Download (PDF 429 KB): https://research.fs.usda.gov/download/treesearch/4534.pdf
Abstract
Two methods of training data collection for automated image classification were tested in Virginia as part of a larger effort to develop an objective, repeatable, and low-cost method to provide forest area classification from satellite imagery. The derived forest area estimates were compared to estimates derived from a traditional photo-interpreted, double sample. One method used maplets digitized from ancillary imagery. Seed pixels, the other approach, used only available ground plot data and the image to be classified. Both methods of training data collection resulted in classification accuracy approaching 89 percent, and area estimation precision surpassing the FIA standard of 3 percent per million acres of timberland. However, the precision estimate was met in large part from the additional ground truth data collected supplemental to the national standard sample frame of one plot per 6,000 ac. The seed pixel approach is recommended over maplets, because it does not require ancillary imagery and is less costly in analyst time.
Citation
Scrivani, John A.; Wynne, Randolph H.; Blinn, Christine E.; Musy, Rebecca F. 2001. Phase I Forest Area Estimation Using Landsat TM and Iterative Guided Spectral Class Rejection: Assessment of Possible Training Data Protocols. In: Reams, Gregory A.; McRoberts, Ronald E.; Van Deusen, Paul C., eds. 2001. Proceedings of the second annual Forest Inventory and Analysis symposium; 2000 October 17-18; Salt Lake City, UT. Gen. Tech. Rep. SRS-47. Asheville, NC: U.S. Department of Agriculture, Forest Service, Southern Research Station. pp. 11-14
