一种基于对抗学习的高分辨率遥感影像语义分割无监督域自适应方法

An adversarial learning-based unsupervised domain adaptation method for semantic segmentation of high-resolution remote sensing images

语义分割模型在高分辨率遥感影像中表现良好的关键是训练集和测试集之间域的高度一致。然而,不同数据集之间存在域偏差,包括地理位置、传感器成像方式和天气条件的差异,导致在一个数据集上训练的模型在另一个数据集上预测时准确性会显著下降。域自适应是解决上述问题的有效策略,该文从域自适应模型的角度,基于对抗学习方法提出了一种用于高分辨率遥感图像语义分割任务的无监督域自适应框架。该框架对全局域对齐模块和局部域对齐模块分别融入熵值加权注意力和逐类别域特征聚合机制,缓解源域和目标域之间的域偏差;此外,引入了对象上下文表征(object context representation,OCR)模块和空洞空间金字塔池化(atrous spatial pyramid pooling,ASPP)模块,以充分利用影像中的空间级和对象级上下文信息,并提出了OCR/ASPP双分类器组合策略,以提高分割精度和准确性。实验结果表明,该方法在公开的2个数据集中实现了优越的跨域分割性能,并超过了同类型的其他方法。

The key to the high performance of semantic segmentation models for high-resolution remote sensing images lies in the high domain consistency between the training and testing datasets.The domain discrepancies between different datasets,including differences in geographic locations,sensors’imaging patterns,and weather conditions,lead to significantly decreased accuracy when a model trained on one dataset is applied to another.Domain adaptation is an effective strategy to address the aforementioned issue.From the perspective of a domain adaptation model,this study developed an adversarial learning-based unsupervised domain adaptation framework for the semantic segmentation of high-resolution remote sensing images.This framework fused the entropy-weighted attention and class-wise domain feature aggregation mechanism into the global and local domain alignment modules,respectively,alleviating the domain discrepancies between the source and target.Additionally,the object context representation(OCR)and Atrous spatial pyramid pooling(ASPP)modules were incorporated to fully leverage spatial-and object-level contextual information in the images.Furthermore,the OCR and ASPP combination strategy was employed to improve segmentation accuracy and precision.The experimental results indicate that the proposed method allows for superior cross-domain segmentation on two publicly available datasets,outperforming other methods of the same type.