基于无人机多光谱遥感的台风灾后玉米倒伏信息提取

为快速获取台风过后玉米倒伏信息，该研究以生态无人农场大田玉米作为研究对象，利用无人机搭载多光谱相机获取玉米田块图像。采用主成分分析（Principal Component Analysis，PCA）变换多光谱图像，保留信息量最多的前3 个主成分波段；应用最小噪声分离变换（Minimum Noise Fraction Rotation，MNF）对48项纹理特征降维，保留信息量最多的前6项特征；计算选择10种植被指数；对多光谱图像进行低通、高通滤波，将以上特征作为全特征集。使用支持向量机递归（Support Vector Machines-Recursive Feature Elimination，SVM-RFE）、 ReliefF和套索算法（Least Absolute Shrinkage and Selection Operator，Lasso）筛选出3种特征子集，建立5种监督分类模型，对4种数据集进行训练。ReliefF特征子集训练的5种监督分类模型测试集最低分类准确率为89.02%，SVM-RFE和Lasso特征子集训练的5种监督分类模型测试集最低分类准确率均为95.38%，与全特征相比仅相差0.58%，表明通过特征筛选方法可在取得较高分类精度同时大幅减少特征输入数量；运用3种特征筛选方法与不同分类模型的最佳组合提取验证区域玉米倒伏信息，通过混淆矩阵验证结果可知，K最邻近模型结合SVM-RFE特征筛选方法分类精度最高，达93.49%，Kappa系数为0.9，表明了分类模型普适性较强。该研究使用最少特征数量参与分类，且获得最高分类识别精度，可为无人机多光谱技术快速、准确提取台风灾后玉米倒伏信息提供技术支持。

Extraction of maize lodging information after typhoon based on UAV multispectral remote sensing

Abstract: Bending of the lower part of the stalk (lodging) has posed a great threat to the yield, quality, and mechanical harvesting capacity in maize production. It is a high demand to quickly identify the lodging of maize subjected to the large wind load. In this study, an unmanned aerial vehicle (UAV)-based multispectral remote sensing was utilized to extract the maize lodging information after typhoon. A field test was conducted at the ecological unmanned farm of Shandong University of Technology of China. A quadrotor UAV carrying a 6-channel multispectral camera was also used to capture the image of the maize field block. A Pix4Dmapper software was selected to spline the multispectral images, and the band synthesis tool of ENVI software was used to process the six single-band gray images into one image with six bands. Firstly, ten kinds of commonly-used indices of multispectral vegetation were all selected to calculate, where 20 features of near-infrared bands were involved in the classification, due to the sensor included two near-infrared bands (840 and 940 nm). Secondly, a principal component analysis (PCA) was made to transform the original 6-band multispectral image, where the first three principal component bands with the most information were retained to extract texture features. Eight texture features were obtained in each band. The minimum noise fraction rotation (MNF) was applied to reduce the dimensionality of 48 texture features generated by the original 6-band multispectral image, further to screen the first 6 texture features with the most retention information. Finally, a low- and high-pass filtering was used to process the images, where the above 62 features were taken as the full feature set. The numbers of obtained subsets were 10, 13, and 12, respectively, using the support vector machines-recursive feature elimination (SVM-RFE), ReliefF and Least absolute shrinkage and selection operator (Lasso). Five supervised classification models were selected to train the feature subsets of the target region, including SVM, Naive Bayes, K-nearest neighbor (KNN), decision tree, and artificial neural network (ANN). The most suitable classification model for different data sets was selected to classify and evaluate the accuracy of the multi-spectrum of the validation region. The results show that the SVM-RFE and Lasso feature screening effectively reduced the dimension of the data, and the KNN and ANN performed better classification on the two feature subsets screened by the SVM-RFE and Lasso. The overall accuracies of the KNN and ANN model were 90.07%-93.49%, and 89.89%-91.77%, where the Kappa coefficients exceeded 0.85, and 0.86, respectively. There was only a 0.58% difference from the lowest accuracy of the full feature data set, indicating a higher accuracy while a significant reduction in the number of features involved in classification. A confusion matrix verified that the KNN and ANN models can be used to effectively identify the soil background, normal and lying maize, with the highest overall accuracy of 93.49% and 91.77%, respectively, where the Kappa coefficient of above 0.88. The KNN model also presented the best overall classification. Consequently, the least features had participated for the highest classification and recognition accuracy. The finding can provide technical support to the rapid and accurate extraction of maize lodging information after typhoon using the UAV multi-spectral remote sensing.