應用深度學習技術於UAV影像調查人工林林分分布之初探

森林經營者為求合理的經營決策，須掌握關於森林資源的適當資訊，作為後續疏伐撫育與整體經營規劃等處理之依據，近年來UAV能拍攝高解析度影像成為評估森林狀況和變化的重要工具，然而森林面積大且複雜性高，隨著深度學習在林業應用中變得越來越流行，因它可如同專家等級的人類以自動化的方法分類影像，進而減少大量調查森林的工作時間。本研究以台灣高雄市六龜林試所人工區域為主，該人工林區域主要種植台灣杉，但受到全球氣候變遷和林業政策轉變以保育為重之影響，使得人工林遭受闊葉林入侵，為了解現今人工林中的林分分布狀況，研究中使用UAV拍攝影像快速獲取現況資訊，並透過深度學習的卷積神經網路（Convolutional Neural Networks, CNN）之Inception模型作為分類影像的神經模型架構。為了讓UAV影像利於結合深度學習技術調查林分之效果，嘗試不同解析度的UAV影像評估分類針葉樹、闊葉樹區域之正確性，根據實驗成果顯示影像對地解析度為1、2、4公分時分類針葉樹的精度為90%，但闊葉樹的分類精度較低（34-75%），經紋理分析發現實驗區屬於人工林為單一樹種台灣杉（針葉樹），所以其紋理的特徵較為相似，另外闊葉樹的紋理不一致性，顯示該區域的闊葉樹品種多樣性，因此造成闊葉樹的分類成果較差，所以本研究藉由資料增強方法改善闊葉樹分類的精度。另外8公分解析度的影像整體分類精度為72%，該整體分類精度相較小於4公分解析度影像有所落差。

關鍵字

林分調查；深度學習； UAV影像

並列摘要

Forest managers must have appropriate information about forest resources in order to make the correct decisions. In recent years, Unmanned Aerial Vehicles (UAV) are becoming an essential tool for evaluating the status and changes in forest ecosystems. However, the forest area is often large and its complexity is high, thus deep learning has become popular for forestry applications. This is because deep learning allows the inclusion of human knowledge into the automatic image processing pipeline. Thus, one can expect that the time required to investigate forests can be greatly reduced. Our study area is located in the planted forest of the Liuguilin Forestry Research Institute in Kaohsiung City, Taiwan. The planted forest area is mainly planted with Taiwania cryptomerioides. In recent years, climate change is bringing about positive changes as well as adjustments in forestry policies to the conservation of trees, thus the planted forest has been invaded by broad-leafed trees. Therefore, in order to obtain the distribution of forest stands given the current conditions, we propose that UAV be used to capture high-resolution images and then use the inception model of deep learning Convolutional Neural Networks (CNN) to classify the images. In this paper deep learning technology is used to investigate forest stand results related to the use of different UAV image resolutions. The results show that the accuracy is 90% when the image resolution is 1, 2, 4 cm, but the classification accuracy of broad-leafed trees is low (34-75%). This is because the study area is predominantly a planted forest of coniferous species that are relatively single and the broad-leafed trees are the most complex, thus the classification results of the broad-leafed trees are bad. Additionally, the overall classification accuracy of an 8 cm resolution image is 72% and its results are different from the overall classification accuracy of images with a resolution of less than 4cm.

並列關鍵字

investigate planted forest ； deep learning ； UAV image

參考文獻

Balenović, I., A. S. Milas, and H. Marjanović. 2017. “A Comparison of Stand-Level Volume Estimates from Image-Based Canopy Height Models of Different Spatial Resolutions.” Remote Sensing 9 (3): 205-232. doi:10.3390/rs9030205

Berni, J. A. J., P. J. Zarco-Tejada, L. Suárez, and E. Fereres. 2009. “Thermal and Narrowband Multispectral Remote Sensing for Vegetation Monitoring from Unmanned Aerial Vehicle.” IEEE Transactions on Geoscience and Remote Sensing 47 (3): 722-738. doi: 10.1109/TGRS.2008.2010457

Maza, I., F. Caballero, J. Capitán, J. R. Martínez-de-Dios, and A. Ollero. 2011. “Experimental Results in Multi-UAV Coordination for Disaster Management and Civil Security Applications.” Journal of Intelligent & Robotic Systems 61 (1): 563-583. doi: 10.1007/s10846-010-9497-5

Zhao, W., and S. H. Du. 2016. “Learning Multiscale and Deep Representations for Classifying Remotely Sensed Imagery.” ISPRS Journal of Photogrammetry and Remote Sensing 113: 155-165. doi: 10.1016/j.isprsjprs.2016.01.004

Krizhevsky, A., I. Sutskever, & G. E. Hinton. 2017. “Imagenet Classification with Deep Convolutional Neural Networks.” Communications of the ACM 60 (6): 84-90. doi: 10.1145/3065386

國際替代計量

應用深度學習技術於UAV影像調查人工林林分分布之初探

全文下載

主題瀏覽