近來,科技的飛速發(fā)展使地理空間專業(yè)人士能夠遠(yuǎn)程收集特定目標(biāo)上的多種類型的傳感器數(shù)據(jù)。這些數(shù)據(jù)帶來的挑戰(zhàn)與機(jī)遇并存。一方面,數(shù)據(jù)的異質(zhì)性給它們的高效處理帶來了挑戰(zhàn),但另一方面,多源數(shù)據(jù)集的涌入和可用性也為組合利用異源和多模式數(shù)據(jù)提供了新的機(jī)會,從而使地理空間應(yīng)用的結(jié)果得到改善。LiDAR數(shù)據(jù)可提供豐富的空間/幾何信息,但由于缺乏光譜信息,其在(復(fù)雜)城市場景識別任務(wù)(如分類)的應(yīng)用范圍有限。高光譜影像數(shù)據(jù)可提供豐富的光譜信息,但會帶來許多問題,例如該數(shù)據(jù)缺少高程信息,同時訓(xùn)練樣本的可用性有限以及來自成像傳感器和環(huán)境的噪聲影響等問題,從而使高光譜數(shù)據(jù)的分類成為挑戰(zhàn)�？紤]到各個傳感器數(shù)據(jù)的局限性和功能,以及分類算法設(shè)計(jì)的本質(zhì)是提取出有效且區(qū)分度高的特征。本研究旨在利用原始LiDAR點(diǎn)云數(shù)據(jù)并結(jié)合該區(qū)域的高光譜數(shù)據(jù),基于深度學(xué)習(xí)的時空光譜特征表示方法獲取城市場景中目標(biāo)的最佳特征從而進(jìn)行分類,并將分類結(jié)果標(biāo)注在原始LiDAR點(diǎn)云上。本文提出了用于初始高光譜特征提取的類-波段選擇和降維方法(CBSR)。然后通過雙分支卷積高斯-伯努利深度置信網(wǎng)絡(luò)(CGBDBN),從LiDAR和高光譜圖像數(shù)據(jù)中提取深...

【文章頁數(shù)】：88 頁

【學(xué)位級別】：碩士

【文章目錄】：
Abstract
摘要
List of Acronyms
1 Introduction
    1.1 Background
    1.2 The State-of-the-art of Point Cloud Labelling
    1.3 Research Content,Workflow and Innovation
        1.3.1 Objectives
        1.3.2 Content
        1.3.3 Workflow
        1.3.4 Innovation
    1.4 Thesis Outline
2 Spatial and Spectral Feature Representation
    2.1 Various Existing Feature Representations
        2.1.1 LiDAR-based features
        2.1.2 Hyperspectral Image-based features
    2.2 Representation Methods
        2.2.1 Classical methods
        2.2.2 Deep learning(DL)methods
    2.3 Datasets
        2.3.1 Data composition and pre-processing
    2.4 Proposed Approach:CBSR for Spectral Feature Representation
    2.5 Proposed Approach:CGBDBN for Spatial Feature Representation
        2.5.1 GBRBMs based on traditional RBMs
        2.5.2 Convolutional GBRBMs
        2.5.3 Convolutional Gaussian-Bernoulli Deep Belief Network
3 Spatial-Spectral Data Fusion and Classification
    3.1 Data Fusion
    3.2 Classification
    3.3 Proposed Method:Spatio-Spectral Feature Representation and Classification through Ensemble Model Stacking
    3.4 2D to3D Projection
4 Experimental Results and Discussion
    4.1 Experimental Setup
        4.1.1 Spatial model
        4.1.2 Spectral model
    4.2 Comprehensive Classification Results
        4.2.1 Houston dataset
        4.2.2 MUUFL Gulfport dataset
    4.3 2D to3D Projection
    4.4 Discussions
        4.4.1 Effects of the image feature patch size
        4.4.2 Performance comparison of the CBSR method to some existing methods
        4.4.3 Performance comparison of the proposed framework to existing methods
        4.4.4 Comparison of ensemble stacking to weighted class probability averaging
        4.4.5 The impact of spatial(DL(L))and spectral(DL(H))feature representations on the proposed spatio-spectral output
5 Conclusions and Recommendation
Acknowledgement
References
Appendix



本文編號：3767684