While anthropogenic methane emissions represent 2/3 of the global total annual emissions, the rest 1/3 constitute natural emissions, are dominated by natural wetlands and significantly vary from year to year depending highly on various climatological parameters.
Currently, process-based models are used to compute wetland extent and CH4 wetland fluxes based on soil carbon content, soil temperature, humidity, etc. but they can be time consuming and may require significant calibration effort.
We substitute these process-based models by a statistical model that can be easily implemented, operates efficiently, and reduces significantly the computation time using novel Machine Learning algorithms.
Currently, process-based models are used to compute wetland extent and CH4 wetland fluxes based on soil carbon content, soil temperature, humidity, etc. but they can be time consuming and may require significant calibration effort.
We substitute these process-based models by a statistical model that can be easily implemented, operates efficiently, and reduces significantly the computation time using novel Machine Learning algorithms.
We tested various machine learning (ML) algorithms (fast-forward multi-linear regression, random forest) to develop a statistical algorithm to compute wetland methane emissions based on an ensemble of predictors.
The algorithm is trained per Plant Functional Type and Koeppen-Geiger climate classification type (PFT+KG) using as dependent variable daily and monthly CH4 wetland emissions produced by ORCHIDEE-WET, while the explanatory (independent) climatological variables are obtained from the ERA5 climate re-analysis dataset.
Statistical evaluation criteria such as the Root Mean Squared Error (RMSE), and Mean Squared Deviation (MSD) that were implemented to evaluate model performance showed that the results from different ML algorithms present very good correlation each time during global modelling and specifically the random forest (RF) approach was able to simulate the results with great efficiency while operating on a daily timestep (R2 > 0.95).
Although the linear regression (MLR) approach showed very good correlation in higher northern latitudes (R2 > 0.96), it presents some discrepancies for the tropics (R2 > 0.88), where more complex approaches are required to simulate wetland methane emissions and are mainly captured by the random forest approach.
Soil temperature is the main driver for higher northern latitudes, while for the tropics and southern high latitudes soil moisture and surface pressure mostly affect the production of wetland methane emissions.
The algorithm is trained per Plant Functional Type and Koeppen-Geiger climate classification type (PFT+KG) using as dependent variable daily and monthly CH4 wetland emissions produced by ORCHIDEE-WET, while the explanatory (independent) climatological variables are obtained from the ERA5 climate re-analysis dataset.
Statistical evaluation criteria such as the Root Mean Squared Error (RMSE), and Mean Squared Deviation (MSD) that were implemented to evaluate model performance showed that the results from different ML algorithms present very good correlation each time during global modelling and specifically the random forest (RF) approach was able to simulate the results with great efficiency while operating on a daily timestep (R2 > 0.95).
Although the linear regression (MLR) approach showed very good correlation in higher northern latitudes (R2 > 0.96), it presents some discrepancies for the tropics (R2 > 0.88), where more complex approaches are required to simulate wetland methane emissions and are mainly captured by the random forest approach.
Soil temperature is the main driver for higher northern latitudes, while for the tropics and southern high latitudes soil moisture and surface pressure mostly affect the production of wetland methane emissions.
No comments:
Post a Comment