Hybrid Dilated Convolution with Multi-Scale Residual Fusion Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Traditional Convolutional Neural Networks (CNNs) for hyperspectral image (HSI) classification often prioritize spectral features over spatial information.
• This limitation can hinder the comprehensive analysis and accurate classification of complex HSI data.

Purpose of the Study:

• To introduce a novel HSI classification model, the Local and Hybrid Dilated Convolution Fusion Network (LDFN).
• To effectively fuse local details and rich spatial features by expanding the receptive field, thereby improving classification performance.

Main Methods:

• The LDFN model integrates standard convolution, average pooling, dropout, and batch normalization operations.
• It employs fusion operations of local and hybrid dilated convolutions to extract rich spatial-spectral information.
• Residual fusion networks aggregate different convolution layers before classification via a softmax layer.

Main Results:

• Experimental results on three widely used hyperspectral datasets (Salinas, Pavia University, and Indian Pines) demonstrate LDFN's effectiveness.
• The proposed LDFN model achieved superior performance compared to existing state-of-the-art classifiers.

Conclusions:

• The LDFN model successfully addresses the limitations of traditional CNNs in HSI classification by incorporating spatial information.
• The fusion of local and hybrid dilated convolutions offers a promising direction for enhancing hyperspectral image analysis and classification accuracy.