Machine Learning-Based Modeling of Net Ecosystem Exchange Using Numerical Weather Data and Satellite Images

Abstract

Recently, the increasing severity of climate change attributable to global warming has emphasized the imperative of carbon absorption to mitigate greenhouse gas emissions. The use of the carbon sink based on the carbon absorption and storage functions of forests is suggested as an effective alternative for domestic greenhouse gas reduction. Additionally, agricultural land cover comprises approximately 38% of the Earth's surface, underscoring the importance of comprehensively understanding the carbon cycle within not only forests but also agricultural landscapes. This significance arises from the fact that agricultural land locally amplifies seasonal variations in carbon dioxide by approximately 25% compared to vegetated areas. Consequently, a comprehensive understanding of both forest and agricultural land carbon cycles is imperative, necessitating quantitative analysis of carbon uptake in agricultural settings. Thus, this study aims to construct a machine learning-based model for estimating the net ecosystem exchange (NEE) of rice paddies in South Korea using ground flux data, meteorological variables, and satellite images. Through quantitative assessment, the NEE was determined, with a mean absolute error of 1.387, root mean square error of 2.203, and correlation coefficient of 0.872. Notably, observed NEE values demonstrating extremes in magnitude were associated with calculation errors, reflecting tendencies of both underestimation and overestimation. This phenomenon is likely attributed to the study's reliance on a limited dataset and the inherent challenges of training models across a broad spectrum of observations. To enhance calculation accuracy, future endeavors should focus on accumulating a more extensive repository of NEE flux observations and leveraging high-resolution satellite imagery and meteorological datasets for refining machine learningbased models.

Full Paper