This literature review surveys the evolution of hyperspectral-LiDAR data fusion from 2004 to 2026, categorizing approaches into traditional statistical methods, deep learning techniques, and proposed optimal estimation frameworks. Despite over 1000 papers addressing fusion for classification, the intersection of joint denoising using minimum-variance fusion (MVF) remains essentially unexplored. The paper presents a taxonomy of feature-level, decision-level, and CNN-based methods, highlighting their limitations in handling heterogeneous noise characteristics. It proposes MVF fusion with adaptive noise-weighted estimation as a novel research direction for simultaneous HS-LiDAR restoration.
Key findings
While hyperspectral denoising (1000+ papers) and LiDAR processing (500+ papers) are mature fields, joint denoising using optimal estimation theory has zero existing publications.
All existing HS-LiDAR fusion research focuses exclusively on classification tasks, ignoring the complementary noise characteristics of Gaussian/sparse HS noise and outlier LiDAR noise.
Deep learning fusion methods (CNNs, attention mechanisms, contrastive learning) dominate recent literature but remain black-box approaches focused on classification rather than restoration.
Minimum-variance fusion (MVF) from structured light reconstruction represents an untapped theoretical framework for adaptive noise-weighted joint denoising.
The field has evolved from handcrafted feature concatenation (2004-2017) to cross-modal transformers (2020-2024), yet statistical optimal estimation remains unapplied to this fusion problem.
Limitations & open questions
The proposed MVF approach lacks experimental validation or empirical results demonstrating denoising performance.
The survey relies on citation counts and landscape analysis rather than comparative experimental benchmarking of existing methods.
Registration requirements and computational complexity of MVF for high-dimensional HS data are not fully addressed.