基于Random Forest的水稻细菌性条斑病识别方法研究

为了快速、准确、有效地识别发病早期的细菌性条斑病,提出基于随机森林(Random forest,RF)算法的水稻细菌性条斑病识别方法,利用光谱成像技术获取该病害的高光谱数据,通过多元散射校正减少和消除噪声及基线漂移对光谱数据的不利影响。利用随机森林特征重要性指标,选取逻辑回归(LR)、朴素贝叶斯(NB)、决策树(DT)、支持向量分类机(SVC)、k最近邻(KNN)和梯度提升决策树(Gradient boosting decision tree,GBDT)算法进行对比试验。同时筛选出12个位于450~664 nm范围内对识别模型有重要影响的光谱波段,并与全波段进行分类结果比较。试验结果表明:RF算法的分类准确率为95.24%,与试验选取的其他算法相比,效果最优,比NB准确率提高了20.97个百分点;与全波段分类结果相比,利用RF算法基于12个波长的识别,波长数减少了98.05%,识别精确率为94.66%,召回率为99.55%,F1值为97.04%,准确率为94.32%。虽然精确率减少了2.97个百分点、准确率减少了0.85个百分点,但召回率增加了4.4个百分点、F1值增加了0.67个百分点,模型精度满足要求。

Identification Method of Rice Bacterial Leaf Streak Based on Random Forest

With the rapid development of rice phenotypic research, rice disease research has also made great progress as an important part of rice phenotypic research. In order to identify bacterial stripe disease quickly, accurately and effectively in the early stages of disease, a method for identifying bacterial stripe of rice based on a random forest algorithm was proposed. The spectral imaging technology was used to obtain hyperspectral data of the disease, and multiple noise correction was used to reduce and eliminate noise and the adverse effects of baseline drift on spectral data. Using the importance index of random forest characteristics, the logistic regression, naive Bayes, decision tree, support vector classifier, k-nearest neighbor and gradient boosting decision tree algorithms were selected for comparative test. At the same time, totally 12 spectral bands which were located in 450~664nm had an important influence on the recognition model were screened out. The results of classification based on the whole band and the 12 important bands were compared. The experimental results showed that the classification accuracy of RF algorithm was 95.24% compared with other algorithms selected in the experiment, the accuracy was higher than that of NB algorithm by 20.97 percentage points. Compared with the whole band classification results, based on these 12 important bands, the number of bands was reduced by 98.05%, the recognition accuracy was 94.66%, the recall rate was 99.55%, the F1 value was 97.04%, and the accuracy rate was 94.32%. Although the accuracy was reduced by 2.97 percentage points, the accuracy rate was reduced by 0.85 percentage points, the recall rate was increased by 4.4 percentage points, the F1 value was increased by 0.67 percentage points, and the model accuracy was basically maintained. Although the accuracy was reduced, the model structure was more streamlined and the computational complexity was reduced. The research result showed that important bands can be used instead of full bands to identify rice bacterial streak disease, which provided new ideas for the identification method of rice bacterial streak disease.