A Method for Selecting Training Data and its Effect on Automated Land Cover Mapping of Large Areas

Many remote sensing projects require the utilization of sample data for training a supervised classification algorithm. The correctness of training data is always highly important for accurate image classification. While data obtained during field studies are suitable for many small scale studies and the classification of high and very high spatial resolution data, automated procedures are necessary to map large areas with medium to coarse resolution images. Most coarse resolution land cover classifications are based on previous studies and mapping efforts. This study illustrates a procedure for training data selection of coarse resolution images if a high resolution map already exists. In a second step the paper analyzes the impact of training data selection parameters on classification accuracy. The study is based on MODIS time series metrics and employs a set of decision trees. This classifier derives a fuzzy image classification. Tests in Germany yielded on average a 10% increase in classification accuracy.