摘要
【目的】准确地估测森林地上生物量(above ground biomass,AGB)对大区域森林资源调查和管理至关重要,机器学习算法能实现森林AGB高精度估测,但超参数的设置能直接影响模型效果。为了提升模型的构建效率和预测精度,研究通过构建超参数优化的机器学习算法进行森林AGB估测,并比较不同超参数下的模型误差变化。【方法】以西藏自治区江达县天然林为研究对象,利用森林资源调查数据提取实测森林AGB数据,结合Sentinel-2多光谱影像提取遥感变量。采用逐步回归法和Boruta法分别进行遥感变量筛选,构建多元线性回归模型、支持向量机模型和随机森林模型进行森林AGB反演。此外,对支持向量机模型和随机森林模型进行超参数优化,以提高模型反演精度。【结果】1)随机森林模型在所有反演模型中实现了最佳的估测精度,模型决定系数达到了20.63,同时实现了最低的均方根误差和相对均方根误差,分别为28.06 t/hm和23.03%。均方根误差相比多元线性回归模型和支持向量机模型分别降低了22.2%和12.1%。2)超参数优化可以有效地提高模型估测精度。通过分析不同参数组合下的误差变化趋势,确定最佳的参数组合,能有效地降低模型估测误差。3)较高的森林AGB值主要分布在东部、南部和东南部地区,中部地区和北部部分地区森林AGB值较小。超参数优化的随机森林模型森林AGB反演结果与研究区实际森林分布情况具有较好的一致性,整体反演效果较好。【结论】利用超参数优化的随机森林模型结合Sentinel-2遥感影像能实现较好的森林AGB反演效果,能为森林资源动态监测提供有效参考。
【Objective】Accurately estimating forest aboveground biomass(AGB)is crucial for large-scale forest resource investigation and management.Machine learning algorithms can achieve high-precision estimation of forest AGB,but the setting of hyperparameters can directly affect model performance.To improve the efficiency and prediction accuracy of the model,machine learning algorithms optimized by hyperparameters were constructed for forest AGB estimation,and the model error changes under different hyperparameters were compared.【Method】The study takes natural forests in Jiangda county,Tibet Autonomous Region as the research object,and the measured forest AGB data was extracted by forest resource survey data.Sentinel-2 multispectral images were used to extract remote sensing variables.The remote sensing variables were screened using stepwise regression method and Boruta method respectively,and multiple linear regression model(MLR),support vector machine(SVM)model and random forests(RF)model were constructed for forest AGB inversion.In addition,hyperparameter optimization was performed for the SVM model and RF model to improve the model accuracy.【Result】(1)The RF model achieved the best estimation accuracy among all inversion models.The RF achieved a coefficient of determination of 0.63,while achieving the lowest root mean square error(RMSE)and relative root mean square error(rRMSE)of 228.06 t/hm and 23.03%,respectively.The RMSE was reduced by 22.2%and 12.1%compared to the MLR model and SVM model,respectively.(2)Hyperparameter optimization can effectively improve the model estimation accuracy.By analyzing the error variation trend under different parameter combinations and determining the best parameter combination,the model estimation error was effectively reduced.(3)The higher forest AGB values were mainly distributed in the eastern,southern and southeastern regions,and the forest AGB values were smaller in the central region and some northern regions.The forest AGB inversion results of the hyperparametric optimized random forests model were in good agreement with the actual forest distribution in the study area,and the overall inversion effect is satisfactory.【Conclusion】The RF model with hyperparameter optimization combined with Sentinel-2 remote sensing images can achieve a better inversion of forest AGB,which can provide an effective reference for forest resource dynamics monitoring.
作者
熊向阳
杨小周
赵银超
李伟坡
XIONG Xiangyang;YANG Xiaozhou;ZHAO Yinchao;LI Weipo(Northwest Surveying and Planning Institute of National Forestry and Grassland Administration,Xi’an 710048,Shaanxi,China;Lutou Experimental Forest Farm,Central South University of Forestry and Technology,Yueyang 410400,Hunan,China)
出处
《中南林业科技大学学报》
CAS
CSCD
北大核心
2024年第5期102-111,共10页
Journal of Central South University of Forestry & Technology
基金
国家自然科学基金青年项目(31901241)。
