基于深度学习的树冠分割及生物量估算  

Canopy Segmentation and Biomass Estimation Based on Deep Learning

作　　者：赵子琪 李丹丹[2] 赵鼎 程志博 郭晓杰 ZHAO Ziqi;LI Dandan;ZHAO Ding;CHENG Zhibo;GUO Xiaojie(College of Computer and Control Engineering,Northeast Forestry University,Harbin 150040,China;College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin 150040,China)

机构地区：[1]东北林业大学计算机与控制工程学院,哈尔滨150040 [2]东北林业大学机电工程学院,哈尔滨150040

出　　处：《森林工程》2024年第5期145-155,共11页Forest Engineering

基　　金：国家重点研发计划项目(ZKJB230100003)。

摘　　要：地上生物量反映植被生长状况和碳储量的大小,该参数的准确性对于碳循环研究以及减缓气候变化至关重要。以芭蕉树为研究对象,提出一种利用深度学习实现芭蕉树冠检测分割和地上生物量估算的新思路。首先,以深度学习算法YOLOv8s-seg为基础框架改进,并应用无人机遥感影像,实现芭蕉树冠检测分割;然后,提取芭蕉树冠覆盖面积,结合实测地上生物量数据进行拟合,使用线性回归、K最近邻算法(K-NearestNeighbor,KNN)、支持向量机、随机森林和XGBoost(eXtreme Gradient Boosting)方法分别建立芭蕉地上生物量估算模型;最后,对模型估测结果进行比较分析确定最优模型。结果表明,改进后的YOLOv8s-seg模型可以快速有效地对芭蕉树冠进行检测分割。通过验证发现,基于XGBoost的地上生物量估算模型拟合效果和预测误差优于其他模型,决定系数(R^(2))为0.8814,均方根误差(RMSE)为231.37 kg,平均绝对误差(MAE)为140.47 kg,能够更准确地预测地上生物量,更适于进行芭蕉地上生物量的反演,进一步验证了利用无人机和深度学习方法提取树冠信息实现估算地上生物量的可行性。Aboveground biomass reflects the growth of vegetation and the magnitude of carbon storage,and the accuracy of this parameter is crucial for carbon cycle research and climate change mitigation.In this study,a new idea of using deep learning to realize banana canopy detection segmentation and aboveground biomass estimation was proposed.Firstly,the deep learning algorithm YOLOv8s-seg was used as the basic framework improvement,and UAV remote sensing images were applied to realize banana canopy detection segmentation.Then,the canopy coverage area of banana trees was extracted,combined with the measured aboveground biomass data for fitting,and the aboveground biomass estimation model of banana was established by linear regression,K-Nearest Neighbor(KNN),support vector machine,random forest and XGBoost(eXtreme Gradient Boosting).Finally,the model estimation results were compared and analyzed to determine the optimal model.The results showed that the improved YOLOv8s-seg model can quickly and effectively detect and segment banana canopies.Through verification,it was found that the fitting effect and prediction error of the aboveground biomass estimation model based on XGBoost were better than those of other models,with R2 of 0.8814,root mean square error(RMSE)of 231.37 kg,and mean absolute error(MAE)of 140.47 kg,which could predict the aboveground biomass more accurately and was more suitable for the inversion of the aboveground biomass of bananas,which further verified the feasibility of using UAV and deep learning methods to extract canopy information to estimate the aboveground biomass.

关 键 词：YOLOv8s-seg 无人机遥感 地上生物量 树冠覆盖面积 XGBoost 

分 类 号：S758.5[农业科学—森林经理学]