In cooperation with the USDA Forest Service Southern Research Station, an algorithm has been developed to replace the current aerial-photography-derived FIA Phase 1 estimates of forest/non-forest with a Landsat Thematic Mapper-based forest area estimation. Corrected area estimates were obtained using a new hybrid classifier called Iterative Guided Spectral Class Rejection (IGSCR) for portions of three physiographic regions of Virginia. Corrected area estimates were also derived using the Landsat Thematic Mapper-based Multi-Resolution Land Characteristic Interagency Consortium (MRLC) cover maps. Both satellite-based corrected area estimates were tested against the traditional photo-based estimates. Forest area estimates were not significantly different (at the 95% level) between the traditional FIA, IGSCR, and MRLC methods, although the precision of the satellite-based estimates was lower. The estimated percent forest area and the standard error (respectively) of the estimates for each region and method are as follows; Coastal Plain- Phase 1 66.06% and 1.08%, IGSCR 68.88% and 2.93%, MRLC 69.84% and 3.08%. Piedmont- Phase 1 63.87% and 1.91%, IGSCR 65.52% and 3.50%, MRLC 59.19% and 3.83%. Ridge and Valley- Phase 1 69.74% and 1.22%, IGSCR 70.02%, and 2.43%, MRLC 70.53% and 2.52%. Map accuracies were not significantly different (at the 95% level) between the IGSCR method and the MRLC method. Overall accuracies ranged from 80% to 89% using FIA definitions of forest and non-forest land use. Given standardization of the image rectification process and training data properties, the IGSCR methodology is objective and repeatable across users, regions, and time and outperforms the MRLC for FIA applications. / Master of Science

The goal of this project was to develop an operational Landsat TM image classification protocol for FIA forest area estimation. A hybrid classifier known as Iterative Guided Spectral Class Rejection (IGSCR) was automated using the ERDAS C Toolkit and ERDAS Macro Language. The resulting program was tested on 4 Landsat ETM+ images using training data collected via region-growing at 200 random points within each image. The classified images were spatially post-processed using variations on a 3x3 majority filter and a clump and eliminate technique. The accuracy of the images was assessed using the center land use of all plots, and subsets containing plots with 50, 75 and 100% homogeneity. The overall classification accuracies ranged from 81.9-95.4%. The forest area estimates derived from all image, filter and accuracy set combinations met the USDA Forest Service precision requirement of less than 3% per million acres timberland. There were no consistently significant filtering effects at the 95% level; however, the 3x3 majority filter significantly improved the accuracy of the most fragmented image and did not decrease the accuracy of the other images. Overall accuracy increased with homogeneity of the plots used in the validation set and decreased with fragmentation (estimated by % edge; R2 = 0.932). We conclude that the use of random points to initiate training data collection via region-growing may be an acceptable and repeatable addition to the IGSCR protocol, if the training data are representative of the spectral characteristics of the image. We recommend 3x3 majority filtering for all images, and, if it would not bias the sample, the selection of validation data using a plot homogeneity requirement rather than plot center land use only. These protocol refinements, along with the automation of IGSCR, make IGSCR suitable for use by the USDA Forest Service in the operational classification of Landsat imagery for forest area estimation. / Master of Science

The impacts of training data sample size and sampling method on the accuracy of forest/nonforest classifications of three mosaicked Landsat ETM+ images with the nearest neighbor decision rule were explored. Large training data pools of single pixels were used in simulations to create samples with three sampling methods (random, stratified random, and systematic) and eight sample sizes (25, 50, 75, 100, 200, 300, 400, and 500). Two forest area estimation techniques were used to estimate the proportion of forest in each image and to calculate forest area precision estimates. Training data editing was explored to remove problem pixels from the training data pools. All possible band combinations of the six non-thermal ETM+ bands were evaluated for every sample draw. Comparisons were made between classification accuracies to determine if all six bands were needed. The utility of separability indices, minimum and average Euclidian distances, and cross-validation accuracies for the selection of band combinations, prediction of classification accuracies, and assessment of sample quality were determined. Larger training data sample sizes produced classifications with higher average accuracies and lower variability. All three sampling methods had similar performance. Training data editing improved the average classification accuracies by a minimum of 5.45%, 5.31%, and 3.47%, respectively, for the three images. Band combinations with fewer than all six bands almost always produced the maximum classification accuracy for a single sample draw. The number of bands and combination of bands, which maximized classification accuracy, was dependent on the characteristics of the individual training data sample draw, the image, sample size, and, to a lesser extent, the sampling method. All three band selection measures were unable to select band combinations that produced higher accuracies on average than all six bands. Cross-validation accuracies with sample size 500 had high correlations with classification accuracies, and provided an indication of sample quality. Collection of a high quality training data sample is key to the performance of the nearest neighbor classifier. Larger samples are necessary to guarantee classifier performance and the utility of cross-validation accuracies. Further research is needed to identify the characteristics of "good" training data samples. / Ph. D.