上海航天（中英文）2025，Vol.42Issue(5)：101-111,11.DOI:10.19328/j.cnki.2096-8655.2025.05.012
基于进化多任务剪枝的SAR图像变化检测
SAR Image Change Detection Based on Evolutionary Muti-tasking Pruning
摘要
Abstract
Currently,synthetic aperture radar(SAR)images are usually processed by the end-to-end deep neural network(DNN)methods for change detection.However,such methods are dependent on the datasets to be detected,and require customized model design and parameter optimization for the specific datasets.In this paper,an SAR image change detection method based on evolutionary multi-task optimization pruning is proposed.First,a general change detection model is trained using mixed datasets to address the issue of insufficient high-quality samples from single datasets.Second,for specific change detection tasks,an evolutionary multi-task neural network pruning approach is adopted to extract a specialized model from the general model.Finally,the specialized model is fine-tuned with limited samples.The experiments on six representative datasets demonstrate that the specialized model extracted by the proposed method has a size less than 10%of the general model,and can achieve comparable detection results with those of conventional deep neural network(DNN)approaches.关键词
合成孔径雷达(SAR)图像/变化检测/神经网络剪枝/进化多任务优化/低轨(LEO)遥感卫星Key words
synthetic aperture radar(SAR)image/change detection/neural network pruning/evolutionary multi-tasking optimization/low-Earth-orbit(LEO)remote sensing satellite分类
航空航天引用本文复制引用
吴涛,黄祖镇,黄龙,徐一凡,王海涛,蔡津剑,黄鹏辉..基于进化多任务剪枝的SAR图像变化检测[J].上海航天（中英文）,2025,42(5):101-111,11.基金项目
国家自然科学面上基金资助项目(62171272) （62171272）
