基于随机森林回归算法的小麦叶片SPAD值遥感估算

Estimation of Wheat Leaf SPAD Value Using RF Algorithmic Model and Remote Sensing Data

摘要:

使用机器学习中的随机森林(RF)回归算法构建小麦叶片SPAD值遥感反演模型。以2010—2013年江苏地区试验点稻茬小麦3个生育期（拔节、孕穗、开花）的叶片为材料，结合我国自主研发的环境减灾卫星HJ-1对研究区域进行同步监测，分析了各生育期叶片SPAD值与8种植被指数间的相关性；以0.01水平下显著相关的植被指数作为输入参数，使用RF回归算法构建了每个生育期的小麦SPAD反演算法模型，即RF-SPAD模型，以支持向量回归(SVR)和反向传播(BP)神经网络算法构建的SVR-SPAD模型和BP-SPAD模型作为比较模型，以R2和均方根误差(RMSE)为指标，分析了每个生育期3个模型的学习能力和回归预测能力，结果表明：RF-SPAD模型在3个生育期都表现出最强的学习能力，R2和RMSE在拔节期分别为0.89和1.54，孕穗期分别为0.85和1.49，开花期分别为0.80和1.71；RF-SPAD模型在3个生育期的回归预测能力都高于BP-SPAD模型，高于或接近于SVR-SPAD模型，R2和RMSE在拔节期分别为0.55和2.11，孕穗期分别为0.72和2.20，开花期分别为0.60和3.16。

Abstract:

As one of the machine learning algorithms, random forest (RF) regression was proposed firstly to construct remote sensing monitoring model to inverse leaf SPAD value in different growth stages of wheat. The experiment was carried out during 2010—2013 in Jiangsu province. Based on the wheat leaves and synchronous China’s domestic HJ-CCD multi-spectral data in the jointing stage, the booting stage and the anthesis stage respectively, the relationships between SPAD and eight vegetation indices were analyzed at corresponding period. According to the selected vegetation indices which were significantly related to the leaf SPAD value in the 0.01 level, the model for estimating leaf SPAD value at each period was built by using RF algorithm, namely the RF-SPAD model. At the corresponding period, SVR-SPAD model which was based on the support vector regression (SVR) and BP-SPAD model which was based on the back propagation (BP) neural network were constructed as compared models. SVR and BP neural network were both machine learning algorithms. Based on R2 and RMSE, the learning abilities and generalization abilities of three models at each period were analyzed. The results showed that the RF-SPAD model at three stages presented the strongest learning ability, which its R2 was the highest as well as RMSE was the lowest, concretely, R2 and RMSE were 0.89 and 1.54 in jointing stage, 0.85 and 1.49 in booting stage and 0.80 and 1.71 in anthesis stage respectively. RF-SPAD model’s prediction ability was equal to or higher than the reference models which R2 and RMSE were 0.55 and 2.11 in jointing stage, 0.72 and 2.20 in booting stage, 0.60 and 3.16 in anthesis stage respectively.