Author(s): Han Wang; Shunyu Yao; Qing Li; Tao Sun; Changjun Liu

Linked Author(s):

Keywords: Change detection; Convolutional neural networks (CNN); Remote-sensing image with high resolution; Self-attention mechanism; Vision Transformer (ViT)

Abstract: With the rapid advances in monitoring technology, remote-sensing images have become gradually available with a dramatic improvement in both quality and resolution, which can significantly contribute to natural disaster analysis such as flash floods, drought, water and soil erosion, etc., strongly influenced by human activities. However, the accuracy and efficiency in detecting anthropogenic influence remain challenging due to the complex, interdependent abundance of feature details within these high-quality images. Therefore, this study proposes a new, change detection framework to enhance the capability of capturing both local features and their long-range spatiotemporal relationships in context by renovating the Vision Transformer (ViT) structure and incorporating it with convolution neural networks (CNNs). The model framework is then trained by thousands of remote sensing images and the performance is finally evaluated by comparing to several widely used networks.

DOI: https://doi.org/10.3850/iahr-hic2483430201-468

Year: 2024