International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 11 Issue: 04 | Apr 2024
p-ISSN: 2395-0072
www.irjet.net
Comprehensive Approach to Hyperspectral Image Analysis: Multiscale Feature Extraction and Spectral Imaging for Classification B.Nancharaiah1, M.V.Subrahmanyam2, R.Manasa3,J.snehitha4,K.Jagan Mohan Achari5 1B.Nancharaiah:Professor, Dept. of Electronics and Communication Engineering,
Usharama College of Engineering and Technology,Telaprolu,Andhra Pradesh-521109,India. 2M.V.Subrahmanyam,3R.Manasa,4J.Snehitha,5K.Jagan Mohan Achari- Students of Usharama College of Engineering
and Technology,Telaprolu,Andhra Pradesh-521109,India. ---------------------------------------------------------------------***---------------------------------------------------------------------
image (HSI) processing technology has seen significant development and widespread application in various domains as rural planning, environmental monitoring, urban planning, vegetation coverage assessment, mineral extraction, national defense infrastructure and precision agriculture.
Abstract - We present a novel approach for
hyperspectral image classification, utilizing a dual-branch architecture for concurrent spatial and spectral feature extraction. Prior to feature extraction, we employ principal component analysis (PCA) to reduce data dimensionality, with varying degrees of downsampling across the two branches. Spatial information is captured through a multiregion piecewise Gaussian pyramid downsampling method, generating multiscale and multiresolution image data. Enhanced ResNet networks are then employed to extract spatial features, enabling the extraction of contextually specific features inherent to hyperspectral images. Meanwhile, spectral information is processed using a unique imaging spectral data technique, involving initial PCA-based dimensionality reduction followed by expansion into N×N images. A dedicated ResNet network, tailored with a distinct number of layers, is utilized for spectral feature extraction, addressing the challenge of variations in spectral data. Subsequently, the spatial and spectral features extracted from the dual-branch network are integrated and fed into a fully connected network for classification, resulting in significantly enhanced classification accuracy. Experimental validation on two benchmark datasets demonstrates the effectiveness of our proposed method, showcasing substantial improvements in classifier accuracy compared to existing approaches.
Consequently,classification techniques in the realm of hyperspectral imagery have also advanced rapidly, playing a crucial role in remote sensing applications. HSI classification, a focal point of recent research, entails assigning specific class labels to individual pixels based on their spatial and spectral characteristics. However, the complexity of HSIs, characterized by numerous spectral bands with high correlation information redundancy, poses computational challenges. Moreover, distinguishing between different materials with similar or identical spectra further complicates classification tasks. Currently, methods for hyperspectral image (HSI) classification can be broadly categorized into those leveraging spectral information alone and those incorporating joint spatial–spectral features. Spectral information-based classification methods rely solely on the spectral dimension of HSIs, disregarding spatial pixel correlations. Examples include support vector machine, random forest, sparse representation, and similar techniques. In contrast, approaches integrating joint spatial–spectral features have shown improved classification performance, utilizing methods such as edgepreserving filtering, multiscale adaptive strategies, lowrank Gabor filtering, and hierarchical guided filtering with nearest-neighbor regularization subspaces. However, these methods heavily rely on manually crafted features, leading to limited classification performance as they may not fully capture the complex content within HSIs. In comparison to traditional shallow methods, deep learning techniques offer enhanced representation and generalization capabilities, capable of extracting deeper image features and achieving more discriminative representations. Consequently, deep learning methods have gained significant traction in HSI classification, including convolutional neural networks (CNNs), two-channel networks, spectral spatial attention networks, and related approaches. Despite their
Keywords:Deep learning, Gaussian pyramid, hyperspectral image (HSI) classification, multiscale feature extraction.
1. INTRODUCTION With the continuous advancements in science and technology, hyperspectral imaging, also referred to as imaging spectroscopy, has experienced rapid progress. This technology involves remote sensing satellites capturing tens of thousands of narrow spectral bands emitted or reflected from a given area. This capability enables the acquisition of more detailed spatial and spectral information compared totraditional panchromatic and multispectral remote sensing images, allowing for improved differentiation between various materials. Due to its advantages in attribute recognition, hyperspectral
© 2024, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 775