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

## TL;DR

This paper introduces ELDGG, a new LiDAR-guided GAN that improves the fusion and classification of hyperspectral images by enhancing spatial reconstruction and cross-modal feature interaction.

## Contribution

The novel CPAF-Module and L-GNIF Unit enable dynamic fusion and high-fidelity spatial reconstruction in HSI-LiDAR fusion.

## Key findings

- ELDGG outperforms existing methods in HSI-LiDAR fusion quality and classification accuracy.
- The proposed framework achieves artifact-free feature reconstruction and enhanced perceptual signals across multiple scales.

## Abstract

To address the prevalent issues in the classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) data fusion, such as insufficient dynamic adaptive interaction of cross-modal features, and difficulties in high-fidelity spatial detail reconstruction, this paper proposes an end-to-end LiDAR-dynamic-guided GAN for hyperspectral image hierarchical reconstruction and classification (ELDGG). The core framework of the network consists of a guided hierarchical reconstruction generator (GHR-Generator) and a perception-enhanced spectral regularization discriminator (PSR-Discriminator). First, we propose the cross-modal parameter-adaptive fusion module (CPAF-Module), which leverages the global context of LiDAR data to generate dynamic convolutional operators tailored for HSI features, addressing the limitations of static fusion methods. Second, to enhance the reconstruction quality of spatial details, we design the LiDAR-guided neural implicit field reconstruction unit (L-GNIF Unit). By learning a continuous mapping from coordinates to features, it achieves high-fidelity and artifact-free feature space reconstruction. Furthermore, we innovatively integrate spectral normalization constraints with a multi-level feature matching mechanism to construct the PSR-Discriminator. This discriminator provides more comprehensive perceptual signals across three scales: shallow textures, mid-level structures, and deep semantics. The entire framework is optimized through end-to-end training and a joint multi-task optimization loss function, ensuring that the generated fused features exhibit both authenticity and class discriminability. On this basis, we further design a spatial-spectral refinement classifier (SSR-Classifier) to accurately decode the deeply optimized feature maps, ultimately producing high-precision land cover classification results. Experiments demonstrate ELDGG’s superiority over state-of-the-art methods in both fusion quality and classification accuracy.