以卷積神經網路對多分類項目進行辨識研究

The study of multi-category image analysis by using convolutional neural network

彭信瑋(Hsin-Wei Peng)

指導教授 ： 萬絢

嶺東科技大學/資訊學院/資訊管理系碩士班/碩士(2022年)

摘要

集水區水源保育範圍可能因違規開發或颱風豪雨造成坡地崩塌及土石滑動等土砂災害，常直接或間接影響水源水質或造成嚴重水庫淤積，威脅水資源之運用與永續，目前集水區內土地之管理，採人力定時巡查有無違規利用情形，以維護集水區之生態保育，因此本研究與影像辨識技術結合，透過該技術可以有效率化的進行土地管理，大大節省人力的支出；本研究使用了位於烏來山區的影像數據，這個區域的影像數據有許多月份，而從中取了兩個月份來進行分析研究，波段方面採用了基本的4波段，地貌方面採用了樹木、草皮、建物等7個分類項目，分類器方面則使用機器學習中最常見的支援向量機(Support Vector Machine, SVM)與深度學習中的卷積神經網路(Convolution Neural Networks, CNN)來進行資料的訓練與測試，CNN模型是參考AlexNet模型，此模型使用了5層的卷積層與3層池化層來當作整體的架構，而最後的結果會透過誤差矩陣與主題圖進行比對，比較兩個分類器兩者的差異。

關鍵字

機器學習 ； 深度學習 ； 誤差矩陣

並列摘要

The water conservation area in the watershed area may be caused by illegal development or typhoon heavy rains, which may cause soil and sand disasters such as landslide and soil and rock sliding. It often directly or indirectly influence the water quality of the water source or cause serious reservoir siltation which threaten the use and sustainability of water resources. The current water catchment area for the management of area relied on manpower to regularly observed whether there is any illegal use in order to maintain the ecological conservation of the watershed area. Therefore, this study is combined with the image recognition technology, through which the land management can be efficiently carried out, and the labor expenditure can be greatly saved. This study uses image data located at the Wulai Mountains. There are many months of image data in this area, and two months are taken from them for analysis. The band adopts the basic 4 bands, and the landform adopts 7 classification categories such as trees, turf, and buildings. The first step uses the Support Vector Machine (SVM) in machine learning. The Convolution Neural Networks (CNN) is a deep learning approach for the second step of data training. For testing, the CNN model refers to the AlexNet model. This model uses 5 layers of convolution layers and 3 layers of pooling layers as the overall architecture. The final result will be compared with the theme map through the error matrix, and the two classifiers will be compared. The difference between CNN and SVM is drawn by confusion matrix and thematic map.

並列關鍵字

Machine learning ； Deep learning ； Confusion matrix

參考文獻

邱盛鴻（2017）。地方審計人員應用GIS情形與影響因素之研究。東海大學公共事務碩士在職專班碩士論文，台中市。

魏振庭（2015）。基於領土防禦之改良式獅子演算法應用於支援向量機分類器之特徵選取。國立中興大學資訊管理學系所碩士論文，台中市。

馬弘霖（2019）。都市高光譜影像地貌辨識- 支持向量機、隨機森林和類神經網路之比較。嶺東科技大學資訊管理系碩士班碩士論文，台中市。

賴之康（2020）。運用卷積神經網路偵測網站頁面異常研究。國立中央大學資訊管理學系碩士論文，桃園縣。

繆忠錡（2016）。頻帶維度縮減和部分濾波法於高光譜影像目標偵測。國立臺灣海洋大學通訊與導航工程學系碩士論文，基隆市。