Concurrent estimation of water-quality indicators and uncertainties using Mixture Density Networks
Arun M Saranathan, Science Systems and Applications, Inc., fnu.arunmuralidharansaranathan@nasa.gov (Presenter)
Nima Pahlevan, SSAI / NASA GSFC, nima.pahlevan@nasa.gov
Space-airborne optical sensors have proven to be a valuable source of information for monitoring the health and outlook of various water bodies. Machine learning models like Mixture Density Networks (MDNs) have emerged as powerful tools for estimating Water Quality Indicators (WQIs) such as Chlorophyll-a, Total Suspended Solids (TSS), and Colored Dissolved Organic Materials (acdom) present in the water column from reflectance (Rrs) data. While the MDN has shown promising prediction performance, the performance guarantees do not extend to scenarios wherein the test data is dissimilar to the data encountered by the model in training. Therefore, the use of satellite based MDN predictions requires the estimation of an additional uncertainty metric that can reliably inform water resource managers on the model’s confidence in a specific prediction, as well as the possible scale of errors. Attempts to leverage the Bayesian nature of the MDN output to estimate the confidence associated with each MDN prediction have shown some success. While such metrics are very informative as to the model’s confidence in a specific prediction, based on factors such as the similarity of test and training data, etc., it does not provide the end-user an accurate estimate of the possible error in the model prediction. To bridge this gap, a simple statistical analysis of the error vs uncertainty trends in different intervals of the WQIs to identify an appropriate “scaling factor” that maps the model uncertainty to an approximation of expected error with an appropriate confidence interval. The scaled uncertainty is compared to the unscaled version quantitatively (for in situ samples) and qualitatively (for predictions on satellite data) to highlight its utility.
Associated Project(s):
Poster Location ID: 1-16
Presentation Type: Poster
Session: Poster Session 1
Session Date: Tue (May 9) 5:00-7:00 PM
CCE Program: OBB