基于随机森林算法的冬小麦生物量遥感估算模型对比

摘要

为了寻求高效的冬小麦生物量估算方法，该研究获取了2014年陕西省杨凌区拔节期、抽穗期和灌浆期的冬小麦生物量和对应的RADARSAT-2全极化雷达、GF1-WFV多光谱数据，并利用随机森林算法（random forest，RF）将光谱、雷达后向散射、光学植被指数和雷达植被指数结合进行冬小麦生物量回归建模。将相关系数分析（correlation coefficient， r）、袋外数据（out-of-bag data，OOB）重要性和灰色关联分析(grey relational analysis, GRA)与随机森林算法（RF）进行整合，构建了3种冬小麦生物量估算模型：r-RF、OOB-RF和GRA-RF，并分别利用3种估算模型对冬小麦生物量进行了估算。结果表明：r-RF、OOB-RF和GRA-RF3种模型分别采用3、4、10组数据时，验证决定系数分别为0.70、0.70和0.65，平均绝对误差分别为0.162、0.164和0.172 kg/m2，均方根误差分别为0.218、0.221和0.236 kg/m2，r-RF和OOB-RF比GRA-RF对冬小麦生物量有更好而的预测能力。研究结果证实了随机森林算法对冬小麦生物量进行遥感估算的潜力。%Biomass is one of important agricultural crop parameters, and has significant meanings in agriculture production management and decision-making. The estimated of biomass by remote sensing is of great importance for the real-time and dynamic crop information and can be acquired by remote sensing detection technology. The random forest algorithm (RF), from which machine learning is used, shows considerable potential for estimated crop parameters. However, there has been little study on estimation of crop parameters by RF model, especially using a combined of optical and SAR Data. In this study, we focused on analyzing different RF models data selection impact on the accuracy of estimation winter wheat biomass using spectral reflectance, radar backscatter, spectral Vis (Vegetation Indices) and radar Vis. In this paper, RF model, optical and SAR data were used to estimate the biomass of winter wheat. The objective of the study was to demonstrate the feasibility of random forest algorithm for monitoring on winter wheat biomass, meanwhile, the method of remote sensing data selection was compared. In the most important winter wheat producing region in China, Guanzhong plain, field experiments were carried out in Yangling district, Shannxi province. The synchronous RADARSAT-2 SAR data and GF1-WFV multiple spectral data which were close to the sampling time were obtained as the remote sensing data in this experiment. Firstly, the biomass of winter wheat at elongation, heading and filling stage were measured. Remote sensing data were pretreated as spectral reflectance, radar backscatter, spectral Vis and radar Vis. Then, the correlation coefficient analysis (r), the importance of out-of-bag data (OOB) and grey relational analysis (GRA) were used in the study. According to the above three analysis methods to select the data, and the input data were sorted according to the analysis results. Three models served for biomass estimation of winter biomass based on random forest algorithm: r-BF, OOB-RF and GRA-RF. The three models were validated using the in situ measured data, and 17 experiments of each model were designed to verify the accuracy of the model changes. As accurate valuation methods, the determination coefficients (R2), the corresponding mean absolute errors (MAE) and the root mean square errors (RMSE) for estimated biomass were calculated respectively with the measured data. The r-RF (R2=0.70, MAE=0.162 kg/m2, RMSE=0.218 kg/m2) and OOB-RF (R2=0.70, MAE=0.164 kg/m2, RMSE=0.221 kg/m2) models achieved similarly very high accuracy, and the accuracy increased with the increase of the input variables, and then decreased. GRA-RF (R2=0.65, MAE=0.172 kg/m2, RMSE=0.236 kg/m2) model was worse than the previous two, the r-RF and OOB-RF showed a more robust predictive ability than GRA-RF model. Most importantly, the results indicated that it is necessary to select the appropriate data inputs to increase the accuracy of the RF model, rather than the input of many vegetation indices. The potential of random forest algorithm to estimate the biomass of winter wheat was show in this research. Our results indicated that the RF could be used to robustly estimate winter wheat biomass. This study may provide a guideline for improving the estimations of biomass of winter wheat using RF model.