融合神经网络与集成学习的复杂地形竹林生物量估测OACHSSCD

Accurate estimation of Aboveground Biomass(AGB)in bamboo forests is of paramount importance for assessing regional carbon budgets and supporting sustainable forest management,particularly given the high carbon sequestration potential of bamboo ecosystems.However,in mountainous regions characterized by complex terrain,such estimations face severe methodological challenges.These include the rapid saturation of optical remote sensing signals caused by the unique high-density,single-layer canopy structure of bamboo,as well as the spatial discontinuity of vertical structure data acquired by spaceborne LiDAR systems.Focusing on the Dendrocalamus giganteus forests in Xinping County,Yunnan,this study proposes a novel collaborative inversion framework that synergistically integrates Empirical Bayesian Kriging Regression Prediction(EBKRP)with a Heterogeneous Stacking Ensemble Learning strategy.To address the spatial limitations of LiDAR data,the framework first employs the EBKRP model,utilizing Sentinel-2 spectral features and topographic factors as auxiliary covariates,to spatially extrapolate discrete structural parameters from ICESat-2 and GEDI footprints into spatially continuous feature layers.Building upon this data foundation,the study constructs a robust two-layer Stacking ensemble model designed to maximize model heterogeneity.A significant methodological innovation of this framework is the strategic incorporation of a Multi-Layer Perceptron(MLP)into the base learner pool,alongside traditional machine learning algorithms including k-Nearest Neighbors(kNN),Random Forest(RF),Gradient Boosting Regression Tree(GBRT),and XGBoost.The introduction of the MLP is motivated by its distinct inductive bias compared to tree-based models;while traditional decision trees rely on recursive partitioning and orthogonal thresholds,the MLP utilizes a feedforward neural network architecture with non-linear activation functions to perform global approximation.This capability allows the model to effectively capture the highly non-linear mapping relationships between biomass and rugged terrain features that traditional shallow models often fail to resolve.A Ridge Regression meta-learner is then applied to integrate the predictions from these heterogeneous base learners,thereby achieving complementary advantages and reducing estimation bias.The experimental results demonstrate that the Stacking model achieved exceptional performance,with a Coefficient of Determination(R2)of 0.78 and a Root Mean Square Error(RMSE)of 12.89 Mg/hm2,significantly outperforming all individual base learners.Notably,the MLP emerged as the most effective single model,surpassing RF and kNN,which empirically validates the superiority of neural networks in handling complex topographic effects.The study estimated the average AGB density of the Dendrocalamus giganteus forest in the study area to be 77.0 Mg/hm2,with a total carbon stock of 1.126 × 106 Mg.Conclusively,this research confirms that a heterogeneous ensemble strategy incorporating neural networks provides a robust,high-precision paradigm for forest biomass mapping in complex mountainous environments,offering valuable methodological insights for future remote sensing applications.