Abstract:

Remote sensing semantic image segmentation technology has significant applications in resource management, natural disaster management, and environmental monitoring and protection. However, different remote sensing image datasets often exhibit issues such as spectral confusion between different objects and spectral variations within the same object. These issues significantly reduce the generalization performance of deep learning models, and cross-domain performance degradation in remote sensing semantic image segmentation algorithms poses a significant challenge. To address these issues, optimizations are performed from two perspectives: neural network architecture and domain adaptation strategies. First, a TransConv network based on a hierarchical multihead self-attention mechanism and multiscale feature fusion is proposed. This network effectively enhances feature extraction and fusion capabilities through sliding window patching, multilayer self-attention modules, and a lightweight feedforward neural network, thereby improving the model's generalization performance. Second, a self-training-based domain adaptation technique is introduced, which optimizes the image input, model parameters, and learning process. As a result, labeled source domain knowledge is successfully transferred to the unlabeled target domain, significantly improving the segmentation performance in the target domain. Experimental results demonstrate that the improved TransConv network significantly outperforms other algorithms in terms of generalization performance. In addition, it excels in domain adaptation tasks with the self-training-based domain adaptation technique. The proposed approach thus enhances the accuracy and generalization capability of remote sensing image semantic segmentation, reduces the impact of erroneous pseudo-labels, and addresses the class imbalance problem, providing more reliable technical support for practical applications.

Key words: remote sensing image, Convolutional Neural Network (CNN), Transformer network, hierarchical attention, domain adaptive

摘要：

遥感图像语义分割技术在资源管理、自然灾害管理、环境监测和保护等领域具有重要应用价值, 然而不同的遥感图像数据集往往存在大量的异物同谱和同物异谱等现象, 极大地降低了深度学习模型的泛化性能, 同时遥感图像语义分割算法中存在跨域预测性能下降的问题。为了解决上述问题, 从神经网络模型架构和域自适应策略两个方面进行优化。首先, 提出了基于层级多头自注意力机制与多尺度特征融合的TransConv网络, 通过滑动窗口切块、多层自注意力模块和轻量前馈神经网络, 有效提升特征提取和融合的能力, 从而增强模型的泛化性能。其次, 提出一种基于自训练的域自适应技术, 该技术通过优化图像输入、模型参数和学习过程, 将带标注的源域知识成功迁移至未标注的目标域, 大幅提高了目标域的分割性能。实验结果表明, 改进后的TransConv网络不仅在泛化性能上显著优于其他算法, 基于自训练的域自适应技术也在域自适应任务中表现出色, 提升了遥感图像语义分割的准确性和泛化能力, 减少了错误伪标签的影响和解决了类不平衡问题, 为实际应用提供了更为可靠的技术支持。

关键词: 遥感图像, 卷积神经网络, Transformer网络, 层级注意力, 域自适应