Optimizing Carbon Stock Estimation in Tree Vegetation through PRISMA Hyperspectral Remote Sensing Data and Machine Learning Approach: A Case Study in Mount Merbabu National Park

Abstract

Abstract The forest ecosystem's pivotal role in the carbon cycle and its impact on the global carbon balance underscore the significance of understanding and mitigating factors that contribute to carbon emissions. This study employs a combination of hyperspectral remote sensing (PRISMA) and machine learning techniques (Random Forest) to estimate the carbon stock of tree vegetation. Recognizing the necessity for variable optimization, the research focuses on identifying the most optimal variables from PRISMA hyperspectral imagery to model tree vegetation carbon stock. Additionally, the study evaluates the accuracy of the model by employing two variable selection methods: Stepwise Regression and Boruta. The research contributes to a comprehensive understanding of tree vegetation carbon dynamics by conducting estimation and mapping in Mount Merbabu National Park. Results indicate that the Random Forest-Boruta model consistently outperforms the Random Forest-Stepwise model, demonstrating superior accuracy and precision. Specifically, Random Forest-Boruta I (α = 0.01) exhibits a Root Mean Square Error (RMSE) of 2.25 ton/pixel, a normalized RMSE (nRMSE) of 22.77%, a Standard Error of Estimate (SEE) of 2.6 ton/pixel, maximum accuracy at 65.52%, and a Bias of 0.23. These findings provide valuable insights for policymakers and environmental stakeholders, offering a robust framework for managing and preserving forest ecosystems as part of global climate change mitigation strategies.