基于多元HoG及无人机航拍图像的植被类型识别

Vegetation type recognition based on multivariate HoG and aerial image captured by UAV

使用无人机进行低空航拍,快速取得大范围的植被图像,结合多元HoG特征进行植被类型识别。首先,利用Gabor滤波器提取图像的纹理信息,HSV和Lab颜色空间转化提取图像的颜色信息。其次,将图像分割为N个单元格(cell),基于纹理与颜色信息计算每个单元格的方向梯度直方图(HoG)特征,形成多元HoG特征。最后,以单元格为分类单位,结合随机森林机器学习算法,建立植被类型识别模型。以福建省安溪县山区为研究区域,结果表明:利用无人机低空航拍的光学影像结合多元HoG特征进行植被类型识别是可行的;对于植被与非植被识别,其最高分类正确率达到96.04%; 20 m航拍下,植被类型识别率最高,为82.44%,随着航拍高度的升高,模型识别效果呈现下降趋势。进一步采集福建省长汀县山区的植被航拍影像为测试数据,证明模型对于不同地区植被类型识别的稳定性,其识别精度最高可达73.31%,正确率无显著差异。本研究采用无人机载光学相机获取植被光学图像数据,数据获取方便且所需费用较低;提出的植被类型识别模型具有较高的精度;对于不同地区的植被类型识别具有较好的稳健性,可方便应用于野外森林树种监控与管理。根据不同高度模型识别结果,航拍高度不宜过高,航拍高度以20 m为宜。

The paper used the unmanned aerial vehicle(UAV)which can quickly obtain a wide range of vegetation images to capture the low-altitude aerial photography and associated with the multivariate HoG features to recognize vegetation types.Firstly,the texture information of the image were extracted by Gabor filter and using HSV and Lab color space transform to extract the color information.Next,the image were divided into N cells,then computing the gradient of histogram of each cell based on the texture and color information to form the multivariate HoG features.Finally,taking the cell as the classification unit,combining the random forest machine learning algorithm,vegetation types recognition model were established.Taking the Anxi County in Fujian Province as the test area,the results showed that it was feasible to identify the vegetation types by using the optical image of low-altitude aerial photography;the highest recognition accuracy in recognizing vegetation and non-vegetation was 96.04%;the recognition accuracy rate was 82.44%in recognizing vegetation types at 20 m and decreased with the height increase.Moreover,the vegetation images of aerial photography were further collected,which was captured at the Changting County of Fujian Province,as another data to prove the robustness of the model of vegetation type recognition in different areas,the highest recognition accuracy was 73.31%,the accuracy was similar compared with the result of Anxi County.In this paper,the vegetation images by using the unmanned aerial vehicle with optical camera were conveniently,econmically obtained;the proposed vegetation type recognition model can be easily applied to the field vegetation type monitoring and management as the high precision and its robustness in various regions.In addition,according to the recognition result at various aerial height,the proposed aerial height should not be too high,20 m as the reference height of aerial photography was suitable.