Gaussian Processes for Satellite Data
Abstract
Using convolutional neural networks (CNNs) on satellite data sets outperforms previous meteorological methods for classifying weather events from space-based sensor data. However, CNNs alone lack the ability to quantify the uncertainty about their prediction given the input to the model. Gaussian processes (GPs) area mathematical technique for making predictions while providing an estimate of the functional uncertainty about the underlying model; however, they are more computationally expensive to train than traditional CNNs. We train a ResNet50 CNN augmented with GPs against a large satellite dataset but test the model across not only held out data from the training dataset but also two other large datasets of a tropical cyclone and mesoscale convective system to test model performance across a variety of weather events. We compare the accuracy of a ResNet50 CNN augmented with GPs for the output layer of the neural network to the Goddard Profiling Algorithm, a differential equation-based approach that is not based on neural networks, a fully-deterministic neural network, as well as several Bayesian neural networks(BNNs) and find that Gaussian Processes outperform the GPROF and deterministic approaches but fall short of the top-end BNN results. Additionally, more work is needed to compare if the uncertainty estimates from the GPs are better calibrated than the uncertainty estimates from the BNNs.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2023
- Accession Number
- AD1224750
Entities
People
- David W. Martin
Organizations
- Naval Postgraduate School