Improving litterfall production prediction in China under variable environmental conditions using machine learning...

Area of Science:

• Ecology
• Forestry
• Climate Change Science

Background:

• Litterfall production is vital for forest ecosystems and the global carbon cycle.
• Abiotic and biotic factors influencing litterfall have been studied using traditional statistical models.
• Machine learning models offer advanced methods for predicting forest litterfall.

Purpose of the Study:

• To evaluate machine learning models (RF, LightGBM, CatBoost, MERF) for predicting total annual litterfall production in Chinese forests.
• To identify key abiotic and biotic drivers of litterfall production.
• To project future litterfall production under different climate change scenarios.

Main Methods:

• Utilized 968 records from 314 forest sites across China.
• Compared linear mixed models with machine learning approaches, including Random Forest (RF), LightGBM, CatBoost, and Mixed Effect Random Forest (MERF).
• Analyzed the influence of climate-related features, forest age, and forest type on litterfall production.

Main Results:

• Machine learning models generally outperformed linear mixed models.
• MERF models demonstrated the highest predictive performance (R² = 0.7), attributed to their ability to capture nonlinear relationships.
• Mean annual temperature and forest age were positively correlated with litterfall production.
• Forest type significantly impacted litterfall, especially in needleleaf forests.
• Future litterfall production is projected to be highest under RCP 8.5, followed by RCP 4.5 and RCP 2.6.

Conclusions:

• MERF models provide a robust framework for predicting forest litterfall production.
• Climate variables and forest age are critical determinants of litterfall.
• Accurate estimation of current and future litterfall is essential for effective forest management and mitigating climate change impacts.