本研究透過資料探勘辨識都市高光譜影像對不同類別的都市地貌進行分類,研究的區域為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等種地貌。目的是比較支持向量機、隨機森林和類神經網路在都市高光譜影像的分類效能,比較隨機森林特徵選取、線性判別分析特徵轉換和自編碼特徵轉換在都市高光譜影像的降維效果。辨識結果由高到低分別為自編碼特徵轉換+類神經網路 (100%)、自編碼特徵轉換+支持向量機 (100%)、自編碼特徵轉換+隨機森林 (99.07%)、支持向量機 (98.7%)、類神經網路 (98.15%)、隨機森林特徵選取+類神經網路 (97.96%)、隨機森林特徵選取+支持向量機 (97.96%)、線性判別分析特徵轉換+類神經網路 (97.59%)、線性判別分析特徵轉換+支持向量機 (97.41%)、線性判別分析特徵轉換+隨機森林 (95.74%)、隨機森林特徵選取+隨機森林 (94.81%)、隨機森林 (93.89%)。
The purpose of this study is to use urban hyperspectral imagery through data mining to classify the urban landscape hyperspectral imagery with various categories. The research area of this study is near Taiyuan North Road, North District of Taichung City. It has 12 kinds of categories: landforms including roads, river embankments, sidewalks, rivers, shadows, trees, turf, playground, basketball court red, basketball court green, cement roof and tin roof. It is decided to compare the support vector machine, random forest and neural network in urban hyperspectral image classification for parallel study. The classification performance for extracting attribute of random forest feature selection, linear discriminate analysis feature transformation and self-encoding feature transformation are also employed. The classification outcomes in decending order are: self-encoding feature conversion + neural network (100%), self-encoding feature conversion + support vector machine (100%), self-encoding feature conversion + random forest (99.07%), support vector machine (98.7%), neural network ( 98.15%), random forest feature selection + neural network (97.96%), random forest feature selection + support vector machine (97.96%), linear discriminate analysis feature conversion + neural network (97.59%), linear discriminate analysis feature conversion + support vector machine (97.41%), linear discriminate analysis feature conversion + random forest (95.74%), random forest feature selection + random forest (94.81%), random forest (93.89%).