EXTRACTING FOREST AGE IN A PACIFIC-NORTHWEST FOREST FROM THEMATIC MAPPER AND TOPOGRAPHIC DATA

The feasibility of extracting forest age of young stands (< 50 yr) in a Pacific Northwest Forest using Landsat Thematic Mapper (TM) spectral bands and topographic information was explored using a neural network approach. Understanding the changes of forest fragmentation through t ime are important for assessing alterations in ecosystem processes (fo rest productivity, species diversity, nutrient cycling, carbon flux, h ydrology, spread of pests, etc.) and wildlife habitat and populations. The study area teas the H.J. Andrews Experimental Forest on the Blue River Ranger District of the Willamette National Forest in western Ore gon. Timber harvesting has occurred in this forest over the past 45 ye ars and has a recorded forest management history. The study area was e xtracted from a georeferenced TM scene acquired on 7 July 1991. A coin cident digital terrain model (DTM) derived from digital topographic el evation data was also acquired. Using this DTM and an image processing software package, slope and aspect images were generated over the stu dy area. Sites were chosen to cover the entire range of forest stand a ge and slope and aspect. The oldest recorded clearcut stands were logg ed in 1950. A number of sites were chosen as primary forest which had no recorded history of cutting. Various feed-forward neural networks t rained with back propagation were tested to predict forest age from TM data and topographic data. The results demonstrated that neural netwo rks can be used as an initial model for inferring forest age. The best network was a 6-->5-->1 structure with inputs of TM Bands 3, 4, and 5 , elevation, slope and aspect. The rms values of the predicted forest age were on the order of 5 years. TM Bands 1, 2, 6, and 7 did not sign ificantly add information to the network for learning forest age. Furt hermore, the results suggest that topographic information (elevation, slope, and aspect) can be effectively utilized by a neural network app roach. The results of the network approach were significantly better t han corresponding linear systems.