基于机器学习算法的天祝藏族自治县草地地上生物量反演

使用机器学习算法快速、准确、大范围监测草地地上生物量（AGB）是目前研究热点，但不同机器学习算法因训练样本、超参数设置不同而存在较大差异。基于实测草地AGB和同期遥感数据、气象数据、地形数据，选择与草地AGB相关性较强的13个因子作为深度神经网络 （DNN）、随机森林算法（RF）、梯度提升回归树（GBRT）、支持向量机（SVR）、人工神经网络（ANN）和高斯过程回归（GPR）算法的输入变量，建立草地AGB预测模型并从模型预测精度、稳定性、样本敏感性等方面综合评价6种模型应用潜力，分析2020年天祝藏族自治县生长季（4-9月）内草地AGB时空变化特征及其对气候的响应。结果表明：1）DNN估算草地AGB的综合性能最佳，但稳定性较差，对样本敏感性较高；GPR综合性能次于DNN，稳定性和精度均较好；GBRT、RF模拟精度较高，稳定性差；SVR和ANN精度相对其他模型较差，SVR稳定性较高，ANN稳定性较差。2）天祝藏族自治县草地AGB集中在50~250 g·m^-2，不同月份草地AGB空间异质性较大，整体表现为从西北向东南呈下降趋势。3）山地草甸、高寒草甸和温性草原中的AGB变化与气温表现出较为明显的正相关关系。降水量对高寒草甸、温性草原和山地草甸的影响不明显，但对温性荒漠草原类的影响较大，AGB随降水量减少呈现减少态势。以上研究结果可为监测草地生物量的方法选择和参数设置提供一定技术支持和参考依据。

Inversion of grassland aboveground biomass in Tianzhu Zangzu Autonomous County based on a machine learning algorithm

Effective， accurate， and large-scale monitoring of grassland aboveground biomass （AGB） using machine learning algorithms is currently a very active field of research， but different machine learning algorithms vary greatly in performance depending on training samples and hyper-parameter settings. The research utilized grassland AGB data collected field， combined with remote sensing data， meteorological data， terrain data for the same period. Thirteen indicators with strong correlation with grassland AGB were selected as input variables for analysis using deep neural network （DNN）， random forest （RF）， gradient boosting regression tree （GBRT）， vector support machine （SVR）， artificial neural network （ANN） and Gaussian process regression （GPR） algorithms for AGB inversion. At the same time， the application potential of 6 models were evaluated from the aspects of model prediction accuracy， stability， sample sensitivity， and characteristics of AGB space-time changes in grasslands during the 2020 Tianzhu County growth season （April-September） and their response to climate. It was found that： 1） The multivariate performance of DNN in grassland AGB inversion was the best， but the stability was poor， and the sensitivity to samples was high； the comprehensive performance of GPR was inferior to DNN， and its stability and accuracy were good； the simulation accuracy of GBRT and RF was high， and its stability was poor； the accuracy of SVR and ANN was relatively poor and the stability of SVR was high， and the stability of ANN was poor. 2） The grassland AGB ranged from 50 to 250 g·m^-2. The spatial heterogeneity of AGB was large and variable over time. In general AGB showed a downward trend from northwest to southeast. 3） Except for temperate desert steppe， the AGB in mountain meadow， alpine meadow and temperate steppe species showed a significant positive correlation with air temperature. Precipitation had no obvious effect on AGB of alpine meadow， temperate grassland and mountain meadow， but had a great effect on AGB of temperate desert grassland. With decrease in precipitation， AGB tended to decrease. The above results provide technical information to support decisions on choice of method and parameter setting when remotely monitoring grassland biomass.