Scholarlink

<p>Accurately estimating above-ground biomass (AGB) is critical for understanding carbon storage and ecosystem dynamics, essential for sustainable forest management and climate change mitigation. In heterogeneous ecosystems like Miombo woodlands, traditional methods often face challenges due to high structural variability, limited data availability, and reliance on coarse-resolution satellite imagery, leading to reduced accuracy. This study assessed the performance of four machine learning models Extreme Gradient Boost (XGBoost), Random Forest (RF), Gradient Boosting (GBM), and Support Vector Machine (SVM) in predicting AGB using UAV-derived spectral and height data. RF emerged as the best-performing model, explaining 77% of the variance (R<sup>2</sup> = 0.77), with an RMSE of 48.7&#xa0;Mg/ha. XGBoost followed, achieving R<sup>2</sup> = 0.65 and RMSE = 52.9&#xa0;Mg/ha. GBM and SVM underperformed (R<sup>2</sup> = 0.28 and 0.29, respectively), likely due to their limitations in handling small, non-linear datasets.&#xa0;The generated AGB maps revealed significant spatial variability across the study areas, with most regions having AGB between 74 and 145&#xa0;Mg/ha and hotspots exceeding 192&#xa0;Mg/ha. To enhance the accuracy and scalability of AGB estimation in Miombo woodlands, future research should explore integrating UAV image-derived textural variables, deep learning models, and additional remote sensing data such as LiDAR or high-resolution satellite imagery.</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

Comparative evaluation of machine learning models for UAV-derived biomass estimation in Miombo Woodlands

  • 期刊论文

  • 发表日期：

    2025年2月19日

查看全文
  • Goodluck S. Melitha,
  • Japhet J. Kashaigili,
  • Wilson A. Mugasha

摘要

Accurately estimating above-ground biomass (AGB) is critical for understanding carbon storage and ecosystem dynamics, essential for sustainable forest management and climate change mitigation. In heterogeneous ecosystems like Miombo woodlands, traditional methods often face challenges due to high structural variability, limited data availability, and reliance on coarse-resolution satellite imagery, leading to reduced accuracy. This study assessed the performance of four machine learning models Extreme Gradient Boost (XGBoost), Random Forest (RF), Gradient Boosting (GBM), and Support Vector Machine (SVM) in predicting AGB using UAV-derived spectral and height data. RF emerged as the best-performing model, explaining 77% of the variance (R2 = 0.77), with an RMSE of 48.7 Mg/ha. XGBoost followed, achieving R2 = 0.65 and RMSE = 52.9 Mg/ha. GBM and SVM underperformed (R2 = 0.28 and 0.29, respectively), likely due to their limitations in handling small, non-linear datasets. The generated AGB maps revealed significant spatial variability across the study areas, with most regions having AGB between 74 and 145 Mg/ha and hotspots exceeding 192 Mg/ha. To enhance the accuracy and scalability of AGB estimation in Miombo woodlands, future research should explore integrating UAV image-derived textural variables, deep learning models, and additional remote sensing data such as LiDAR or high-resolution satellite imagery.