<p>Hyperspectral image classification faces significant challenges due to high-dimensional data and variability, often leading to inefficiencies in processing and reduced classification accuracy. Existing methods struggle with handling high-dimensional data, exhibit limited generalization capabilities, and lack efficiency in feature selection. To address these issues, this research proposes an advanced deep learning-based approach that integrates noise reduction, contrast enhancement, and min–max normalization for preprocessing. Feature extraction is performed using an Improved ResNet50 model, while optimal feature selection is achieved through the Improved Artificial Rabbits Optimization (IARO) algorithm, which effectively reduces dimensionality while preserving essential features. Classification is carried out using a Multi-Scale Residual Depthwise Separable Convolutional Network (MRDSCNN), leveraging residual learning and depthwise separable convolutions to enhance performance. The proposed approach significantly improves data quality and model accuracy, providing a robust and efficient solution for hyperspectral image classification. These innovations collectively address the limitations of traditional methods, offering a more reliable and effective classification system.</p>