Hyperspectral image (HSI) classification is a crucial task in various remote sensing applications. Deep learning, particularly Convolutional Neural Networks (CNNs), have emerged as powerful tools for extracting complex features from HSI’s high-dimensional data. However, two significant challenges hinder accurate classification: limited labelled training data and reliance solely on spectral information. This paper proposes a novel logarithmic group 3D-CNN architecture specifically designed for HSI classification. Our key contribution is a unique multi-scale spectral-spatial feature extraction approach. Utilizing multi-kernel convolutions arranged in a pyramidal structure, we capture spectral and spatial features at various scales, ensuring comprehensive feature representation. Furthermore, we introduce a novel strategy as a part of spectral-spatial feature extraction by leveraging logarithmic channel division within standard convolutions. This innovative approach significantly reduces the number of model parameters and FLOPs, leading to a lightweight and efficient network. To validate the effectiveness of our proposed architecture, extensive evaluations are conducted on benchmark datasets with varying train-test splits. The results demonstrate that our model performs competitive classification compared to existing state-of-the-art methods. This accomplishment highlights the potential of the logarithmic group 3D-CNN architecture for accurate and efficient HSI classification, particularly in scenarios with limited training data.