高光譜影像對水稻田影像判釋之研究:以倒傳遞類神經網路與細菌覓食演算法比較之實證

高光譜是一種先進的影像材料，這種材料資訊豐富的影像在判釋上亦需要花費較多的時間，若能建構一決策系統進行分析，準確的判讀出地表上所代表的物種，就能大幅減少實地探勘的人力、物力與時間。本研究主要探討如何從高光譜影像中篩選出重要的光譜資訊，並以水稻田為主要判釋對象，搭配監督式及非監督式學習的分類器進行判釋。本研究擬倒傳遞類神經網路與細菌覓食演算法對於高光譜影像進行影像判釋，先以亂度基礎分類法進行影像屬性之篩選，再將上述兩種演算法進行計算，進而設計以下四種研究案例: (a)原始波段搭配倒傳遞類神經網路 (b)亂度基礎分類法篩選出重要光譜資訊搭配倒傳遞類神經網路 (c) 原始波段搭配細菌覓食演算法(d) 亂度基礎分類法篩選出重要光譜資訊搭配細菌覓食演算法，最後使用誤差矩陣表以及主題圖呈現出分類後之成果進行比較。

關鍵字

高光譜影像；亂度基礎分類法；倒傳遞類神經網路；細菌覓食演算法

並列摘要

The hyperspectral is an advanced material which render more accurately classification result. However, due to spectral information is very rich, the computation also takes more time. Therefore, if we can construct a decision system to accurately interpreting the surface of ground, it can be applied significantly to reduce manpower and time. This study focused on how to extract the important factors of hyperspectral image to classify the paddy rice area with applying supervised and unsupervised learning algorithm. In this study, back-propagation neural network and bacterial foraging optimization for hyperspectral image for image classification. The prior processing of image data used entropy-based classification to extract the influenced factors of image band properties. Then the two algorithms are applied into following four case studies: (a) the original band with a back-propagation neural network (b) entropy-base-classification filter out important information with back-propagation neural network (c) the original band with bacterial foraging optimization (d) entropy-base-classification filter out important information with bacterial foraging optimization. Finally, the error matrices are present and thematic maps are drawn among four cases and outcomes are compared.