Superpixel Guided Deformable Convolution Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional neural networks (CNNs) excel at nonlinear feature extraction in hyperspectral image (HSI) classification.
• Regular CNNs struggle with fixed convolution kernels, leading to over-smoothed boundaries and degraded performance in HSI classification.
• Accurate classification of HSI is crucial for various applications, including land cover mapping and environmental monitoring.

Purpose of the Study:

• To propose a novel Superpixel Guided Deformable Convolution Network (SGDCN) for improved HSI classification.
• To address the limitations of regular convolutions in capturing spatial-spectral information at object boundaries.
• To enhance the accuracy and robustness of HSI classification, particularly at the boundaries of ground objects.

Main Methods:

• Superpixel region fusion filter (SRF-Filter) to create homogeneous regions with multi-scale spatial features.
• Superpixel guided deformable convolution (SGD-Conv) to adapt convolution shapes to land cover boundaries for pure feature extraction.
• Superpixel joint bilateral filter (SPJBF) to mitigate pixel-level and region-level misclassifications by leveraging superpixel homogeneity.

Main Results:

• The proposed SGDCN achieved superior classification performance on three HSI datasets.
• The method demonstrated significant improvements compared to twelve state-of-the-art HSI classification techniques.
• SGDCN effectively extracts pure spatial-spectral features and enhances classification accuracy at object boundaries.

Conclusions:

• The SGDCN offers a significant advancement in hyperspectral image classification by effectively integrating superpixel segmentation with deformable convolutions.
• The proposed network architecture successfully overcomes the limitations of traditional CNNs in handling complex spatial-spectral information and object boundaries.
• SGDCN provides a robust and accurate solution for HSI classification, paving the way for more precise land cover analysis.