Title

綠環境生態系列研究:在農作物以高光譜影像多元分類的輔助因子探討

Translated Titles

Green-environment ecology series study: The multi-classification approach among various corps by hyperspectral image data

Authors

萬絢(Shiuan Wan);張士勳(Shih-Hsun Chang);陳瑋(Wei Chen);林若琪(Ruo-Chi Lin)

Key Words

高光譜影像 ; 影像分類 ; 支持向量機 ; hyperspectral image data ; image classification ; support vector machine

PublicationName

水保技術

Volume or Term/Year and Month of Publication

12卷1期(2018 / 05 / 01)

Page #

10 - 15

Content Language

繁體中文

Chinese Abstract

高光譜影像(Hyperspectral Image)在具有大空間解析度、高頻譜解析度之特性,使的光譜影像具有更豐沛且細膩的光譜資訊,理論上有助於分辨更細微差異之地物而若農作物地區地貌(Ground truth)複雜,傳統光譜影像因為光譜解析度較低、波段數相對較少許多,在影像分類的案例中大多需要加入紋理資訊、植生指標等輔助因子區分不同物種差異性,以達到更理想的分類精度,本研究將探討高光譜影像中含有的大量的波段資訊是否仍需要紋理資訊之輔助因子之必要是本研究的主要重點。外埔水稻田的CASI遙測影像為實驗圖資,並區分為三種影像判釋策略:(1)紋理資訊,(2)植生指標,(3)維度縮減。使用二種紋理資訊(Contrast、Energy)、與維度縮減之方法(粗集合理論),分類器使用支持向量機(Support Vector Machine ,SVM)逐一對各種不同策略之分類精度評估。高光譜平均整體正確率94.765%,但透過SVM繁雜計算,計算時間長,屬性刪減可以加速SVM(節省時間約60%)運算但精度約降低0.7%加入紋理資訊則計算時間更長但約可提升0.6%的判釋精度。

English Abstract

Hyperspectral Image has the characteristics of large spatial resolution and high spectral resolution. Spectral images have a richer and more delicate spectral information that theoretically helps to distinguish the detail land-cover of category. If the ground truth of the crops is complex, the traditional spectral images have a relatively small number of bands because of their low spectral resolution. In the case of image classification, most of the cases need to include ancillary information such as texture information and vegetation index to distinguish between target categories for achieving a better classification accuracy outcomes. This goal of this study is to use hyper-spectral image data to analyze the multi-category image data of Golden Galley of Chia Yi. To improve the classification results, this study adopt (1) texture information: Contrast and Energy (2)vegetation index: Normalized Difference Vegetation Index (NDVI) and Modified Soil adjusted vegetation index, (SAVI). The dimension reduction obtaining by Rough Set is used to extract the features of bands. On the other hand, an approach of this study is to observe the classification performance of classifier: Support Vector Machine, (SVM). Different cases of strategies are considered by former situations and the outcomes are compared respectively.

Topic Category 工程學 > 水利工程
工程學 > 市政與環境工程