ESMII-Net: An edge-synergy and multidimensional information interaction network for remote sensing change detection

In recent advancements, deep learning-based methods for change detection have demonstrated rapid capabilities to identify alterations across extensive regions, underscoring significant research and application potential in remote sensing change detection. Nonetheless, these methods currently encount...

Full description

Saved in:
Bibliographic Details
Main Authors: Yixin Chen, Xiaogang Ning, Ruiqian Zhang, Hanchao Zhang, Xiao Huang, You He
Format: Article
Language:English
Published: Elsevier 2025-05-01
Series:International Journal of Applied Earth Observations and Geoinformation
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S1569843225001542
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In recent advancements, deep learning-based methods for change detection have demonstrated rapid capabilities to identify alterations across extensive regions, underscoring significant research and application potential in remote sensing change detection. Nonetheless, these methods currently encounter limitations in feature extraction, often leading to blurred edges and challenges in identifying small-scale changes. To overcome these challenges, we introduce the Edge-Synergy and Multidimensional Information Interaction Network (ESMII-Net) specifically designed for remote sensing change detection. We achieve feature enhancement through the Multidimensional Information Interaction Fusion Module (MIIFM) and, by integrating the edge aware decoder and the Edge-Synergy Module (ESM), guide the model to acquire effective edge information, thereby improving change detection performance. Furthermore, during the loss function formulation, we have incorporated a Small Object Enhancement Factor (SOEF) to prioritize small object detection. An edge-awareness map is also utilized within the model to accurately delineate change edges and assess their influence on adjacent changed pixels. The efficacy of our model and its innovative components has been validated through experimental results on two public datasets, showcasing improved capabilities in detecting edges and small objects.
ISSN:1569-8432