深度学习方法下GEDI数据的天然云杉林地上生物量反演

Study on using deep learning method to retrieve the biomass of natural Picea forest from GEDI data

森林作为陆地最大碳库,对人类的生活与发展至关重要,精准掌握森林资源动态变化并对其进行现代化可持续发展已成为当下研究热点。本文以天山山脉的天然云杉林为研究对象,利用地面实测数据、直升机机载激光雷达点云数据以及全球生态系统动态调查激光雷达(Global Ecosystem Dynamics Investigation,GEDI)数据,构建多源融合数据框架,通过使用AutoKeras框架下的深度学习算法,实现GEDI数据的多个相对高度百分位数(Relative Height Percentile,RH)与其光斑内地上生物量的回归模型预测,验证GEDI数据在较大范围的地上生物量反演方面的可行性,主要结论如下:(1)GEDI数据用于森林地上生物量估测研究具有较高可行性,通过自动化深度学习算法,训练集、验证集、整体数据的决定系数(Coefficient of Determination,R2)分别为0.69、0.63和0.67,平均绝对误差(Mean Absolute Error,MAE)分为3.73 mg·hm^(-2)、4.22 mg·hm^(-2)和3.89 mg·hm^(-2),具有较高的预测精度。(2)直升机激光雷达作为GEDI数据估算地上生物量的中间技术,整个研究区内的单木识别准确率高于0.75。最终本次研究通过多模态数据融合,定量化描述研究区单木基础结构参数的同时,验证GEDI数据在获取森林地上生物量方面的潜力,也为相近区域大面积的森林碳源汇、生物量、蓄积量估算、森林管理与经营、生物多样性保护等多个项目研究提供理论基础,具有一定的指导意义和基础数据支撑作用。

As the largest carbon reservoir on land,forests play a crucial role in human life and development.Understanding the dynamic changes in forest resources and modernizing their sustainable development is currently a significant research focus.This study focuses on natural Picea forests in the Tianshan Mountains and uses ground measurement data,helicopter airborne LiDAR point cloud data,and Global Ecosystem Dynamics Investigation(GEDI)data to construct a multisource fusion data framework.By utilizing deep learning algorithms within the AutoKeras framework,the study aims to predict the regression model of multiple relative height quantiles of GEDI data and their aboveground biomass in the study area,thereby validating the feasibility of GEDI data for large-scale aboveground biomass retrieval.The main conclusions are as follows:(1)GEDI data are highly feasible for estimating forest aboveground biomass.Through automated deep learning algorithms and training and verification sets,the overall data achieve a coefficient of determination(R2)of 0.69,0.63,and 0.67,respectively,along with a mean absolute error of 3.73 mg·hm^(-2),4.22 mg·hm^(-2),and 3.89 mg·hm^(-2),demonstrating high prediction accuracy.(2)Helicopter LiDAR,an intermediate technology for estimating aboveground biomass using GEDI data,exhibits a single tree recognition accuracy of over 0.75 across the study area.The study successfully utilizes multimodal data fusion to quantitatively describe the structural parameters of the single tree foundation in the study area while verifying the potential of GEDI data for obtaining forest aboveground biomass.Moreover,the study provides a theoretical basis for estimating carbon sources and sinks,biomass,stock,forest management,biodiversity protection,and other projects in similar areas,offering essential guidance,and fundamental data support.