Work

Hyper-neighborhood context-aware transformer network for high-resolution remote sensing change detection

Year: 2025

Type: article

Abstract: • HCTNet integrates CNN and Transformer for enhanced long-range contextual modeling. • HCTNet integrates diverse information from multiple hyper-neighborhood samples. • HCTNet adaptively incorporates relevant neighborhood information. • HCTNet fuses extensive contextual and local multi-scale features. • HCTNet achieves state-of-the-art performance on three datasets. Change detection in high-resolution remote sensing images is pivotal for applications such as urban development, environmental monitoring, and disaster response. However, the current deep-learning-based methods that primarily rely on cropped samples exhibit significant limitations in effectively modeling ultra-long-range contextual information, particularly in the context of large-scale land-use and land-cover changes in high-resolution remote sensing images. To overcome these limitations, we introduce the Hyper-neighborhood Context-aware Transformer Network (HCTNet), which synergizes Convolutional Neural Networks (CNNs) and Transformer frameworks to enhance ultra-long-range contextual modeling while maintaining the capability to extract fine-grained details. HCTNet incorporates three core components: the Hyper-neighborhood Context Branch (HCB) for aggregating extensive contextual information, the Hyper-neighborhood Context-aware Position Embedding (HCPE) for adaptive integration of neighborhood information, and the Hyper-neighborhood Context Transformer (HCT) for fusing local multi-scale features with broad contextual insights. These innovations enable HCTNet to effectively capture both large-scale scene changes and the intricate details of these changes. Comprehensive experiments on two publicly available datasets, namely HRSCD and WHU-CD, along with a newly constructed dataset, the China Wide Scope Change Detection (CWSCD), demonstrate the superior performance of HCTNet, achieving F1 scores of 93.41% on WHU-CD, 68.12% on HRSCD, and 72.01% on CWSCD dataset.

Source: International Journal of Applied Earth Observation and Geoinformation

DOI: 10.1016/j.jag.2025.104861