基于U-Net和分水岭算法的无人机单木树冠提取方法

A Method for Individual Tree-Crown Extraction from UAV Remote Sensing Image Based on U-Net and Watershed Algorithm

高分辨率无人机遥感影像单木树冠参数信息提取方法是森林资源精准监测和生态功能评估的重要基础,而自然光照条件下粘连和遮挡单木树冠的准确分割是直接决定单木树冠信息提取精度的关键。针对自然光照条件下山地森林无人机遥感影像中单木树冠相互粘连、遮挡难以分割,以及传统算法泛化能力弱等问题。本研究结合深度学习和标记控制分水岭算法的优点,提出了一种基于U-Net和标记控制分水岭(marker-controlled watershed,MCW)算法(简称U-Net+MCW算法)的山地森林单木树冠提取方法。以新疆山地森林优势树种天山云杉(Picea schrenkiana var.tianschanica)为研究对象,在南山实习林场采集积雪背景下无人机遥感影像作为试验数据,构建了基于深度神经网络U-Net和标记控制分水岭算法的单木树冠提取模型。首先,从无人机遥感影像中选取1 000张训练样本,128张测试样本,并对样本进行标注,通过数据增强将1 000张训练样本扩增为16 000张,按照4∶1分为训练集和验证集,对U-Net模型进行训练,在训练过程中赋予2个或多个树冠间的相邻边界像素较大权重。然后,利用训练好的U-Net模型对测试集样本进行单木树冠提取。最后,在深度神经网络U-Net单木树冠提取的基础上,采用MCW算法对提取结果进行优化,并对单木树冠提取效果进行精度评估。结果表明,U-Net+MCW算法对于单木尺度的F测度为74.04%,比单一使用U-Net模型提高了28.52%,以该方法提取遥感影像中的天山云杉树冠信息为基础,计算其单木树冠面积和冠幅的精度分别为81.05%和89.94%。因此,U-Net+MCW算法能够有效解决自然光照条件下,由于原始图像背景复杂且树冠内部亮度变化不均匀和树冠间粘连、遮挡等因素,导致的单个树冠内、树冠聚集处或连接重叠区域出现的树冠错分割、过分割、合并等问题,是一种低成本、高效率的单木树冠提取方法,能够满足中小尺度山地森林资源调查和监测要求。

Extraction of parameter information for individual tree crown from high-resolution remote sensing images taking by the unmanned aerial vehicle(UAV)is an important foundation for precise monitoring and ecological function evaluation of forest resources.And accurate segmentation of adjoining and occluded tree crowns in natural lighting conditions is the key to directly determining the accuracy for extraction of individual tree crown information.However,in the process of UAV remote sensing images under natural lighting,there are several issues such as the difficulty in separating adjoining and occluded crowns of individual trees and weak generalization ability of traditional algorithms.Aiming at solving these problems,a method of individual tree crown extraction in mountainous forest based on the U-Net and the marker-controlled watershed(MCW)algorithm was proposed in this paper combining the advantages of deep learning and the MCW algorithm(U-Net+MCW algorithm for short).Taking Picea schrenkiana var.tianschanica,the dominant tree species in mountainous forests of Xinjiang as research object,and using the UAV remote sensing images collected in the Nanshan Forest Farm in snow background as experiment data,the network of U-Net was improved and an extraction model of individual tree crown was constructed based on the deep neural network U-Net and the MCW algorithm.First,1000 training samples and 128 testing samples were selected from the UAV remote sensing images and labeled.Through data enhancement,the number of training samples was extended from 1000 to 16000.The samples were then divided into the training set and the validation set in a ratio of 4∶1 to train the U-Net model.In the training process,the pixels along the boundaries of two or multiple tree crowns were endowed with larger weights.Next,the trained UNet model was used to extract the individual tree crown in the testing set samples.Finally,on the basis of individual tree crown extraction by the deep neural network U-Net,the MCW algorithm was adopted to optimize the extraction results and evaluate the precision of individual tree crown extraction.The results showed that the U-Net+MCW algorithm had a F-measure of 74.04%in the individual tree level,which was improved by 28.52%compared to that of pure U-Net model.Based on the tree crown information of P.schrenkiana var.tianschanica extracted from remote sensing images with this method,the accuracy of individual tree crown area and crown wide on the level of individual trees was calculated to be 81.05%and 89.94%,respectively.Therefore,the U-Net+MCW algorithm could effectively solve the problems such as false segmentation,over-segmentation and combining of tree crowns in image backgrounds,individual tree crown and clustering,connecting or overlapping tree crown areas.The U-Net+MCW algorithm is a lowcost and high-efficiency individual tree crown extraction method,which can meet the requirements of the survey and monitoring of small and medium-scale mountainous forest resources.