Change Detection Network Based on Transformer and Transfer Learning

The purpose of the change detection(CD) task is to contrast the change information of a specific object in remote sensing images from different time periods. The deep-learning-based change-detection algorithm can extract pixel-level semantic segmentation results for changed objects. Currently, deep...

Full description

Saved in:
Bibliographic Details
Main Authors: Hua Li, Jingyu Li, Guanghao Luo, Liang Zhou, Hao Wu, Zhangcai Yin
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11020642/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The purpose of the change detection(CD) task is to contrast the change information of a specific object in remote sensing images from different time periods. The deep-learning-based change-detection algorithm can extract pixel-level semantic segmentation results for changed objects. Currently, deep learning based change detection algorithms have achieved excellent detection results through the meticulous design of feature extraction and change judgment modules. However, challenges, such as limited generalization ability owing to small-scale change-detection datasets and noise generated by the change judgment module, still exist. Therefore, it is crucial to enhance the generalization capability of the change detection model and improve its ability to detect challenging information in remote sensing building change tasks. In this study, we propose a swin transformer based change detection network (STTL-CD) that incorporates Transfer Learning. Based on the idea of transfer learning, we transferred the weights of the backbone network trained on a large semantic segmentation dataset to initialize our feature extraction network, and enhance the generalization ability of the change detection network. To address the issue of noise during change detection, we established a joint loss function that increases the weight ratio of difficult and easy samples, thereby compelling the model to focus on discerning challenging samples. Our experimental results demonstrate that STTL-CD achieves IoU(0.8571), F1(0.9231) and IoU(0.8902), F1(0.9422) scores in the LEVIR-CD and WHU-CD datasets, surpassing the state-of-the-art change-detection methods developed in recent years.
ISSN:2169-3536