Deep learning (DL) has profoundly transformed Hyperspectral Image (HSI) classification by capturing robust features in an end-to-end manner. However, traditional DL models face significant hurdles, including the need for massive amounts of labeled data and the manual tuning of complex structural hyperparameters.

This post delves into two advanced methodologies designed to overcome these challenges: Differentiated-Scale Restricted Graph Convolutional Networks (DSR-GCN) for few-shot learning and Attention-Based Second-Order Pooling Networks (A-SPN) for simplified, yet powerful, feature extraction.

1. Overcoming Few-Shot Limitations with DSR-GCN

A major challenge in HSI classification is the scarcity of labeled training samples. While Graph Convolutional Networks (GCNs) show immense potential for few-shot learning by modeling spatial structures, traditional GCNs rely on single-scale graph constructions that often ignore subtle relations between small regions or cause oversmoothing.

The DSR-GCN Framework

To address this, the Differentiated-Scale Restricted Graph Convolutional Network (DSR-GCN) introduces a novel internal and external solution:

1. Differentiated-Scale Graph Construction: Instead of relying on a single segmentation scale, DSR-GCN builds both a refined localized graph (small scale) and a rougher initial graph (big scale). This allows the model to leverage and balance information from both levels.
2. Restricted Adaptive Weighted Feature Fusion: The network adaptively fuses features from both scales using a learnable weight. A restricted fusion loss is applied to ensure that the features from different scales remain consistent when projected to the category dimension, avoiding contradictions.
3. Spectral-Spatial Siamese Network (S3Net): To remedy the loss of pixel-level features often caused by superpixel-level graph convolution, a lightweight S3Net is integrated. S3Net uses contrastive learning to push heterogeneous samples apart while pulling homogeneous samples closer, providing highly discriminative pixel-level features.

Performance

Evaluated across multiple datasets (e.g., Indian Pines, University of Pavia), DSR-GCN demonstrated significant superiority under few-shot scenarios (e.g., only 5 labeled samples per class). For instance, on the Indian Pines dataset, DSR-GCN achieved an Overall Accuracy (OA) of 82.29%, outperforming other models by 6.20% to 38.58%.

2. Simplifying Architecture and Enhancing Features with A-SPN

While Convolutional Neural Networks (CNNs) are the standard for HSI classification, they typically rely on first-order pooling operations (like max or average pooling) that ignore crucial spectral correlations. Furthermore, tuning their complex structural parameters (layer depth, kernel size) is computationally laborious.

The A-SPN Framework

To resolve this, the Attention-Based Second-Order Pooling Network (A-SPN) was proposed, replacing complex hierarchical CNNs with a plain, highly efficient architecture:

1. First-Order Feature Operator: A simplified operator extracts spectral-spatial information using PCA, patch extraction, batch normalization, and dropout, avoiding the gradient vanishing issues of deeper networks.
2. Attention-Based Second-Order Pooling (A-SOP): The core innovation is the A-SOP operator. Traditional second-order pooling captures the correlation between spectral channels but ignores the diversity of neighboring pixels within a spatial patch. A-SOP solves this by introducing a spatial attention mechanism. It calculates a correlation matrix and uses a learnable cosine distance function to assign different weights to neighboring pixels based on their relevance to the central pixel.
3. Classification: The network concludes with Frobenius normalization, vectorization, and a fully connected layer with softmax loss.

Advantages and Performance

A-SPN effectively discards redundant features and intensifies class-oriented second-order statistics. Because it avoids interleaved convolutional layers, it is free of complex hyperparameter tuning (requiring only patch size adjustments) and trains incredibly fast.

In experiments on the University of Pavia dataset, A-SPN achieved an OA of 99.65%, with training times significantly faster than traditional 2D and 3D CNNs. A-SOP proved highly capable of modeling both inter-class discriminative features and intra-class consistencies.

3. Summary

Both DSR-GCN and A-SPN represent significant leaps forward in HSI classification.

• DSR-GCN tackles the lack of labeled data by rethinking graph construction, using differentiated scales and siamese networks to maximize spatial and spectral information extraction.
• A-SPN challenges the necessity of complex deep CNNs by proving that a simplified network, powered by attention-weighted second-order pooling, can achieve state-of-the-art accuracy with a fraction of the computational effort.