07/20/2023), Planet (≈3 m) and Sentinel (10 m). The team supplied training data from manual maps, converted to pixel-level samples, and used Google Earth Engine to develop and test models. Goodwin said dense Phragmites stands were identified with high user accuracy, with producer accuracy lower (about 83% in some tests), while treated Phragmites and small linear fragments were more likely to be missed at coarser resolutions.
“Models generally map dense phragmites with about 80% accuracy,” Goodwin said, noting random-forest approaches tended to outperform manual mapping on some metrics and were substantially faster to produce. He cautioned that models misclassify in certain conditions—algae, shallow flooding and treated areas produce spectral signatures that confuse classifiers—and that training data selection affects outcomes: training on very obvious patches can overstate modeled performance on mixed-field validation points.
The presenters recommended mixing high-resolution imagery and field-collected training/validation points and exploring object-based classification (structure plus spectral bands) from drone or manned low-altitude flights to resolve mixed vegetation classes such as cattails and bulrush. The team also proposed using periodic, watershed-scale mapping to track new infestations and to evaluate treatment effectiveness over time.
Participants discussed practical next steps: targeted drone surveys for problem areas, using covariance tower ET measurements to quantify phragmites water use before and after treatment, and scaling the workflow to the wider watershed while acknowledging budget and logistics constraints. Goodwin thanked partners including USFWS and noted that robust training data remain the single most important factor for reliable automated mapping.
