| This thesis presents a semi-automated method of tree stand classification implementing data mining models to classify tree stands as PFTs and SFUs. The study area consists of eleven FMUs in the province of Ontario, inventoried by the OMNR as of 2009, and provided in ESRI compatible shapefiles. Models were built by machine learning algorithms using variables derived from a climate model, DTM, spatial coordinates, stand densities, and tree heights. Machine learning algorithms include C4.5 Decision Trees with and without bagging, M5 model trees, Random Forests, Multilayer Perceptron ANNs, and both supervised and unsupervised Self Organizing Maps. WEKA 3.6.5 was used to create all classifier models, except for SOMs which were created using R 2.04, and run on the SHARCNET Brown cluster. Models created by the algorithms were validated using ten-fold cross validation (except for SOM models), and metrics on model accuracy and class true positive rates were calculated. Three models were created to explore different aspects of classification. Model 1 classified the six most dominant PFTs across all 11 FMUs. Model 2 attempted to assess the affect additional terrain data derived from the DTM had on model accuracies in the eight FMUs where it was available. Classification involved both PFTs and SFUs. Model 3 classified PFTs and SFUs in three northern FMUs where species composition was quite different than in the previous models. Binary classifications proved to give the highest model accuracies for all algorithms, with Random Forest models producing the highest accuracies for all classifications. Of the PFTs, the Mixedwood PFT consistently provided the lowest true positive rates and Tolerant Hardwood PFTs achieved the highest true positive rates in the multi-class models. Different SFUs were classified in different regions with various degrees of success. Future work is discussed, including the possibility of removing the manual calculations required in this method through the use of emerging technologies.;Keywords: Data Mining, Forestry, Ecology, Supervised Classification, Habitat Modeling, Climate Model. |
