GhoMR

GhoMR enhances hyperspectral image (HSI) classification by employing multi-receptive lightweight residual modules to extract and hierarchically combine spectral-spatial features.

Key Features:

• Multi-receptive lightweight residual modules: Modules combine residual learning with multiple receptive fields (RFs) to extract diverse features from HSI data.
• Multiple receptive fields and hierarchical combination: Distinct RFs extract features that are hierarchically combined by other RFs to form higher-level representations.
• Ghost modules for parameter efficiency: Ghost modules generate a limited set of primary features and apply inexpensive transformations to reduce feature redundancy and network parameters.
• GhoMR-Net architecture: A streamlined neural network implementation that integrates the proposed modules for HSI classification.
• Benchmark datasets: Evaluated on Indian Pines, University of Pavia, and Salinas Scene hyperspectral datasets.
• Evaluation metrics: Performance assessed using overall accuracy (OA), Kappa coefficient (Kappa), and average accuracy (AA).
• Comparative benchmarking: Compared against ten state-of-the-art architectures, demonstrating a lightweight structure with comparable or superior classification accuracy.
• Implementation framework: Implemented using the PyTorch framework.

Scientific Applications:

• Remote sensing HSI classification: Classification of land-cover and surface materials from hyperspectral imagery.
• Spectral-spatial feature extraction: Extraction and combination of spectral and spatial features for improved discriminative representation in HSIs.
• Benchmarking deep learning models for HSI: Comparative evaluation of lightweight CNN architectures on standard hyperspectral datasets.

Methodology:

Uses multi-receptive convolutional neural network modules with residual learning; employs multiple RFs for feature extraction and hierarchical combination; integrates Ghost modules to generate primary features and apply inexpensive transformations to reduce parameters; evaluates the GhoMR-Net on Indian Pines, University of Pavia, and Salinas Scene using OA, Kappa, and AA, and compares results with ten state-of-the-art architectures.