Probabilistic Stacked Ensemble and Automated Machine Learning Approach in Remote Sensing Data

Abstract

Background/Aim: In the classification of remote sensing data, Machine Learning (ML) models when used alone fall short in capturing spatial and spectral features. These models, especially in the classification of hyperspectral images with similar spectral characteristics, produce high generalization errors. In this study, a probability-based Stacked Ensemble and Automated Machine Learning (AutoML) approach that carries the confidence information of the base learners into the combination stage is proposed to overcome these issues.Methods: The Stacked model is an ensemble method that combines the outputs of multiple ML or deep learning models with a second meta model. Naive Bayes, Decision Trees, Artificial Neural Networks, K-Nearest Neighbors and Support Vector Machines were used as base learners. The class probability scores produced by these models were presented as input to a meta-learner whose hyperparameters were optimized by AutoML, instead of being used directly as output.Results: The proposed framework achieved the highest Macro-F1 score on both datasets (0.9297 on Indian Pines and 0.9573 on Pavia University), significantly outperforming every individual base learner as well as the tree-based ensembles, the voting strategies and the deep learning architectures included in the comparison.Conclusion: The findings show that by combining the strengths of different models, more accurate and more class-balanced predictions can be made compared to single models.