Carbon stock assessment and estimation using machine learning: A case study on acacia forest plantation in Rupat Island peat ecosystem, Indonesia

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a framework for using random forest to estimate carbon stocks in peat forests, useful for carbon cycle modelers.

🏢実務担当者:Demonstrates how ML can improve accuracy of carbon stock estimates for plantation management and carbon offset projects.

🏛政策担当者:Offers scientific evidence for national carbon accounting and emission reduction strategies in agricultural sector.

• Peatland carbon stock up to 806.45 ± 136.48 Mg C ha -1 ; SOC contributes 93.40% at 0–40 cm • TCS declines with Acacia age, which influences carbon content by 24% (p < 0.001) • RF machine learning captures carbon dynamics at vegetation, soil, and accumulation levels • Improving ML models is essential to enhance RF performance on peat data characteristics • RF spatial mapping estimated 13.94 million Mg TCS across a 16,534 ha plantation area Tropical peat ecosystems in Indonesia play an important role in controlling emissions, but can also exacerbate climate change if managed unsustainably. Sustainable management of Acacia forest plantations implies maintaining available carbon stocks. However, the quantification of carbon stocks modeled using machine learning is still very limited in Indonesia. Therefore, this study used carbon stock plots in cultivated peat ecosystems for Acacia on Rupat Island, Indonesia. The results of this study show that the available carbon stock in peat with an average pH of 3.76 has a total carbon stock (TCS) of 806.45 ± 136.48 Mg C ha⁻¹ (maximum 1185.23 Mg C ha⁻¹). Carbon stocks from peat soils contributed the most to TCS, with an average percentage of 93.40% at a peat depth of 0-40 cm. Dynamically, TCS is negatively affected by Acacia age, with an influence of 24% (p-value <0.001), indicating that the sustainability of Acacia forest plantations is primarily determined by the dynamics of carbon stocks, which will continue to decline in the future. Based on testing three prediction models using random forest machine learning on AGC-SOC-TCS, the following R-squared values were obtained: 0.65 (RMSE 10.9 Mg C ha⁻¹), 0.35 (RMSE 143 Mg C ha⁻¹), and 0.27 (RMSE 136 Mg C ha⁻¹). This research confirms that the machine learning approach can be an option and is recommended for studying carbon dynamics in peat ecosystems at the regional and national levels. In addition, this information can also serve as science evidence in national action to reduce carbon emissions in the agricultural sector.

• openalex https://doi.org/10.1016/j.envc.2026.101518first seen 2026-06-02 04:43:47 · last seen 2026-06-13 04:30:32