Enhancing Hyperspectral image classification through transformer-based contextualization and novel logarithmic convolutional techniques

Hyperspectral image (HSI) classification is essential in remote sensing and geospatial analysis, particularly for precise land-cover classification. This technique offers detailed spectral information, which allows for the accurate identification and differentiation of various materials and surface types. Recently, there has been a shift towards employing DL techniques to enhance HSI classification. Notably, convolutional neural networks (CNNs) have significantly improved the extraction and analysis of local features, leading to better classification accuracy. However, traditional CNNs face challenges in capturing long-range dependencies and global contextual information within HSI data, which can hinder classification performance. Additionally, these networks tend to be computationally expensive in terms of model parameters and training requirements. To address these limitations, this paper proposes a novel lightweight framework that integrates local and global features while capturing long-range dependencies for enhanced HSI classification. Our approach begins with a combination of novel logarithmic-based 3D and 2D group convolutions to extract correlated spectral–spatial features from the HSI data and address the issues of computational cost and model complexity. To enhance spatial positional encoding and semantic feature mapping, we introduce the Spatial-Context Patch Embedding (SCPE) approach, which improves spatial correlation between patches and reduces complexity. Our novel Spatial Contextual Sine–Cosine Positional Embedding (${\left(SC\right)}^{2}PosEmbed$) consists of two components: the Global Sine–Cosine Position Embedding (GSCPosEmbed) and the Spatial Contextual Position Embedding (SCPosEmbed). This module captures detailed positional information, improving the representation of spatial features in HSI. Finally, we capture the long-term dependencies of spectral and spatial information through the incorporation of a vision Transformer. We evaluate our proposed framework on four publicly available datasets and compare its performance against state-of-the-art methods. The experimental results highlight the superior performance of our model in HSI classification tasks, demonstrating its effectiveness in capturing both local and global features while maintaining computational efficiency. Our framework represents a promising advancement in HSI analysis, offering improved classification accuracy and broader applicability across various domains.