BENCHMARKING BAYESIAN DEEP LEARNING METHODS WITH MULTI-SPECTRAL SATELLITE IMAGERY

Abstract

The deep convolutional neural network (DCNN) is the current state-of-the-art approach for automatic image classification tasks. Historically, Bayesian deep learning methods have been applied to these models in narrow scopes. This thesis has created and tested several Bayesian deep learning models to perform classification on operational meteorological multi-spectral satellite data while quantifying the uncertainty in the model predictions. This large-scale dataset is used to compare the performance of Bayesian models against a DCNN and the current algorithm used by the National Aeronautics and Space Administration (NASA) to perform precipitation classification on the dataset. The use of a large-scale, operational dataset to benchmark Bayesian deep learning methods is the first application of its kind and represents a novel contribution to the fields of Bayesian deep learning and computer science. Several novel benchmarks were developed for use in this work. The best performing Bayesian model achieved 92 percent classification accuracy with demonstrated calibrated uncertainty on test data. All Bayesian models are shown to outperform current state-of-the-art DCNNs and the current operational algorithm. Furthermore, it is demonstrated that Bayesian model uncertainties can be used to screen uncertain predictions, and these uncertainties can be mapped spatially to identify specific regions of data that can be used to further improve the model performance.