A Vision Transformer-Based Approach to Remote Sensing Scene Classification: Exploring the Impact of Dropout Regularization
Year: 2025
Type: conference-paper
Abstract: Remote sensing image classification is an important problem for environmental monitoring and urban planning. While convolutional neural networks (CNNs) have shown better results in several remote sensing tasks, they are unable to capture long-range dependencies and global spatial relationships. Vision Transformers (ViTs) have become popular because they can represent local and global features more accurately. However, ViTs face challenges such as spatial information handling and overfitting. We propose a ViT-based architecture integrating hierarchical feature extraction and spatial encoding to improve the performance of the remote sensing scene classification. Additionally, we systematically analyze dropout regularization at critical layers—patch embedding, spatial encoding, and multi-head attention—to mitigate overfitting. The proposed approach is evaluated on two public datasets—NWPU-RESISC45 and UCM. Experimental results shows our methods are effectiveness in remote sensing image classification.
Authors: Odbal, Huakun Zhang
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413845123