Integration of UAV-LiDAR and Sentinel-2A Data for Modeling Robinia pseudoacacia Aboveground Biomass via Artificial Neural Networks

Aboveground forest biomass (AGB), a critical ecosystem component, plays essential roles in maintaining ecological balance, conserving biodiversity, mitigating climate change, and supporting sustainable development. Optical remote sensing characterizes horizontal forest structures through multispectral and textural data, whereas UAV-LiDAR captures vertical structural parameters. Integrating these complementary technologies improves biomass estimation by enhancing model fit and accuracy. This multimodal approach significantly advances aboveground biomass quantification.This study investigated aboveground biomass (AGB) estimation in artificial Robinia pseudoacacia forests within the Caijiachuan watershed of Jixian County, Shanxi Province, utilizing multisource remote sensing data. Field survey data served as ground truths, whereas Sentinel-2A multispectral imagery and UAV-LiDAR point clouds provided optical and radar remote sensing data, respectively. After extracting feature variables from both datasets, we developed and compared AGB estimation models via four machine learning approaches—partial least squares regression (PLS), random forest regression (RF), backpropagation neural network (BPNN), and convolutional neural network (CNN)—to evaluate each method's performance with single-source and combined data inputs. The optimal model was subsequently applied to generate a spatial distribution map of Robinia pseudoacacia AGB density across the watershed, followed by analysis of biomass distribution patterns relative to topographic factors (elevation, slope, and aspect).Our results provide important insights as follows:(1) Compared with the single-variable model, the model incorporating both variables demonstrates superior performance (MAE = 2.53, MBE = -0.31, RMSE = 3.96, NMSE = 0.06) (MAE = 4.57, MBE = -2.92, RMSE = 5.61, NMSE = 0.11).(2) Among the four modeling approaches evaluated—partial least squares (PLS) regression, random forest (RF) regression, backpropagation neural network (BPNN), and convolutional neural network (CNN)—the CNN algorithm demonstrated superior performance in reducing estimation errors for Robinia pseudoacacia aboveground biomass.(3) The mean aboveground biomass (AGB) density of Robinia pseudoacacia in the Caijiachuan plantation watershed was 61.49 Mg·ha⁻¹. Approximately 87% of the total biomass occurred within the 1,050–1,200 m elevation range. Slope gradient analysis revealed maximum AGB accumulation on moderately steep and steep slopes, whereas minimum values were observed on gentle slopes. The aspect-dependent distribution revealed peak biomass on semii-shaded northeast-facing slopes, whereas the lowest biomass levels occurred on north-facing shaded slopes.These findings provide an effective method for assessing the aboveground biomass of Robinia pseudoacacia forests in the region, as well as valuable data for evaluating the carbon sequestration capacity of soil and water conservation forests.