B53G-06:
Data Mining Approach for Evaluating Vegetation Dynamics in Earth System Models (ESMs) Using Satellite Remote Sensing Products

Friday, 19 December 2014: 2:55 PM
Shijie Shu1, Forrest M Hoffman2,3, Jitendra Kumar3, William Walter Hargrove4 and Atul K Jain1, (1)University of Illinois at Urbana Champaign, Department of Atmospheric Sciences, Urbana, IL, United States, (2)University of California Irvine, Department of Earth System Science, Irvine, CA, United States, (3)Oak Ridge National Laboratory, Climate Change Science Institute, Oak Ridge, TN, United States, (4)USDA Forest Service Southern Research Station, Eastern Forest Environmental Threat Assessment Center, Asheville, NC, United States
Abstract:
Uncertainties in data retrieved from remote sensor present challenges to using such observational products to constrain Earth system model (ESM) results. While simple statistics can be applied to compare models with observations, advanced data mining methods, like unsupervised cluster analysis, offer powerful tools for summarizing model-data differences in the spatial and temporal patterns of ecological characteristics. We compared modeled land surface phenology with MODIS 16-day composited Normalized Difference Vegetation Index (NDVI) (MOD13C1) and Global Inventory Modeling and Mapping Studies (GIMMS) NDVI3g products spanning years 2001 to 2012. Annual traces of NDVI from two ESMs (CESM-CLM and ISAM) were calculated using modeled radiation from the output of historical simulations and corrected to better match observed properties by considering instrumental bandwidths and solar angle. Multivariate Spatio-Temporal Clustering (MSTC) was applied to annual traces of MODIS and GIMMS NDVI to create phenological regions (phenoregions) and analyzed using Mapcurves, a tool designed for comparing categorical maps, to check the consistency of the spatial patterns of observations before assessing model output. To evaluate modeled phenology, MSTC was first applied to obtain representative centroids of modeled NDVI and separately of satellite NDVI. The Mapcurves method was applied to compare the spatial patterns of modeled phenology to remote sensing observations. Next, modeled NDVI were projected onto the centroids defining phenoregions of observed NDVI, and observed NDVI were projected onto the centroids of modeled NDVI. Mapcurves was then applied to compare the spatial patterns of these classifications. Results showed a general agreement in the spatial pattern of phenoregions from models to satellite observations, except in high-latitude regions and agricultural areas. MSTC averages out small deviations between modeled and observed phenology, which are exhibited across all biome types. However, Mapcurves results showed a relatively low goodness of fit score for modeled phenology projected onto observations. This study demonstrates the utility of a data mining approach for cross-validation of observations and evaluation of model performance.