.png){kind=icon}
.png){kind=icon}
Return
EvoGrad-RSFormer: Evolutionary-Gradient Transformer Architecture Search for Remote Sensing Scene Classification
.png){kind=icon}
PRE
AI
G
Y
X
L
R
L
DOI：10.1109/lgrs.2026.3708498.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Remote sensing scene classification faces significant challenges due to the complex spatial patterns and high intraclass variance. Though Vision Transformers (ViTs) have shown promise in capturing global contextual information, manually designed Transformers are often suboptimal in the remote sensing scene classification task. Neural architecture search (NAS) has been proposed to automatically design neural networks without extensive human expertise. However, existing convolutional neural network (CNN)-based NAS methods and transformer architecture search (TAS) methods often suffer from high computational cost and spatial complexity in the field of remote sensing scene classification. This letter introduces EvoGrad-RSFormer, a novel evolutionary-gradient search framework for automatically designing efficient Transformers for remote sensing scene classification. The proposed method designs an evolutionary-gradient algorithm to search for optimal architectures. Considering the high spatial complexity of high-spatial-resolution (HSR) remote sensing scene classification images, a novel slow-fast evaluation strategy is designed to reduce the search cost by exploring the categorical diversity and capturing subtle spatial relationships. Furthermore, a tailored fitness function is used to ensure the optimal architectures perform well under resource constraints. Experimental results on three HSR datasets demonstrate the superiority of EvoGrad-RSFormer compared to state-of-the-art NAS methods.
Keywords:
Evolutionary gradient
high-spatial-resolution (HSR)
multiobjective
remote sensing
scene classification
transformer architecture search (TAS)
Journal
I
IF:
4.4
Papers:
486
Citations:
0
Organization
Cited Papers