5:15 PM - 6:45 PM
[MIS01-P13] Machine learning approach for estimating Net Ecosystem Exchange in Northern Eurasia
Keywords:Carbon dioxide fluxes, FLUXNET, MODIS, Machine learning , Gradient boosting
Accurate estimation of carbon dioxide (CO2) emission and uptake dynamics across diverse ecosystems, from local to regional scales, is crucial for comprehending historical climate change trends and predicting future scenarios. This is especially important for countries, committed to the Paris Agreement and seeking to substantiate their carbon emission regulations with robust data. The aim of this study is to develop a robust machine-learning (ML) model capable of accurately reproducing the pattern of Net Ecosystem Exchange (NEE) for any region of Northern Eurasia and in particular for Russia. By integrating meteorological and remote sensing data with the data of global network for greenhouse gas (GHG) flux monitoring stations (FLUXNET), the model designed to show the complex dynamics of NEE fluxes across the different ecosystems of Northern Eurasia. The developed ML model provides numerical values of NEE flux for all regions of Russia, creating an understanding of the carbon balance in different ecosystems.
Developed ML model is based on the CatBoost algorithm, the product of Russian IT Company, Yandex. CatBoost represents a proficient implementation of the gradient boosting technique, renowned for its efficacy in addressing similar regression challenges. The training dataset included data of NEE fluxes from the FLUXNET database, NASA POWER meteorological parameters and MODIS remote sensing data: Leaf Area Index (LAI), Fraction of Photosynthetically Active Radiation (fPAR), and Gross Primary Productivity (GPP). As a result, the model trained on this comprehensive dataset demonstrates good capacity to accurately interpret CO2 emission and uptake dynamics. The results of the final modeling revealed the following quality metric values on the test dataset: R-squared (R2) = 0.86 and Mean Absolute Percentage Error (MAPE) = 10.2%. These findings enable to assert that the model accurately represents NEE flux values. Moreover, the developed model has been integrated into a web service designed to automate the collection of remote sensing and meteorological data for the designated region of Russia. This integration facilitates the generation of maps illustrating the spatial distribution of NEE fluxes, which are essential for drawing conclusions regarding the region's carbon balance.
Developed ML model is based on the CatBoost algorithm, the product of Russian IT Company, Yandex. CatBoost represents a proficient implementation of the gradient boosting technique, renowned for its efficacy in addressing similar regression challenges. The training dataset included data of NEE fluxes from the FLUXNET database, NASA POWER meteorological parameters and MODIS remote sensing data: Leaf Area Index (LAI), Fraction of Photosynthetically Active Radiation (fPAR), and Gross Primary Productivity (GPP). As a result, the model trained on this comprehensive dataset demonstrates good capacity to accurately interpret CO2 emission and uptake dynamics. The results of the final modeling revealed the following quality metric values on the test dataset: R-squared (R2) = 0.86 and Mean Absolute Percentage Error (MAPE) = 10.2%. These findings enable to assert that the model accurately represents NEE flux values. Moreover, the developed model has been integrated into a web service designed to automate the collection of remote sensing and meteorological data for the designated region of Russia. This integration facilitates the generation of maps illustrating the spatial distribution of NEE fluxes, which are essential for drawing conclusions regarding the region's carbon balance.