高光譜影像分類新興深度學習變換器架構之比較：以CTMixer、MAEST與SSTN為例;Comparison of Novel Transformer-based Deep Learning Architecture for Hyperspectral Image Classification: A Case Study of CTMixer, MAEST and SSTN

NCUIR > Research Center > Master of Science Program in Remote Sensing Science and Technology > Electronic Thesis & Dissertation > Item 987654321/98485

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98485

Title:	高光譜影像分類新興深度學習變換器架構之比較：以CTMixer、MAEST與SSTN為例;Comparison of Novel Transformer-based Deep Learning Architecture for Hyperspectral Image Classification: A Case Study of CTMixer, MAEST and SSTN
Authors:	陳映慈;CHEN, YING-TZU
Contributors:	遙測科技碩士學位學程
Keywords:	深度學習;高光譜影像;變換器架構;影像分類;Deep Learning;Hyperspectral Image;Transformer Architecture;Image Classification
Date:	2025-07-26
Issue Date:	2025-10-17 12:50:16 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著遙測（Remote Sensing, RS）與深度學習（Deep Learning, DL）技術的迅速發展，加速了地表物質分類與識別方法的演進與實務應用。高光譜影像（Hyper-Spectral Image, HSI）因具備豐富的光譜與空間資訊，廣泛應用於農業監測、地質探勘及地物分類等領域。然而，其高維度特性與地物類別間的光譜相似性，對於分類任務構成極大挑戰。近年來，深度學習技術迅速發展，特別是結合變換器 (Transformer) 架構的模型在 HSI 分類上展現高度潛力。鑑於這些方法不僅應用於不同數據集，所採用的訓練參數（如訓練週期與學習率）亦有所差異，因此有必要針對其分類表現與運算效率進行更全面的分析與比較。本研究選擇三種近年提出的 Transformer-based 架構進行比較：融合卷積神經網路(CNN) 與 Transformer 結構的 CTMixer、採用遮罩自編碼器設計的 MAEST、以及採用改良式 Swin Transformer 的 SSTN。透過統一的實驗設計與兩組訓練參數配置（CFG1 與 CFG2），分別於 Indian Pines、Pavia University 與 Houston 2013 三組公開數據集上進行分類實驗。研究評估指標涵蓋整體準確率、平均準確率、Kappa 係數、類別分類表現、模型推論時間，並輔以分類地圖進行視覺化分析，以檢視模型在空間邊界識別與分布一致性方面的表現。研究結果顯示，SSTN 在分類準確性與穩定性方面表現最佳，CTMixer 對參數設定較為敏感，適合應用於結構清晰的場景；而MAEST 則具備良好的推論效率，但在光譜相似與樣本不均的情境時，分類表現相對較弱。分類地圖的視覺化進一步揭示各模型在邊界清晰度與區域連續性上的差異：SSTN 可有效維持空間一致性，而 MAEST 在大面積地物的分類上則較易產生破碎現象。綜合比較指出，各架構各具優勢與適用場景，研究結果可作為未來高光譜影像分類任務中，模型選擇與架構設計之有力參考。;In recent years, deep learning techniques have advanced rapidly, with Transformer-based models in particular demonstrating high potential in hyperspectral image classification. Given that these methods are not only applied to different datasets but also adopt varying training parameters (such as training epochs and learning rates), a more comprehensive analysis and comparison of their classification performance and computational efficiency is necessary. This study compares three recently proposed Transformer-based architectures: CTMixer, which integrates Convolutional Neural Network (CNN) and Transformer structures; MAEST, which employs a masked autoencoder design; and SSTN, which adopts a modified Swin Transformer. Through a unified experimental setup and two training parameter configurations (CFG1 and CFG2), classification experiments are conducted on three publicly available datasets: Indian Pines, Pavia University, and Houston 2013. Evaluation metrics include overall accuracy, average accuracy, kappa coefficient, class-wise classification performance, and model inference time. In addition, classification maps are visualized to examine the models′ ability in boundary recognition and spatial distribution consistency. The experimental results show that SSTN achieves the best performance in terms of classification accuracy and stability. CTMixer is more sensitive to parameter settings and is better suited for scenes with clear spatial structures. MAEST demonstrates good inference efficiency but exhibits weaker classification performance in scenarios with high spectral similarity and sample imbalance. The visualization of classification maps further reveals differences in boundary clarity and regional continuity among the models: SSTN effectively maintains spatial consistency, whereas MAEST tends to produce fragmented regions when classifying large-area objects. In summary, each architecture has its strengths and suitable application scenarios. The results of this study can serve as a valuable reference for model selection and architectural design in future HSI classification tasks.
Appears in Collections:	[Master of Science Program in Remote Sensing Science and Technology ] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	5	View/Open

社群 sharing

Loading...