A machine learning approach for improved GCC predictive model of GPP in semi-arid ecosystems along an ecological state transition gradient in the southwestern US

Authors: Denham, Sander; Browning, Dawn; BARNES, MALLORY - Indiana University

Submitted to: American Geophysical Union
Publication Type: Abstract Only
Publication Acceptance Date: 9/1/2023
Publication Date: 12/11/2023
Citation: Denham, S.O., Browning, D.M., Barnes, M.L. 2023. A machine learning approach for improved GCC predictive model of GPP in semi-arid ecosystems along an ecological state transition gradient in the southwestern US. American Geophysical Union. Abstract.

Interpretive Summary:

Technical Abstract: Dryland ecosystems make up ~40% of the land surface and are thus an important contribution to the global carbon budgets. Despite their importance, these ecosystems are largely underrepresented in models due to the heterogeneity in vegetation composition and their propensity towards ecological state transitions caused by various environmental stressors. These circumstances make it difficult to assess seasonal productivity dynamics which are important for carbon monitoring and mitigation efforts. The characteristics of the landscape also make it difficult to rely on available satellite-based products alone for an accurate representation of these landscapes and for determining the magnitude of various ecosystem processes (i.e., carbon and water cycling). As such, we aimed to leverage near-surface digital imagery (PhenoCam) and eddy-covariance (EC) flux tower measurements to quantify the relationships between the green chromatic coordinate (GCC) and the sum of daytime gross primary productivity (GPP) along a gradient ecological state transition: a grassland (BOER), woody shrubland (DUNE), and a more novel composite of the two (NOVEL) for 2022. Subsequently, we explored the additional meteorological variables required to improve model estimates of GPP and how the order of variable dependence varied across sites. Notably, the strongest relationship between GCC and GPP occurred at the NOVEL site (R2 = 0.71) which improved to R2 = 0.90 with the addition of air temperature, soil moisture content, vapor pressure deficit, net radiation, and precipitation, which were selected using a random forest model. The next strongest GCC vs GPP relationship occurred at BOER (R2 = 0.65) followed by DUNE (R2 = 0.36). Both BOER and DUNE model predictions were improved with the addition of the same meteorological variables at NOVEL however, the extent of the improvement fell short of what was achieved at NOVEL, and with different degrees of variable dependence. These important results highlight the utility of GCC derived from digital imagery in dryland ecosystems with modest vegetation cover and the ability to make predictions of ecosystem processes (i.e., GPP) which can be achieved at a lower cost than expensive eddy-covariance systems and are a more accurate representation of these processes in the systems that are characterized by high heterogeneity which is more difficult to capture solely with satellite data products (i.e., MODIS, LANDSAT).