集成U-Net方法的无人机影像胡杨树冠提取和计数

Extraction and Counting of Populus Euphratica Crown Using UAV Images Integrated with U-Net Method

塔里木河流域的胡杨林是该荒漠区域典型的森林资源,胡杨树冠大小和株数信息对塔里木河流域森林资源监测、生态保护和恢复具有重要意义。由于该流域乔灌草植物群落分布的复杂性,传统方法很难实现胡杨树冠的精准分割和大范围的株数提取。以塔里木河中游胡杨林为研究区,选取几块典型胡杨林区域,提出集成深度学习和分水岭分割的处理方法,对密集胡杨树冠的精准分割和单株胡杨的提取进行了深入探讨。首先,将无人机影像(空间分辨率0.16 m)无缝拼接生成正射影像;采用U-Net卷积神经网络对胡杨树冠覆盖区域进行精准分割;在U-Net模型分割的基础上使用标记分水岭方法对密集胡杨树冠进行自动再分割和单株计数,计算出所选研究区的胡杨株数并精准定位。结果表明U-Net卷积神经网络对胡杨的所有树冠区域提取的平均精度可达94.1%,在胡杨树冠覆盖区域识别分割的基础上,用标记分水岭分割方法对胡杨单木计算总体精度为93.3%。研究认为,结合深度学习和标记分水岭方法为自动化大范围森林资源监测提供了新思路和借鉴经验。

The Populus euphratica forest in the Tarim River Basin is a typical forest resource in the desert area.The canopy size and plant number information of Populus euphratica is of great significance for forest resource monitoring,ecological protection and restoration in the Tarim River Basin. Due to the complexity of the distribution of arbor,shrub and grass communities in the area,it is difficult to achieve accurate segmentation of canopy in dense Populus euphratica and large-scale plant number extraction. Taking the Populus euphratica forest in the middle of Tarim River as the research area,several typical Populus euphratica forest areas were selected,and the integrated processing methods of fusion deep learning and watershed segmentation were proposed. The precise segmentation of dense Populus euphratica and the extraction of Populus euphratica were carefully discussed in depth. First,the drone images(spatial resolution 0.16 m)are seamlessly stitched together to generate an orthophoto. Then U-Net convolutional neural network was used to accurately segment the canopy cover area of Populus euphratica. Furthermore,the marker segmentation method was used to automatically re-segment and count the intensive Populus canopy,and the number of Populus euphratica in the selected study area was calculated and accurately positioned. The results show that the average accuracy of the extraction of all canopy regions of Populus euphratica by integrated U-Net convolutional neural network is up to 94.1%. The overall accuracy of the calculation of Populus euphratica by the marker watershed segmentation method is 93.3%. The study suggests that the combination of deep learning and marker watershed methods can provide new ideas and lessons for the automation of large-scale forest resource monitoring.