A³ CLNN: Spatial, Spectral and Multiscale Attention ConvLSTM Neural Network for Multisource Remote Sensing Data...

Area of Science:

• Remote Sensing
• Geospatial Artificial Intelligence
• Data Fusion

Background:

• Exploiting information from multiple remote sensing data sources is challenging.
• Hyperspectral images (HSIs) and Light Detection and Ranging (LiDAR) data offer complementary information.

Purpose of the Study:

• To develop a novel approach for effectively fusing HSI and LiDAR data.
• To enhance feature extraction and classification accuracy in multisource remote sensing.

Main Methods:

• Proposed a dual-channel spatial, spectral, and multiscale attention convolutional long short-term memory neural network (dual-channel A³CLNN).
• Incorporated spatial, spectral, and multiscale attention mechanisms for HSI and LiDAR data.
• Developed a composite attention learning mechanism with a three-level fusion strategy.
• Utilized a stepwise training strategy inspired by transfer learning.

Main Results:

• The dual-channel A³CLNN demonstrated superior feature representation capabilities.
• Achieved more competitive classification performance compared to state-of-the-art methods.
• Experimental results validated on multiple multisource remote sensing datasets.

Conclusions:

• The proposed dual-channel A³CLNN effectively exploits the complementarity of HSI and LiDAR data.
• The developed attention and fusion mechanisms significantly improve classification accuracy.
• This approach offers a promising solution for multisource remote sensing data analysis.