基于无人机遥感可见光影像的农作物分类

CROP CLASSIFICATION BASED ON UAV REMOTE SENSING IMAGES

关键词:  无人机遥感农作物分类支持向量机可见光影像差异数字表面模型（DDSM）

Key words:  UAV  crop classification  SVM  RGB  DDSM

摘要:

无人机遥感具有高空间、高时间分辨率的优点，并可同时获得光谱和空间信息，因此在农作物分类中备受研究者的青睐。与侧重于从高分辨率RGB图像中提取纹理特征的分类方法不同，文章重点研究如何利用作物在光谱和空间维度上的联合特征尤其是作物高程特征，以实现农作物精细分类。［方法］首先，我们进行研究区域选择和地面实际情况调查，用无人机遥感系统进行可见光影像采集；其次，确定研究区域内农作物分类类别，分别对可见光遥感影像进行可见光植被指数计算及纹理滤波；针对数字表面模型（DSM）数据特点，对两期DSM 数据进行差值处理，获得差异数字表面模型数据（DDSM），提取作物高度信息，并根据农作物冠层特性对差异数字表面模型进行滤波处理；最后，进行特征优选及组合，使用SVM 方法进行农作物分类。［结果］确定最优分类特征为RGB、红波段对比度、绿波段二阶矩、蓝波段方差、DDSM、DDSM 方差、DDSM 对比度，分类精度由7186%提高到9230%，验证了由DSM 影像提取的空间特征可以提高农作物分类精度。［结论］该研究探索了一种基于可见光及空间联合特征的农作物精细分类方法，方法简单可行，设备成本低，在基于无人机低空遥感的样方调查领域中有较大的应用前景。

Abstract:

UAV remote sensing has the advantages of high spatial and temporal resolution, and can obtain spectral and spatial information at the same time. Therefore, it has been favored by researchers in crop classification in recent years. Differencing from the classification methods that focuses on extracting texture features from high resolution RGB images, this study focuses on how to use the combined features of spectral and spatial dimensions, especially the elevation features of crops, to achieve fine classification of crops. In this study, UAV mounted visible camera was used to obtain high resolution images of crops in the target area. Combining photogrammetry technology, crop height was reconstructed from UAV images, and height features were used for crop classification. Furthermore, using DSM (Digital Surface Model) data of two periods, the differential digital surface model (DDSM) was generated to highlight the difference characteristics of crop growth, and this feature was applied to crop classification. At the same time, the spectral and spatial characteristics of crops were analyzed and extracted. By comparing and analyzing the coefficient of variation and the coefficient of difference between classes of different characteristics, a SVM classification model based on feature combination was constructed for crop classification. The experimental results showed that the classification accuracy of the proposed method could be improved from 76.00% to 91.90% by introducing crop elevation features and combining them with visible light features. This study explores a fine classification method of crops based on visible light and spatial characteristics. The method is simple and feasible, and the equipment cost is low. It has a great application prospect in the field of sample survey based on UAV low altitude remote sensing.