Examining interactions between and among predictors of net ecosystem exchange: A machine learning approach in a semi-arid landscape

Authors: ZHOU, QINTAO - University Of Oklahoma; Fellows, Aaron; Flerchinger, Gerald; ALEJANDRO, FLORES - Boise State University

Submitted to: Science of the Total Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/4/2019
Publication Date: 2/18/2019
Citation: Zhou, Q., Fellows, A., Flerchinger, G.N., Alejandro, F.N. 2019. Examining interactions between and among predictors of net ecosystem exchange: A machine learning approach in a semi-arid landscape. Science of the Total Environment. 9(1):2222. https://doi.org/10.1038/s41598-019-38639-y.
DOI: https://doi.org/10.1038/s41598-019-38639-y

Interpretive Summary: Semiarid regions play a key role in governing the global carbon exchange between land surfaces and the atmosphere, underscoring the need to estimate the temporal patterns of carbon cycling in semiarid ecosystems. This study developed and tested a model that used satellite-based remotely-sensed data to identify important predictors of carbon exchange and estimate net carbon uptake across semiarid rangelands. Predicted results show good agreement with the observed data. Scientist can use this information to further understanding the role of these ecosystems in modulating carbon and water exchange, which is critical to quantifying carbon and water balances across landscapes.

Technical Abstract: Net ecosystem exchange (NEE) is an essential climate indicator of the direction and magnitude of carbon dioxide (CO2) transfer between land surfaces and the atmosphere. Improved estimates of NEE, particularly if they can be developed at high spatial resolutions in space and time, can serve to better constrain spatiotemporal characteristics of terrestrial carbon fluxes, improve verification of land models, and advance monitoring of Earth’s terrestrial ecosystems. There are two different approaches to estimate NEE, 1) field-based measurement via eddy flux towers, and 2) deriving estimates from multispectral remote sensing data. Both methods are limited by either area or data gaps due to weather conditions. Using a machine learning approach, in this study we explore the accuracy with which the time series of NEE observed at a flux tower could be predicted by variables that could, in principle, be remotely sensed. The particular machine learning approach, random forests, allow us to determine the relative importance of different predictor variables and quantify the uncertainty of the model estimate in time. While the model is built with a mixture of remote sensing and ground-based data, all predictor variables correspond to different existing and/or planned remote sensing missions. We develop and test the model in a complex landscape of the Reynolds Creek Experimental Watershed (RCEW) in southwest Idaho, USA. Remote sensing data are in the form of 4-day composite estimates of the absorbed fraction of photosynthetically active radiation (fPAR) and leaf area index (LAI), derived from the MODerate-resolution Imaging Spectroradiometer (MODIS) at 1 km spatial resolution (MCD15A3 product). Point measurement data (i.e., daily NEE, soil moisture, downward solar radiation, precipitation, and mean air temperature) are from the U.S. Department of Agriculture, Agriculture Research Service, Northwest Watershed Research Center (NWRC). We explored the linear regression relationship between fPAR, LAI, precipitation, downward solar radiation, mean air temperature and soil moisture with ground-based NEE datasets. In an exploratory analysis of predictor and response variables we found NEE has a strong relationship with LAI (r2=0.45), mean air temperature (r2=0.61) and downward solar radiation (r2=0.64). We then developed a Random Forest (RF) model to predict observed NEE at two sites in the RCEW with a suite of these predictor variables. Across the RF model the most important predictors include LAI, downward solar radiation, and soil moisture. Predicted results show good agreement with the observed data for the NEE (r2=0.87). We then validated the temporal pattern of the NEE generated by the RF model for two independent years at the two sites not used in the development of the model.