Multi-source remote sensing image classification based on two-channel densely connected convolutional networks.

Area of Science:

• Geospatial analysis
• Computer vision
• Machine learning

Background:

• Traditional remote sensing classification methods using medium- or low-resolution images suffer from low accuracy and poor automation.
• Deep learning improves classification accuracy but deep networks risk overfitting.
• Multi-source remote sensing data potential is often underutilized due to ineffective feature organization.

Purpose of the Study:

• To propose a novel Two-channel Densely Connected Convolutional Networks (TDCC) for automated ground surface classification.
• To effectively utilize multi-source remote sensing data, specifically hyperspectral images (HSI) and Light Detection and Ranging (LiDAR).
• To overcome the limitations of traditional methods and deep learning overfitting in remote sensing classification.

Main Methods:

• Multi-source remote sensing data (HSI and LiDAR) were pre-processed and re-sampled.
• Two-channel densely connected convolutional networks were developed for extracting spatial-spectral features from HSI and LiDAR data.
• A feature fusion network was designed to integrate HSI and LiDAR features for pixel-wise classification.

Main Results:

• The proposed TDCC method demonstrated competitive classification performance compared to state-of-the-art methods.
• The TDCC achieved superior results in terms of Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficient.
• The method proved effective for classifying complex ground surfaces.

Conclusions:

• The TDCC method offers an effective approach for remote sensing image classification by leveraging multi-source data.
• The proposed network architecture successfully extracts and fuses spatial-spectral features from HSI and LiDAR.
• TDCC provides a robust solution for accurate and automated classification of complex terrestrial environments.