This site requires Cookies enabled in your browser for login.
. . .
Alias List Editor
AOSM2022: Uncertainty estimations for mapping lake ice using random forest on MODIS TOA reflectance data
Section 1: Publication
Authorship or Presenters
Nastaran Saberi, Claude Duguay, Andrea Scott
Uncertainty estimations for mapping lake ice using random forest on MODIS TOA reflectance data
Hydrology and Terrestrial Ecosystems
poster plus 2-minute lightning talk
Nastaran Saberi, Claude Duguay, Andrea Scott (2022). Uncertainty estimations for mapping lake ice using random forest on MODIS TOA reflectance data. Proceedings of the GWF Annual Open Science Meeting, May 16-18, 2022.
AOSM2022 Core modelling, TTSW
Section 2: Abstract
Plain Language Summary
Lake ice coverage products are a requirement identified by the climate community for improving numerical weather prediction and atmospheric reanalysis products, as well as for climate monitoring as determined by the Global Climate Observing System (GCOS). There are many suitable sources of observations available for mapping and monitoring lake ice coverage such as optical satellite data with the most practical ones from the Moderate Resolution Imaging Spectroradiometer (MODIS) over the last two decades. Considering the limitation of the presence of cloud cover and daylight dependency to capture imagery by optical sensors, the high revisit time of NASA’s Terra and Aqua satellites that carry MODIS allows for the production of lake ice maps required for operational and research-based projects.
Building on our previous research findings concluded from a GWF-supported project on lake ice cover mapping of Lake Erie from RADARSAT data, we are proposing a method to characterize inherent uncertainties (aleatoric) and model uncertainties (epistemic) for the production of daily lake ice maps. Random Forest (RF) is used for classifying lake ice, water, and cloud and for measuring and quantifying predictive uncertainty. As RF is an ensemble-based approach, it allows learning different hypotheses (different trees); and therefore, it provides different expected outcome. The total uncertainty in a prediction can be calculated by the (Shannon) entropy of the predictive posterior distribution, whereas calculating the entropy of each probability distribution and then computing the average gives the aleatoric uncertainty. Epistemic uncertainty is then calculated by subtracting aleatoric from total uncertainties. Uncertainty estimates expands product usability, making researchers aware of aleatoric and epistemic uncertainty when incorporating ice fractions in their physical/numerical lake models in the form of direct integration of observation error variance or as a quality control flag.
Section 3: Miscellany
University of Waterloo
First Author: Dr. Nastaran Saberi, Reserch associate, Department of Geography and Environmental Management, University of Waterloo
Additional Authors: Prof. Claude Duguay, Department of Geography and Environmental Management, University of Waterloo; Dr. Andrea Scott, Department of Systems Design Engineering, University of Waterloo
Section 4: Download
T-2022-04-24-m1hbH6fohikOEWb957eDYrA Conference Publication 1.0