ECE-based Deep Ensemble for Neural Network Calibration in Satellite Image Classification

We address the problem of deep neural network calibration in the context of satellite image classification, a problem highly related to the reliability of deep learning-based models in remote sensing applications. In this work we propose an Expected Calibration Error (ECE) based deep ensemble, and study different approaches to this technique. The experiments compare the proposed technique with single network models, and a standard average-based deep ensemble, on three different satellite imagery datasets. The achieved results show that it is possible to accomplish better calibrated predictions using the proposed ECE-based approach applied to deep ensembles, when compared with the aforementioned baselines.