光学精密工程2025，Vol.33Issue(22)：3460-3474,15.DOI:10.37188/OPE.20253322.3460
整合主被动遥感数据特征插值与机器学习的森林地上生物量反演与制图方法
A method for forest biomass inversion and mapping by integrating active and passive remote sensing data feature interpolation and machine learning
摘要
Abstract
Aiming at the problem of spatial sparsity of spaceborne lidar data,this paper proposed a multi-source data fusion method based on feature interpolation,which realized regional-scale forest aboveground biomass(AGB).Firstly,the three-dimensional features were extracted from GEDI L2A/L2B and ICE-Sat-2/ATL08 data;the data set of spot-scale feature variables was constructed by combining Sentinel-2 spectral feature variables and terrain factors.Then,the correlation analysis was carried out to eliminate the high-collinearity feature variables,and the three regression algorithms of CatBoost,RF and LightG-BM were compared to identify the optimal model.Subsequently,based on CatBoost feature importance and SHAP analysis,key predictor variables were further identified.Finally,the key feature variables of LiDAR were interpolated to obtain continuous raster features,and then the forest AGB spatial mapping was realized by the optimal regression model.The validation results demonstrated that CatBoost per-formed best in spot-scale modeling(R2=0.88,RMSE=78.74 Mg/ha,rRMSE=20.93%);the spatial mapping accuracy based on feature interpolation and multi-source data fusion is R2=0.82,RMSE=60.90 Mg/ha,and rRMSE=36.54%.Compared to regression mapping using optical remote sensing imagery alone,the rRMSE was reduced by approximately 34.7%.The feature interpolation strategy was used to spatially continuous the key structural variables of the laser spot and fuse them with high-resolution optical and topographic information,which can mitigate sparse laser-footprint sampling and the lack of vertical-structure information in optical images.It enhanced regional forest AGB estimation accuracy.And the method provides a valuable reference for large-scale forest carbon stock assessment and ecosystem monitoring.关键词
全球生态系统动态调查/冰云和陆地高度卫星/激光雷达/光学遥感影像/特征插值/地上生物量制图Key words
Global Ecosystem Dynamics Investigation(GEDI)/ICESat-2/LiDAR/optical remote sensing imagery/feature interpolation/aboveground biomass mapping分类
信息技术与安全科学引用本文复制引用
QI Qi,WANG Hongtao,FENG Baokun,WANG Cheng,WANG Yingchen,ZHANG Shuting..整合主被动遥感数据特征插值与机器学习的森林地上生物量反演与制图方法[J].光学精密工程,2025,33(22):3460-3474,15.基金项目
国家自然科学基金项目(No.U22A20566) （No.U22A20566）
测绘科学与技术"双一流"学科创建项目(No.SYSB202503) （No.SYSB202503）
河南理工大学基本科研业务费专项(No.NSFRF220203) （No.NSFRF220203）
青年创新探索性基金项目(No.NS-FRF2502029) （No.NS-FRF2502029）
