應用深度學習方法進行UAV影像植被區分類之研究

THE STUDY OF APPLYING DEEP LEARNING TO VEGETATION CLASSIFICATION USING UAV IMAGES

林迪詒(Di-Yi Lin) ； 謝嘉聲(Chia-Sheng Hsieh) ； 翁綺君(Chi-Chun Weng)

《中國土木水利工程學刊》 31卷6期 (2019/10) Pp. 579-588

https://doi.org/10.6652/JoCICHE.201910_31(6).0006

摘要

UAV影像具有高機動性、高解析度之優點，被廣泛應用到許多領域，但拍攝的過程易受到環境影響，因此在進行影像處理時需考量UAV影像圖幅小、影像數量多、傾斜之特點，且UAV影像中常會拍攝到大量的植被，因植被分佈不均和易受時空環境影響具高度變動性，易造成影像匹配錯誤、遮蔽地物等問題，導致植被區產生較大的誤差，因此利用分類方法將植被區框選剔除，可獲得高精度之數值地表模型進行應用。由於近年來熱門發展的深度學習技術具有優越的分類效果，因此本研究嘗試以深度學習方法中Mask-RCNN演算法進行影像分類之探討，但該演算法對於影像邊界處容易因資訊不完整而分類錯誤，因此本研究提出重疊切割正射影像方法，避免影像邊界分類錯誤之問題，另因研究區植被種類與網路常用之公開樣本訓練集有所不同，在此利用遷移學習方法解決研究區少量樣本集的問題。利用本研究所提出的研究流程，經實驗成果顯示整體分類精度提高至96%，能有效框選出植被區。

關鍵字

深度學習 ； UAV影像 ； 植被區

並列摘要

The advantages of capturing images by UAV are high maneuverability and high resolution. Therefore, this technique is widely used in many fields. However, the process of taking pictures is susceptible to environmental influences, so that it is necessary to consider the characteristics of small view frames, large number of images, and oblique shooting when performing image processing. In addition, UAV imagery often captures a lot of vegetation. However, vegetation has non-uniform distribution and high variability that is susceptible to space-time environment, which may cause problems such as image matching errors and shielding of ground objects. It causes a large error in the vegetation area. Therefore, using the classification method to remove the vegetation area, it can obtain the high-precision numerical surface model to do more application. In recent years, the deep learning technology have the superior classification effect, this study attempts to explore the image classification by the Mask-RCNN algorithm in the deep learning method. However, this algorithm is easy to classify wrongly near image border due to incomplete information. Therefore, this study proposes an overlapped orthophotos method to avoid the problem. And, the vegetation types in the study area are different from the public sample training sets in the network. Here the transfer learning method is used to solve the problem of a small sample set in this study area. Using the process proposed by this research, the experimental results show that the overall classification accuracy is improved to 96%, and the vegetation area can be effectively selected.

並列關鍵字

deep learning ； UAV images ； vegetation area