基于机器学习算法的草地地上生物量估测——以祁连山草地为例

Abstract

PDF HTML阅读 XML下载 导出引用 引用提醒 基于机器学习算法的草地地上生物量估测——以祁连山草地为例 DOI: 10.5846/stxb202203180669 作者: 作者单位: 作者简介: 通讯作者: 中图分类号: 基金项目: 第二次青藏高原综合科学考察研究（2019QZKK0302）；中国科学院战略性先导科技专项（A类）项目（XDA23100201） Estimation of grassland biomass using machine learning methods: A case study of grassland in Qilian Mountains Author: Affiliation: Fund Project: the Second Tibetan Plateau Scientific Expedition and Research Program (grant no. 2019QZKK0302), the Strategic Priority Research Program (A) of the Chinese Academy of Sciences (grant no. XDA23100201) 摘要 | 图/表 | 访问统计 | 参考文献 | 相似文献 | 引证文献 | 资源附件 | 文章评论 摘要:草地地上生物量（Aboveground Biomass，AGB）是指导畜牧业生产管理的重要指标，是草畜平衡综合分析的基础。目前，有关祁连山草地AGB反演的研究较少，且多源数据间的尺度差异问题并未得到很好的解决。为了解祁连山草地AGB的空间分布状况，利用Sentinel-2多光谱数据、无人机（Unmanned Aerial Vehicle，UAV）数据以及2021年植被生长期实测草地AGB数据实现了空天地一体化监测，通过决策树回归（Decision Tree Regression，DTR）、随机森林回归（Random Forest Regression，RFR）、梯度提升决策回归树（Gradient Boosting Regression Tree，GBRT）以及极致梯度提升（eXtreme Gradient Boosting，XGBoost）共4种算法反演草地AGB的适用性分析，利用最优模型反演了祁连山草地的AGB空间分布状况。结果表明：研究区内多种植被指数所表现出的特性有所差异。祁连山地区AGB在空间分布上呈现出由西北向东南递增的趋势，平均AGB为925.43kg/hm2。6种植被指数与实测AGB之间均表现为显著正相关，适合作为祁连山草地AGB遥感反演的指标；XGBoost模型较其它模型具有最高的R2值（0.78）和精度（74.75%）、最低的均方根误差（RMSE，99.74 kg/hm2）和平均绝对误差（MAE，71.60 kg/hm2），模型反演效果最好；UAV数据能够提供更加详细的空间细节特征，减小Sentinel-2数据和实地采样数据间的尺度差异；因此，基于6种植被指数与祁连山草地AGB间的相关性，构建XGBoost模型反演研究区草地AGB空间分布状况是具有实践意义的。研究结果将为指导祁连山草地畜牧业的发展和维护草地生态系统的平衡提供一定的参考价值与数据支撑。 Abstract:Aboveground biomass (AGB) is an important indicator to guide the management of livestock industry, and it is the basis of comprehensive analysis of the balance between grassland and livestock. To date, only few studies have studied the spatial distribution of grassland AGB in Qilian Mountains, and the scale differences of multi-sources data have not been well solved. Therefore, in order to understand the spatial distribution of AGB in Qilian Mountains, we used the space-air-ground integrated method based on Sentinel-2 multispectral data, Unmanned Aerial Vehicle (UAV) data and the measured AGB data during the growth period of vegetation in 2021. In addition, we analyzed the applicability of Decision Tree Regression (DTR), Random Forest Regression (RFR), Gradient Boosting Regression Tree (GBRT) and eXtreme Gradient Boosting (XGBoost) algorithms for AGB inversion. Finally, we mapped the spatial distribution of AGB in the study area using the optimal model among all the models. The results verified that the effectiveness of vegetation indices varied in study area. Generally, the results indicated that the spatial distribution had an increasing trend from northwest to southeast, with an average AGB density of 925.43 kg/hm2. A significantly positive correlation was found between 6 vegetation indices and measured AGB, and both of the indices was suitable for inversion of grassland AGB in the Qilian Mountains. Moreover, compared with other models, the performance of XGBoost model was the best, with the highest R2 of 0.78 and accuracy of 74.75%, the lowest Root Mean Squared Error (RMSE) of 99.74 kg/hm2and Mean Absolute Deviation (MAE) of 71.60 kg/hm2. In addition, UAV data provided spatial characteristics in detail, which reduced the scale difference between Sentinel-2 and the measured data. Therefore, on the basis of the correlation between 6 vegetation indices and AGB, it is of practical significance to construct the XGBoost model to invert the spatial distribution of AGB in grassland of Qilian Mountains. The results can provide a reference value and data support for guiding the development of livestock industry and maintaining the balance of grassland ecosystem. 参考文献 相似文献 引证文献