An ensemble-based approach for classification of high-resolution satellite imagery of the Amazon Basin

摘要

Land cover mapping of the Amazon Basin provides valuable insight for assessment of regional land use and the extent of human encroachment, which can help environmentalists detect and respond to deforestation. Current methods rely on coarse-resolution imagery, which limits effectiveness at identifying small-scale deforestation. This paper proposes an ensemble-based method utilizing weighted voting of heterogeneous classifiers for accurate and scalable labeling of high-resolution satellite imagery of the Amazon rainforest. The ensemble learner relies on the predictions of multiple classification models trained in parallel on a publicly available dataset of satellite images with crowdsourced labels, which are aggregated and incorporated into the ensemble as meta-features. Performance of the ensemble is evaluated on a hold-out validation set and is found to show a marked improvement over the performance of any individual classifier.