ELDGG: an end-to-end LiDAR-dynamic-guided GAN for hyperspectral image hierarchical reconstruction and classification

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Hyperspectral image (HSI) and light detection and ranging (LiDAR) data fusion faces challenges in dynamic feature interaction and spatial detail reconstruction.
• Existing fusion methods often struggle with static feature integration and high-fidelity spatial detail preservation.

Purpose of the Study:

• To propose an end-to-end LiDAR-dynamic-guided Generative Adversarial Network (ELDGG) for hierarchical HSI reconstruction and classification.
• To improve the dynamic adaptive interaction of cross-modal features and high-fidelity spatial detail reconstruction in HSI-LiDAR fusion.

Main Methods:

• The ELDGG framework utilizes a guided hierarchical reconstruction generator (GHR-Generator) and a perception-enhanced spectral regularization discriminator (PSR-Discriminator).
• Key innovations include the cross-modal parameter-adaptive fusion module (CPAF-Module) for dynamic feature adaptation and the LiDAR-guided neural implicit field reconstruction unit (L-GNIF Unit) for artifact-free spatial detail reconstruction.
• A perception-enhanced spectral regularization discriminator (PSR-Discriminator) with multi-level feature matching and spectral normalization constraints was developed.

Main Results:

• The proposed CPAF-Module effectively leverages LiDAR global context to generate dynamic convolutional operators for HSI features.
• The L-GNIF Unit achieves high-fidelity, artifact-free feature space reconstruction by learning continuous coordinate-to-feature mappings.
• The PSR-Discriminator provides comprehensive perceptual signals across shallow, mid-level, and deep semantic scales.

Conclusions:

• ELDGG demonstrates superior performance over state-of-the-art methods in both data fusion quality and land cover classification accuracy.
• The end-to-end training and joint multi-task optimization ensure generated fused features possess authenticity and class discriminability.
• The spatial-spectral refinement classifier (SSR-Classifier) effectively decodes optimized feature maps for high-precision land cover classification.