基于GF-1遥感影像和relief-mRMR-GASVM模型的小麦白粉病监测

选择合适的建模和特征选择算法对提高作物病害的遥感监测水平有着重要的作用。研究以河北省小麦白粉病为研究对象,基于GF-1/WFV数据共提取了4个波段反射率数据和10个对作物长势和胁迫敏感的植被指数作为初选特征。针对常用的特征提取算法relief算法筛选出的特征存在冗余性的问题,提出了一种relief结合最小冗余最大相关(minimum redundancy maximum relevance,m RMR)的特征降维算法(relief-m RMR)。首先,通过relief算法计算出各特征的权重系数,对特征集进行加权;然后利用m RMR算法选出与类别具有最小冗余性的特征,利用支持向量机(support vector machine,SVM)对河北白粉病进行监测,并用遗传算法(genetic algorithm,GA)优化的SVM(GASVM)建立了白粉病的监测模型(relief-m RMR-GASVM),将监测结果分别与SVM和网格寻优(grid search,GS)算法优化的SVM(GSSVM)的监测结果进行对比分析,同时比较了该方法与Ada Boost、粒子群(Pso)优化的最小二乘支持向量机(least squares support vector machine,Pso-LSSVM)和随机森林(random forest,RF)3种方法的优越性。结果表明,relief-m RMR算法筛选出的特征与GASVM、SVM和GSSVM建立的监测模型精度比传统relief算法筛选特征所建模型的精度分别提高了14.3个百分点、7.2个百分点和7.1个百分点,比传统m RMR算法筛选特征所建模型的精度分别提高了14.3个百分点、14.3个百分点和14.2个百分点。relief-m RMR算法结合GASVM建立的监测模型精度为所有模型中最高,精度为85.7个百分点,分别比SVM和GSSVM所建监测模型精度提高了21.4个百分点和7.2个百分点。此外,GF-1数据结合relief-m RMR-GASVM模型的监测精度分别高出Ada Boost、Pso-LSSVM和RF方法21.4个百分点、14.3个百分点和7.1个百分点。说明GF-1数据结合relief-m RMR-GASVM模型可用于小麦白粉病的遥感监测。

Wheat powdery mildew monitoring based on GF-1 remote sensing image and relief-mRMR-GASVM model

Abstract: Wheat powdery mildew is one of the most serious diseases in wheat, so timely and effective monitoring of wheat powdery mildew is necessary for improving wheat yield and quality. The selection of suitable modeling method and feature selection algorithm play an important role in improving the performance of remote sensing monitoring of crop diseases. In this study, the GF-1 remote sensing image was used to extract 14 characteristic variables, including 4 band reflectance variables and 10 vegetation indices. An approach combining relief and minimum redundancy maximum relevance (mRMR) algorithm (relief+mRMR) is proposed for improving the ability to remove redundancy of relief algorithm. First, the relief algorithm was used to calculate the weight of each feature and filter out the disease independent features. Then the mRMR algorithm was used to remove the redundant features. Finally, the optimal feature set including NIR(near-infrared reflectance), SR (simple ratio index) and NDVI (normalized difference vegetation index) was as the input variables. Meanwhile, the other 2 feature sets such as SR, GNDVI (green normalized difference vegetation index) and TVI (triangular vegetation index) by relief algorithm and TVI, RTVI (ratio triangular vegetation index) and RDVI (re-normalized difference vegetation index) by mRMR algorithm were also obtained. Support vector machine (SVM) is a machine learning method based on statistical learning theory, and its working principle is to minimize the structural risk as the core, and it can improve the generalization ability, and resolve problems such as nonlinearity, small samples. Penalty factor and nuclear parameter should be considered in establishing monitoring model for wheat powdery mildew, and traditional parameter selection was mostly through multiple experiments. Presently, the commonly used grid search (GS) algorithm obtains the optimal parameters, but its efficiency is low and workload is large. The advantage of the genetic algorithm (GA) is to solve the global optimal problem, which is robust and can be independent of the domain of the problem when searching quickly. Therefore, the algorithm of SVM optimized by GA (GASVM) was used to monitor wheat powdery mildew in Hebei, China. For comparison and validation, SVM method, the approach of SVM optimized by GS algorithm (GSSVM), and 3 existing monitoring methods of wheat powdery mildew i.e. AdaBoost, particle group optimized least square support vector machine (PSO-LSSVM) and random forest (RF) were also been used. The results illustrated that the performance of models constructed using the feature set through relief+mRMR algorithm outperformed the models established using the feature set through only relief algorithm or only mRMR algorithm. The result demonstrated that the combination of relief and mRMR algorithms can effectively remove the redundancy between features while selecting high correlated features with disease. Additionally, in 3 monitoring models based on the features selected by relief+mRMR, the relief-mRMR-GASVM monitoring model had the highest overall accuracy of 85.7%, which increased by 21.4 and 7.2 percentage points compared with relief-mRMR-SVM and relief-mRMR-GSSVM monitoring models. The result indicated that the relief-mRMR-GASVM approach can effectively improve the monitoring accuracy and consistency of the wheat powdery mildew model, and further strengthen the reliability of the model in practical application. Furthermore, the monitoring accuracy of GF-1 data combined with relief-MRMR-GASVM model increased by 21.4, 14.3 and 7.1 percentage points compared with AdaBoost, PSO-LSSVM and RF methods, respectively. These results reveal that the GF-1 data combined with the relief-mRMR-GASVM model can be used for remote sensing monitoring of wheat powdery mildew.