都市高光譜影像多目標地貌辨識－機器學習之隨機森林之特徵轉換之探索

Multi-target Landform Image Classification in Urban Hyperspectral Image-The Discussion on Feature Transformation of Random Forest in Deep Learning

萬絢(Wan, Shiuan) ； 鄭育欣(Cheng, Yu-Hsin) ； 馬弘霖(Ma, Hong-Ling)

《嶺東學報》 49期 (2022/12) Pp. 33-50

摘要

本研究的目的是透過資料探勘辨識都市高光譜影像，來探討都市高光譜影像辨識度較差或較混淆的地貌有哪些，了解隨機森林在都市高光譜影像的分類效能，比較線性判別分析（Linear Discriminate Analysis; LDA）特徵轉換在隨機森林於都市高光譜影像的降維效果。本研究的研究地區是台中市北區太原北路附近，其擁有道路、河堤、人行道、河、陰影、樹、草皮、操場、籃球場紅、籃球場綠、水泥屋頂、鐵皮屋頂12種地貌，各取90個點為訓練資料，45個點為測試資料，使用的分類器有隨機森林和使用的特徵選取有線性判別分析LDA特徵轉換，所以整個研究可分為原始影像＋隨機森林、LDA特徵轉換＋隨機森林兩種組合，原始影像以隨機森林辨識正確率為89.81%而經過LDA特徵轉換的隨機森林辨識正確率為95%。

關鍵字

影像判釋 ； 都市高光譜影像 ； 隨機森林 ； 線性判別分析

並列摘要

The goal of this research is to classify urban hyperspectral images through machine learning analysis. It is decided to explore the poorly recognizable or confusing landforms through urban hyperspectral images to have better understand the classification efficiency of random forests in urban hyperspectral images. In the parallel study, the Linear Discriminant Analysis; (LDA) feature transformation applying to a random forest of urban hyperspectral images. The study area is near Taiyuan North Road, North District, Taichung City, which has 12 land-cover categories including roads, river banks, sidewalks, rivers, shadows, trees, turf, playground, basketball court red, basketball court green, cement roof, and iron roof. The study selects 90 points each as training data and 45 points as testing data. The classifier used supports random forest, and the feature selection includes linear discriminant analysis LDA feature conversion and random forest feature selection. The study can be divided into two combinations of two case as: original image + random forest and LDA feature transformation + random forest. The original image has an accuracy rate of 89.81% with random forest, and the identification accuracy of random forest with LDA feature conversion is 95%.

並列關鍵字

Image classification ； Urban hyperspectral image ； Random Forest ； Linear Discriminant Analysis

參考文獻

Wan, S.; Wang, Y. P. The comparison of density-based clustering approach among different machine learning models on paddy rice image classification of multispectral and hyperspectral image data. Agriculture 2020, 10, 1–18, doi:10.3390/agriculture10100465

中科英特（2017），高光譜在精準農業方面的應用，檢自http://www.zkinter.com/ListView.Asp?ID=276

李仁鐘、李秋緣（2017），R 語言資料分析從機器學習、資料探勘、文字探勘到巨量資料分析，博碩文化股份有限公司。

林軒如（2011），應用 GIS 評估台中市集集地震地價災害損失時空特性，國立中正大學應用地球物理研究所碩士論文，嘉義縣。

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。