應用機器學習UAV影像於高美濕地內雲林莞草生長範圍評估

Growth Area Estimation of Bolboschoenus planiculmis in Gaomei Wetland Using Machine-Learning-Based Supervised Classification on UAV Images

鄭尹翔(Yin-Hsiang Cheng)

指導教授 ： 張憲國

國立交通大學/工學院/土木工程系所/碩士(2020年)

摘要

雲林莞草為瀕危物種，於高美濕地具有台灣最大的分布面積。然而，近十多年內因棲地環境變遷與蘆葦及互花米草等植物的入侵，而使其生長範圍逐步外移。有效監測高美濕地內雲林莞草的分布情形是相當重要的學術研究及生態課題。評估往昔徒步實測及應用空拍影像的監督式及非監督式辨識技術的優缺點，而提出以不同取樣方式及機器學習方法，提升影像辨識雲林莞草能力的研究。
本文依據影像辨識物體的紋理特徵，選擇大量樣本以避免主觀的少量選擇樣本造成辨識一致性的差異，並配合Weka開源軟體的應用機器學習進行大量數據的監督式分類。選擇適當視窗大小，測試JpegCoefficientFilter篩選器與SMO分類器可有效的雲林莞草監督式辨識。經測試不同取樣方法發現，矩形框選莞草、沙土及蘆葦三類別的取樣方法具最優辨識能力。
除11月份空拍影像情況外，本法所得各月份雲林莞草生長範圍與步測範圍吻合程度較楊(2018)及楊(2019)具高。本法的相對誤差指標低於15%，覆蓋率指標可達到74%以上，影像辨識正確率可高達85%以上。

關鍵字

影像辨識 ； 機器學習 ； 監督式分類 ； 雲林莞草 ； 高美濕地

並列摘要

Bolboschoenus planiculmis is an endangered species and it has the largest distribution area at the Gaomei Wetland in Taiwan. However, in the past ten years or so, due to the changes in the habitat and the invasion of plants such as Phragmites australis and Spartina alterniflora, its growth area has gradually moved out. Effectively monitoring the distribution of Bolboschoenus planiculmis in Gaomei Wetland is an essential subject in academic research and ecological. After assessing the advantages and disadvantages of foot-measurement and those who used supervised and unsupervised classification on UAV images, we decided to improve the ability of image classification of Bolboschoenus planiculmis by using different sampling methods and machine learning.
To avoid the differences in recognition consistency caused by subjective selection in a small number of samples, this study selects a large number of samples based on the texture characteristics of image recognition objects. Then, we use Weka, an open source software, to perform supervised classification of large amounts of data with machine learning. We choose an appropriate window size and confirm that JpegCoefficientFilter and SMO classifier were effective in supervised identification. After testing different sampling methods, we found that the sampling method which uses rectangular frames to divide all samples into three types, including Bolboschoenus planiculmis, sand and Phragmites australis, has the best identification ability.
Except for the result in November, the relative duplication of the area of Bolboschoenus planiculmis in each month between determined by foot-measurement with that by the present method is higher than those determined by Yang’s (2018) and Yang’s (2019) methods. The relative error of this method is less than 15%, the coverage index can reach more than 74%, and the correct rate of image recognition can reach more than 85%.

並列關鍵字

Image recognition ； Machine Learning ； Supervised Classification ； Bolboschoenus planiculmis ； Gaomei Wetland

參考文獻

1. Butusov, B.O. (2003) Textural classification of forest types from. Mapping science and remote sensing,40(2),91-104.

2. Govender, M., Chetty, K., Naiken, V., and Bulcock, H. (2008) A comparison of satellite hyperspectral and multispectral remote sensing imagery for improved classification and mapping of vegetation. Water SA, 34(2), 147-154.

3. Luca, Z., Thomas, H. (2006) Parallel Software for Training Large Scale Support Vector Machines on Multiprocessor Systems. Journal of Machine Learning Research 7 (2006), 1467–1492.

4. 湯武仁 (2014)利用衛星影像探討高美濕地內雲林莞草生長範圍的變遷研究，國立交通大學土木工程學系所學位論文。

5. 楊承諭 (2018) 高美濕地雲林莞草生長範圍的影像監測及外移現象的研究，國立交通大學土木工程學系所學位論文。