Scholarlink

<p>Forests play a significant role in the carbon cycle by storing atmospheric carbon in different carbon pools of the forest ecosystem. Recent advancements in machine learning (ML) models and open-source satellite datasets provide the potential to estimate carbon stock at the local level in different carbon pools. Therefore, we quantified the accuracy and consistency of Random Forest (RF), Support Vector Regression (SVR), and Boosted Regression Tree (BRT) for predicting above-ground tree carbon (AGTC), soil organic carbon, (SOC) and litter carbon (LC) stock in tropical dry deciduous forests using Sentinel-1, Sentinel-2, and SRTM-DEM-based predictor variables. For each carbon pool, each model was run 100 times, and the average accuracy and coefficient of variation were used to identify the optimum model and assess the consistency of the models. The RF model achieved the highest average accuracy for predicting AGTC and SOC, and the BRT model exhibited the highest average accuracy for predicting LC stock. The average predicted carbon stock was 63.669 ± 25.904 Mg/ha, 10.576 ± 2.031 Mg/ha, and 1.383 ± 0.341 Mg/ha for AGTC, SOC, and LC, respectively. There was a variation in the accuracy of a specific ML model for predicting each carbon pool. Therefore, alongside evaluating model accuracy, it is essential to measure model consistency to obtain the optimum model and accurate predictions for each carbon pool. Also, this study highlights the usefulness of Sentinel-1 and Sentinel-2-based variables for carbon stock prediction in tropical dry deciduous forests.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Accuracy and consistency of the machine learning models for predicting carbon stock in different carbon pools using satellite-based predictor variables

  • 期刊论文

  • 发表日期：

    2025年2月4日

查看全文
  • Dipankar Bera,
  • Nilanjana Das Chatterjee,
  • Vivek Dhiman,
  • Subrata Ghosh,
  • Santanu Dinda

摘要

Forests play a significant role in the carbon cycle by storing atmospheric carbon in different carbon pools of the forest ecosystem. Recent advancements in machine learning (ML) models and open-source satellite datasets provide the potential to estimate carbon stock at the local level in different carbon pools. Therefore, we quantified the accuracy and consistency of Random Forest (RF), Support Vector Regression (SVR), and Boosted Regression Tree (BRT) for predicting above-ground tree carbon (AGTC), soil organic carbon, (SOC) and litter carbon (LC) stock in tropical dry deciduous forests using Sentinel-1, Sentinel-2, and SRTM-DEM-based predictor variables. For each carbon pool, each model was run 100 times, and the average accuracy and coefficient of variation were used to identify the optimum model and assess the consistency of the models. The RF model achieved the highest average accuracy for predicting AGTC and SOC, and the BRT model exhibited the highest average accuracy for predicting LC stock. The average predicted carbon stock was 63.669 ± 25.904 Mg/ha, 10.576 ± 2.031 Mg/ha, and 1.383 ± 0.341 Mg/ha for AGTC, SOC, and LC, respectively. There was a variation in the accuracy of a specific ML model for predicting each carbon pool. Therefore, alongside evaluating model accuracy, it is essential to measure model consistency to obtain the optimum model and accurate predictions for each carbon pool. Also, this study highlights the usefulness of Sentinel-1 and Sentinel-2-based variables for carbon stock prediction in tropical dry deciduous forests.