KMS XINJIANG INSTITUTE OF ECOLOGY AND GEOGRAPHY，CAS
| 塔里木河流域胡杨树冠信息提取和计数研究 | |
| Alternative Title | Extraction and Counting of Poplar Canopy Based on the Combination of Deep Learning and the Watershed Method |
| 李越帅 | |
| Subtype | 硕士 |
| Thesis Advisor | 郑宏伟 |
| 2019-06-30 | |
| Degree Grantor | 中国科学院大学 |
| Place of Conferral | 北京 |
| Degree Discipline | 理学硕士 |
| Keyword | 无人机影像 胡杨 树冠 深度学习 分水岭分割 计数 UAV image Populus euphratica Cown Deep learning Wtershed Sgmentation Counting |
| Abstract | 森林生态系统不仅为人类提供了宝贵的物质和精神财富， 而且与人类命运息息相关， 森林生态系统的稳定是确保该区域生物多样性的前提和关键。 新疆处于内陆干旱区，其森林生态系统稳定性差，很容易遭到破坏。 塔里木河流域的胡杨林是该荒漠地区弥足珍贵的森林资源，对生态系统的稳定及荒漠的绿化起了非常重要的作用， 加强对胡杨林的监测研究具有非常重要的现实意义。 胡杨具有很顽强的生命力， 但是随着时间的推移和生态环境的改变，加上塔里木河河流改道较为频繁， 会造成各种不同生长状态的胡杨林， 增加了监测识别的不确定性， 导致传统方法很难实现较为精准地对大范围胡杨林进行监测评价和生长状态的综合分析。本研究从胡杨树冠信息提取的角度出发， 选取 2017 年 8 月采集的塔里木河中游地区 0.16 m 高分辨率可见光波段无人机数据，使用专业的无人机影像处理软件， 将拍摄区域 14151 景无人机影像无缝拼接，得到完整的一幅塔里木河中游干流区域无人机影像图， 进一步结合最新的深度学习算法对塔里木河中游胡杨林进行处理和分析。首先使用优化的 U-Net 卷积神经网络模型对胡杨树冠覆盖区域进行精准的提取，并与传统的 SVM（支持向量机）和面向对象方法做对比分析；然后将获取的树冠覆盖范围作为掩膜提取出胡杨树冠，利用树冠纹理、 大小、 形状、 灰度值等特征信息，用改进的标记分水岭方法对密集区胡杨树冠进行单木树冠分割； 最后使用该研究方法对整个塔里木河中游研究区的胡杨林进行处理分析，计算整个研究区胡杨林的林分参数信息，并分析胡杨的空间分布状况及影响因素。本研究在传统研究方法的基础上， 结合深度学习方法实现了无人机影像处理的技术流程， 为胡杨林及其他森林生态系统的监测和管理带来了新的技术手段。 通过对塔里木河中游胡杨林的信息提取和分析， 本研究得到主要结论如下：（1） 使用深度学习 U-Net 模型可以对胡杨树冠进行精准的提取，提取树冠的总体精度最高可达 95.6%，相对于 SVM 和面向对象模型平均总体精度分别提高 3.5%和 7%。 实验表明深度学习方法对胡杨树冠提取精度优于传统方法， 在遥感影像分类和信息提取方向具有较强的优势。（2）使用标记分水岭分割方法对胡杨树冠做单木分割， 实验结果表明胡杨单木树冠分割总体精度为 93.3%， 证明了集成应用深度学习和分水岭分割的方法对胡杨树冠提取和株数计算的可行性，本研究方法相较于同类文章具有更高的精度，并且实用性较强。（3）使用本研究方法对塔里木河中上游 32 km2 区域做测试，结果得到胡杨总株数为 620392 株， 该区域树冠郁闭度为 15.457%， 平均单株胡杨覆盖面积约为 7.88 m2，平均冠幅为 3.2 m。 该实验方案可以有效地提高树冠提取精度， 进一步自动化提取林木冠幅、郁闭度、林分密度等参数，提高林业资源调查效率，并且节省大量人工成本。（4） 将塔里木河中游按照河流走向分为 70 个小区展开研究， 发现胡杨林空间分布规律主要是受自然和人为因素影响。 在整体上受自然因素影响，越接近下游胡杨分布越少；局部区域受人为影响因素较大， 大面积开垦为农田的区域胡杨林面积及株数呈现断崖式儿下降。（5） 研究区总面积为 1173.26 km2， 胡杨树冠覆盖总面积为 67.43 km2， 塔河中游胡杨林整体郁闭度为 5.643%， 该研究区胡杨总株数为 8585640 株。 该研究方法可应用于其他大范围的林业资源调查，为推动精准林业的发展提供新思路和借鉴经验。 |
| Other Abstract | Forest ecosystems not only provide valuable material and spiritual wealth to humanbeings, but are also closely related to human destiny. The stability of forest ecosystemsis the prerequisite and key to ensuring biodiversity in the region. Xinjiang is a domesticarid area, its forest ecosystem is poorly stable and easily destroyable. The Populuseuphratica forest in the Tarim River Basin is a valuable forest resource in the desertarea, which plays an important role in ecosystem stability and desert gardening.Strengthening the monitoring and research of Populus euphratica forest has veryimportant practical significance. Populus euphratica has a very strong vitality, but withthe passage of time and changes in the ecological environment, and the frequentrerouting of the Tarim River, it will cause various growth states of Populus euphraticaforest. It is difficult to achieve accurate monitoring and analysis of large-scale Populuseuphratica forests by traditional methods.In this study, the high-resolution UAV data of 0.16m in the middle reaches of theTarim River in August 2017 was selected to study the canopy information extraction ofPopulus euphratica. Using professional UAV image processing software, the 14,151landscape UAV image of the pilot area is spliced into a complete image, and furthercombined with the latest deep learning algorithm for processing and analysis. Firstly,the U-Net convolutional neural network model is used to extract the canopy coveragearea of Populus euphratica and compared with traditional SVM (Support VectorMachine) and object-oriented methods. Secondly, the canopy crown was extracted fromthe canopy and the canopy was divided into single tree canopy by the watershed methodusing the feature information such as the texture, size, shape and gray value of thecanopy. Finally, the research method was used to treat the Populus euphratica forest inthe middle study area of Tarim River. The parameter information of Populus euphraticaforest was calculated and the spatial distribution and influencing factors of Populuseuphratica were analyzed. Based on the traditional research methods, this studycombines the deep learning method to realize the technical process of UAV image processing. It brings new technical means for the monitoring and management ofPopulus euphratica and other forest ecosystems. Through the information extractionand analysis of Populus euphratica forest in the middle reaches the Tarim River, themain conclusions of this study are as follows:(1) Using the deep learning U-Net model, the Populus canopy can be accuratelyextracted with an overall accuracy of 95.6%, which is 3.5% and 7% higher than SVMand object-oriented models, respectively. Experiments show that the deep learningmethod has higher precision in extracting poplar canopy than traditional methods andhas strong advantages in remote sensing image classification and information extraction.(2) The single-wood segmentation of Populus euphratica was carried out by usingthe marker watershed segmentation method. The experimental results show that theoverall accuracy of single-wood canopy segmentation of Populus euphratica is 93.3%.The feasibility of integrating deep learning and watershed segmentation methods forextracting canopy and count the number of plants of Populus euphratica wasdemonstrated. This research method has higher precision and more practicality thansimilar articles.(3) This study was used to verify the area of 32 km2 of the Tarim River. The resultsshowed that the total amount of Populus euphratica was 620,392, and the degree ofcanopy closure was 15.457%. The average coverage of Populus euphratica was 7.88m2 and the average crown was 3.2 m. The experimental scheme can effectively improvethe crown extraction precision, and further automatically extract the parameters such astree crown width, canopy density and stand density. It can also improve the efficiencyof forest resource surveys and save a lot of effort.(4) The middle reaches of the Tarim River were divided into 70 plots accordingto the river trend. It was found that the spatial distribution of Populus euphratica forestwas mainly affected by natural and human factors. In general, under the influence ofnatural factors, the closer to the downstream, the less the distribution of Populuseuphratica; the local area is influenced by human factors, area and the number of plantsof Populus euphratica forest with a large area of open area are significantly reduced.(5) The total area of the study area is 1,173.26 km2, the total area of Populus euphratica is 67.43 km2, the overall canopy degree of Populus euphratica forest is5.643%, and the total amount of Populus euphratica in the study area is 8,585,640. Theresearch method can be applied to other large-scale forest resource surveys, providingnew ideas and lessons for promoting the development of precision forestry. |
| Subject Area | 地图学与地理信息系统 |
| Language | 中文 |
| Document Type | 学位论文 |
| Identifier | http://ir.xjlas.org/handle/365004/15283 |
| Collection | 中国科学院新疆生态与地理研究所 研究系统 |
| Affiliation | 中国科学院新疆生态与地理研究所 |
| First Author Affilication | 中国科学院新疆生态与地理研究所 |
| Recommended Citation GB/T 7714 | 李越帅. 塔里木河流域胡杨树冠信息提取和计数研究[D]. 北京. 中国科学院大学,2019. |
| Files in This Item: | There are no files associated with this item. | |||||
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.
Edit Comment