Satellite‐derived ecosystems classification: image segmentation by ecological region for improved classification accuracy, a boreal case study
An unsupervised image classification technique employing image segmentation by ecological regions is evaluated using percentage accuracies and tau coefficients against an unsegmented two-stage classification. k -fold cross-validation is used to partition the field data into training and testing sets. A Z -test of the tau statistic and its variance is used to test for a significant increase in classification accuracy when using image segmentation. Results show a significant increase in classification accuracy (a = 0.05, one-tailed) over two-stage approaches ( Z = 2.49, Z crit = 1.65 p = 0.0063). This supports our hypothesis that spectral variance within information classes can be explained, in part, by ecological region. Multi-group discriminant analysis is performed using jack pine ( Pinus banksiana ) plant community spectral data, grouped by ecological region. Results show significant spectral differences in a single information class within different ecological regions, which support the image segmentation approach to classification. The minimum mappable unit (MMU) is discussed in the context of Landsat Thematic Mapper (TM). The plant association, or ecosite, is presented as the MMU and the physical and ecological properties are discussed in relation to their spectral properties. The results suggest refinements in data collection and image analysis for remotely sensed data in boreal environments.
No Reference information available - sign in for access.
No Citation information available - sign in for access.
No Supplementary Data.
Document Type: Research Article
Affiliations: Earth Observation Systems Laboratory (EOSL), Department of Earth and Atmospheric Sciences, University of Alberta, Edmonton, Alberta, Canada T6G 2E3
Publication date: 2006-01-20