Mapping Ecological Land Systems and Classification Uncertainties from Digital Elevation and Forest-Cover Data Using Neural Networks

Document Type


Publication Date



Our approaches in this project emphasized mainly the technical aspects of the land-systems classification problem with neural networks. Using digital elevation, its derivatives, and forest cover data as input, we constructed neural networks to classify 27 land-system classes at Duck Mountain, Manitoba, Canada. Training and testing of those neural networks were done using an existing land-systems map prepared through airphoto interpretation and field studies. Two types of data structure were evaluated: polygon and raster forms. Both types of data sets contained the elevation, slope, aspect, dominant forest species and corresponding crown closures, and more general site information on cover type, subtype, site, cutting class, and crown closure. Because the data were obtained from different sources with different scales of measurement, we developed several methods to encode those data into suitable formats for use by the neural networks. With the polygon-based data set, a number of neural network structures and different data encoding methods were tested, and the best overall classification accuracy was only 26.8 percent in agreement with the existing map. The elevation and the forest-cover data were converted into a raster data set with 50-m by 50-m grid cell units. More experiments were done with this data set. Results indicate that a random sampling strategy for training sample selection led to better classification results than a contiguous sampling method. Approximately 10 percent of the total samples were sufficient for network training. The best overall classification accuracy was 52.0 percent when the neural network classification result was compared with the existing map. We developed a method to estimate classification uncertainties based on neural network outputs obtained from evezy mapping unit.

Was this content written or created while at USF?


Citation / Publisher Attribution

Photogrammetric Engineering & Remote Sensing, v. 62, issue 11, p. 1249-1260