Interactive Graph-Based Distillation Integrated Meta-Learning Network for Hyperspectral Image Incremental Classification

doi:10.1109/TGRS.2025.3647656

Interactive Graph-Based Distillation Integrated Meta-Learning Network for Hyperspectral Image Incremental Classification

2026-01-01

0

PRE

AI

DOI：10.1109/TGRS.2025.3647656

Original

Original request for help

Save

Abstract

En 中文

Hyperspectral image (HSI) incremental classification aims to break the limitation of the fixed-category HSI classification models and implement adaptive recognition of new categories through incremental learning. The inherent characteristic of spectral variability in HSI aggravates distribution drift between new and old classes, and exacerbates the forgetting of old class knowledge in the HSI incremental classification. In this article, we propose the interactive graph-based distillation integrated meta-learning network (IGDIMN) for HSI incremental classification, which achieves the efficient memory transfer of old-class knowledge between the initial training phase and the incremental training phase (ILP). Specifically, we present the dual-graph network with dynamic distillation (DGNet-DD), which refines old-class knowledge in the incremental phase based on the hierarchical dual-graph interaction strategy. In the DGNet-DD, the proposed interactive graph distillation (IGD) module performs hierarchical distillation on two generated graphs that reflect instance-level and distribution-level relations of spatial–spectral features. The IGD module facilitates knowledge transfer between the initial training phase (ITP) and ILP stages, preserving learned knowledge and enhancing the capability to recognize new classes. Without relying on old-class exemplar replay, we further propose the prototype-aided memory transfer module (PAMTM) that leverages new-class prototypes to suppress distribution drift between old and new land-cover classes in the ILP. Besides, we adopt a multilevel prototype constraint to further enhance the knowledge transfer capability across different stages. Extensive experiments on three popular HSI datasets validate the superiority of the proposed IGDIMN method compared to other typical HSI incremental classification approaches.

Keywords:

Graph neural networks (GNNs)

hyperspectral image (HSI) incremental classification

knowledge distillation (KD)

meta-learning (ML)

prototype learning (PL)

Interactive Graph-Based Distillation Integrated Meta-Learning Network for Hyperspectral Image Incremental Classification

Abstract

Journal

Organization