Difference-Augmented Continual Learning for Language-Instructed Spectral–Spatial Hyperspectral Image Classification

Peidong Qi

doi:10.1109/LGRS.2026.3662278

Difference-Augmented Continual Learning for Language-Instructed Spectral–Spatial Hyperspectral Image Classification

2026-02-06

0

PRE

AI

DOI：10.1109/LGRS.2026.3662278

Original

Original request for help

Save

Abstract

En 中文

Hyperspectral image (HSI) provides abundant spectral–spatial information for land cover classification. With the growing availability of HSI datasets, continual learning has become a necessity for RS applications. However, it has been demonstrated that spectral–spatial features vary across different HSI datasets. Consequently, classification models are susceptible to the “catastrophic forgetting” problem and fail to accurately classify HSI continually. To address this, a DTI mechanism has been proposed. The model employs the virtual category “Diff Curve” to input text with spectral–spatial difference information, thereby guiding the model to identify the features of new and old categories, and simultaneously adapting to new knowledge while retaining old information to achieve continual learning without overly reusing previous data. Experimental results demonstrate that the average forgetting (AF) of using DTI alone differs by less than 1% from that of traditional replay methods, while the combination of DTI and replay reduces the forgetting rate by 6.53% compared to the replay-only method, fully verifying its superior adaptability and anti-forgetting performance.

Keywords:

Deep learning

hyperspectral image (HSI) classification

life-long learning

remote sensing

Difference-Augmented Continual Learning for Language-Instructed Spectral–Spatial Hyperspectral Image Classification

Abstract

Journal

Organization