.png){kind=icon}
.png){kind=icon}
Return
Spatial–Spectral Flow Guidance and Dynamic Reconstruction for Few-Shot Open-Set Hyperspectral Recognition
.png){kind=icon}
PRE
AI
Y
K
DOI：10.1109/lgrs.2026.3710306.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Real-world hyperspectral image classification faces limited annotated samples and unknown classes. Under few-shot open-set conditions, existing methods often struggle to model complex spectral–spatial structures and maintain stable discrimination. To address this challenge, this letter proposes a few-shot open-set recognition framework based on structure-flow guidance and dynamic reconstruction. A dual-branch feature extractor comprising a spatial flow-guided feature transformer and a Laplacian-enhanced multiscale network is designed. By regularizing the attention distribution with spatial structural variations and explicitly enhancing local high-frequency textures, it extracts robust spectral–spatial features under limited supervision. The two branch features are further fused into unified embeddings through consensus-guided masked aggregation. To delineate known-class boundaries and reject unknown samples, a similarity-conditioned reconstruction decision (SCRD) strategy is further introduced. Its dynamic spectral autoencoder (DSAE) constructs a task-adaptive memory bank from the support prototypes of the current task and performs conditional reconstruction on query features. By jointly evaluating feature similarity and reconstruction error, the proposed strategy exploits conditional reconstructability differences between known and unknown distributions, enabling adaptive unknown-sample rejection. Experiments on the Pavia University and WHU-Hi-LongKou datasets demonstrate improved known-class recognition while achieving a favorable balance between known-class classification and unknown-sample rejection.
Keywords:
Few-shot open-set recognition
hyperspectral image classification
similarity-conditioned reconstruction decision (SCRD)
spatial flow-guided feature transformer
Journal
I
IF:
4.4
Papers:
486
Citations:
0
Organization
Cited Papers