Identifying and eliminating mislabeled training instances

摘要: This paper presents a new approach to identifying and eliminating mislabeled training instances. The goal of this technique is improve classification accuracies produced by learning algorithms improving the quality data. employs an ensemble classifiers that serve as filter for Using n-fold cross validation, data passed through filter. Only instances classifies correctly are final algorithm. We present empirical evaluation task automated land cover mapping from remotely sensed Labeling error arises in these multitude sources including lack consistency vegetation used, variable measurement techniques, variation spatial sampling resolution. Our shows noise levels less than 40%, filtering results higher predictive accuracy not filtering, class or equal 20% allows base-line be retained. suggest effective method labeling errors, further, will significantly benefit ongoing research develop accurate robust remote sensing-based methods map at global scales.