Scholarlink

Context <p>Spatially accurate estimates of forest above-ground biomass (AGB) are indispensable for carbon-stock accounting and sustainable silviculture. Existing mapping approaches face challenges in densely vegetated Coastal Plain forests because of seasonal optical variability, radar–optical saturation, and limited wall-to-wall structural information.</p> Objectives <p>We aimed to (i) develop and evaluate a multisensor, AI-enabled fusion framework for landscape-scale AGB mapping, (ii) quantify the added value of seasonal optical data and photogrammetric canopy-height profiles, and (iii) interpret model drivers using explainable artificial intelligence (AI) to relate predictors to forest structure and composition.</p> Methods <p>We mapped AGB across ~ 10,500 km<sup>2</sup> in southeastern North Carolina using wall-to-wall predictors from optical, radar, and photogrammetric sources. Forest Inventory and Analysis plot data (n = 305) were used to train and evaluate an ensemble of gradient-boosted tree models (CatBoost, LightGBM, XGBoost) and a neural network (RealMLP) via cross-validation. Model behavior was interpreted using feature importance and partial dependence analysis.</p> Results <p>Expanding Sentinel-2 temporal coverage from summer-only to four-season composites improved normalized RMSE by 8.7%. Incorporating canopy-height profiles from NAIP produced the largest accuracy gain, lowering nRMSE by 15.9–18.0% relative to the multisensor baseline, which underscores the critical value of structural information for AGB prediction. Three key predictors illustrated complementary ecological dimensions: the 10th percentile canopy height captured canopy openness, L-band polarimetric alpha indicated volume-scattering regime, and spring red-edge reflectance captured vegetation biochemistry. These findings show that fusing structure, polarimetry, and spectral phenology yields robust AGB maps and improves generalizability across heterogeneous landscapes.</p> Conclusions <p>This transferable, broadly accessible framework integrating structural, polarimetric, and spectral phenology data enables landscape-scale AGB monitoring and supports targeted conservation planning, restoration tracking, and adaptive management for carbon sequestration. The incorporation of high-resolution wall-to-wall structural data is particularly valuable for improving the accuracy and usability of forest AGB maps, thereby informing more responsive decision-making.</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

AI-powered multisensor fusion for forest biomass mapping: photogrammetric canopy profiles improve estimates in Southeastern North Carolina

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月30日

查看全文
  • Chao Wang,
  • Conghe Song,
  • Todd A. Schroeder,
  • Tamlin M. Pavelsky,
  • Zhengxiao Yan,
  • Yulong Zhang,
  • Fangfang Yao,
  • Yuwei Fu

摘要

Context

Spatially accurate estimates of forest above-ground biomass (AGB) are indispensable for carbon-stock accounting and sustainable silviculture. Existing mapping approaches face challenges in densely vegetated Coastal Plain forests because of seasonal optical variability, radar–optical saturation, and limited wall-to-wall structural information.

Objectives

We aimed to (i) develop and evaluate a multisensor, AI-enabled fusion framework for landscape-scale AGB mapping, (ii) quantify the added value of seasonal optical data and photogrammetric canopy-height profiles, and (iii) interpret model drivers using explainable artificial intelligence (AI) to relate predictors to forest structure and composition.

Methods

We mapped AGB across ~ 10,500 km2 in southeastern North Carolina using wall-to-wall predictors from optical, radar, and photogrammetric sources. Forest Inventory and Analysis plot data (n = 305) were used to train and evaluate an ensemble of gradient-boosted tree models (CatBoost, LightGBM, XGBoost) and a neural network (RealMLP) via cross-validation. Model behavior was interpreted using feature importance and partial dependence analysis.

Results

Expanding Sentinel-2 temporal coverage from summer-only to four-season composites improved normalized RMSE by 8.7%. Incorporating canopy-height profiles from NAIP produced the largest accuracy gain, lowering nRMSE by 15.9–18.0% relative to the multisensor baseline, which underscores the critical value of structural information for AGB prediction. Three key predictors illustrated complementary ecological dimensions: the 10th percentile canopy height captured canopy openness, L-band polarimetric alpha indicated volume-scattering regime, and spring red-edge reflectance captured vegetation biochemistry. These findings show that fusing structure, polarimetry, and spectral phenology yields robust AGB maps and improves generalizability across heterogeneous landscapes.

Conclusions

This transferable, broadly accessible framework integrating structural, polarimetric, and spectral phenology data enables landscape-scale AGB monitoring and supports targeted conservation planning, restoration tracking, and adaptive management for carbon sequestration. The incorporation of high-resolution wall-to-wall structural data is particularly valuable for improving the accuracy and usability of forest AGB maps, thereby informing more responsive decision-making.