基于随机森林算法的冬小麦叶面积指数遥感反演研究

【目的】通过利用随机森林算法(random forest,RF)反演冬小麦叶面积指数(leaf area index,LAI),及时、准确地监测冬小麦长势状况,为作物田间管理和产量估测等提供科学依据。【方法】本研究依据冬小麦拔节期、挑旗期、开花期及灌浆期地面观测数据,将相关系数分析(correlation coefficient,r)和袋外数据(out-of-bag data,OOB)重要性分析与随机森林算法(random forest,RF)相结合,在优选光谱指数和确定最佳自变量个数的基础上,构建了两种冬小麦LAI反演模型|r|-RF和OOB-RF,并利用独立数据集对两种模型进行验证;然后,将所建LAI反演模型用于无人机高光谱影像,进一步检验所建模型对无人机低空遥感平台的适用性和可靠性。【结果】|r|-RF和OOB-RF反演模型分别采用相关性前5强、重要性前2强的光谱指数作为输入因子时精度最优,验证决定系数(R~2)分别为0.805、0.899,均方根误差(RMSE)分别为0.431、0.307,表明这两个模型均能对作物LAI进行精确反演,其中OOB-RF模型的反演效果更好。利用无人机高光谱影像数据结合OOB-RF估算模型反演得到冬小麦LAI与地面实测值的拟合方程的决定系数R~2为0.761,RMSE为0.320,数值范围(1.02—6.41)与地面实测(1.29—6.81)亦比较吻合。【结论】本文基于地面数据构建的OOB-RF模型不仅具有较高的反演精度,而且适用性强,可用于无人机高光谱遥感平台提取高精度的冬小麦LAI信息。

Remote Sensing Inversion of Leaf Area Index of Winter Wheat Based on Random Forest Algorithm

【Objective】The objective of this study is to quickly and precisely monitor the growth of winter wheat by inversion of leaf area index (LAI) using random forest algorithm. Thus it could provide a guideline in crop management and mitigation, high and stable yield, agricultural insurance claims, etc.【Method】In this study, field data of canopy reflectance and LAI of winter wheat of four critical growth stages (i.e., jointing period, flag leaf period, flowering period and filling period), were collected under different treatments. The correlation coefficient (r) analysis and the importance analysis of out-of-bag data (OOB) were combined with the random forest algorithm (RF) to determine the more suitable spectral indices and the optimal number of variables for inputs, and then two LAI inversion models (|r| -RF, OOB-RF) were constructed and validated with independent data-sets. Further, the proposed LAI inversion model was applied to the (unmanned aerial vehicle) UAV remote sensing platform to evaluate its performance and reliability for monitoring LAI of winter wheat.【Result】The results showed that the best accuracy of |r|-RF and OOB-RF inversion models was achieved when the top five spectral indices in the correlation and the top two spectral indices in the importance were used as input variables, respectively. The coefficients of determination (R²) of |r|-RF and OOB-RF models during LAI validation were 0.805 and 0.899, and the root mean square errors (RMSE) were 0.431 and 0.307, respectively, which indicated that both |r|-RF and OOB-RF models could well estimate LAI of winter wheat, while the accuracy of the latter was much higher. The retrieved LAI from the UAV hyperspectral images using the OOB-RF model was in well agreement with the ground measured values, with R²=0.761, RMSE=0.320, and the range of estimated values (i.e., 1.02-6.41) also consistent with that actually measured (i.e., 1.29-6.81).【Conclusion】The OOB-RF model constructed in this study not only has high retrieval accuracy, but also can be used to extract high-precision winter wheat LAI information from UAV hyperspectral remote sensing platform.