The power of voting: Ensemble learning in remote sensing

Abstract

Ensemble Learning, the concept of generating, training, and employing multiple machine learning models for inference rather than just one, is of increasing interest. It offers an interesting anthropomorphism by drawing similarities to human committees and the wisdom of crowds. Additionally to such a rather intuitive understanding, it has a sound theoretic foundation and an overwhelming amount of empirical evidence that it outperforms single estimators for a large variety of tasks including regression and classification. While ensembles come with an increased computational load and decreased interpretability, they lead to more accurate and robust predictions, are less prone to overfitting and adversarial attacks, and provide reliable uncertainty estimates of their predictions. This chapter discusses Ensemble Learning from various aspects, presents theoretical considerations and empirical heuristics and provides examples where it was successfully applied to analyze remote sensing and Earth observation data.