Automatically labeling remotely sensed unsupervised clusters using GIS data and a hierarchical key scheme

摘要

Labeling spectral clusters is a challenge for mapping projects that use unsupervised methods for classifying remotely sensed imagery. The most common method of labeling clusters involves manual interpretation, which can be time consuming for large projects. A method was developed for automatically labeling remotely sensed unsupervised clusters using GIS data and a hierarchical key scheme through an exclusion process. The hierarchical key scheme uses polyfurcations to systematically label each cluster. For example, a cluster is first identified as being vegetated or non-vegetated. If it is vegetated (non-vegetated is excluded), then it is identified as a treed or non-treed cluster. If the cluster is treed, then it is labeled as coniferous, broadleaf, or mixed forest type. The classification hierarchy follows the hierarchy of the Canadian National Forest Inventory and the classification legend of a national scale satellite land cover mapping project in Canada, Earth Observation for Sustainable Development of Forests (EOSD). The data processing procedures were integrated into a geo-processing model using ArcGIS model builder, python scripts, and external processes. Classification accuracies were assessed for a test area in Western Newfoundland and Labrador, Canada. The overall accuracy for the finest level of the hierarchical key scheme representing the EOSD land cover classes was 59%. Accuracies increased as the number of classes descreased and the level of the hierarchy became more generalized. Overall accuracies for forest type, cover type, and a simple vegetated and non-vegetated land discrimination were 73%, 96% and 100%, respectively.