Improving Land Cover Classification Over a Large Coastal City Through Stacked Generalization with Filtered Training Samples

Abstract

To improve remote sensing-based land cover mapping over heterogenous landscapes, we developed an ensemble classifier based on stacked generalization with a new training sample refinement technique for the combiner. Specifically, a group of individual classifiers were identified and trained to derive land cover information from a satellite image covering a large complex coastal city. The mapping accuracy was quantitatively assessed with an independent reference data set, and several class probability measures were derived for each classifier. Meanwhile, various subsets were derived from the original training data set using the times of being correctly labeled by the individual classifiers as the thresholds, which were further used to train a random forest model as the combiner in generating the final class predictions. While outperforming each individual classifier, the combiner performed better when using the class probabilities rather than the class predictions as the meta-feature layers and performed significantly better when trained with a carefully selected subset rather than with the entire sample set. The novelties of this work are with the insight into the impact of different training sample subsets on the performance of stacked generalization and the filtering technique developed to prepare training samples for the combiner leading to a large accuracy improvement.