| 摘 要: 针对传统遥感领域使用对比学习方法难以获取多层次、多角度信息的问题,提出了一种自监督对比学习模型MLCL。该模型通过注意力机制捕捉局部细节与整体结构,同时,分别对多层次特征、风格特征和区域特征执行对比学习以获取精准的信息表示。实验在Agricultural-Land-Cover及耕地数据集上进行,MLCL模型的平均交并比(mIoU)值分别达到了78.69%和87.95%,平均准确率(mACC)值为86.43%和93.62%,平均F1值(mF1)值则达到了87.73%和93.59%,均优于Barlow-Twins、BYOL、MoCo和 GLCNet等模型,验证了该方法设计策略的有效性以及对数据量的低依赖性。 |
| 关键词: 遥感 农用地分割 对比学习 多层次特征 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(32201663,32360437) |
|
| A Self-Supervised Semantic Segmentation Model for Agricultural Land Remote Sensing Based on Multi-Level and Multi-Perspective Contrastive Learning |
|
ZHANG Buxun, GUO Xiaoyan
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
3020284554@qq.com; guoxy@gsau.edu.cn
|
| Abstract: To address the challenge of capturing mult-i level and mult-i perspective information with contrastive learning in traditional remote sensing, this paper proposes MLCL—a self supervised contrastive learning model. MLCL employs an attention mechanism to capture both local details and global structures, while performing contrastive learning on mult-i level features, style features, and regional features separately to obtain precise information representations. Experiments on the Agricultura-l Land-Cover and cultivated land datasets show that MLCL achieves mean Intersection over Union (mIoU) scores of 78.69% and 87.95% , mean Accuracy (mACC) values of 86.43% and 93.62% , and mean F1-score (mF1) scores of 87.73% and 93.59% , respectively. These results outperform models including Barlow-Twins, BYOL, MoCo, and GLCNet, demonstrating the effectiveness of the proposed design strategy and its less data-dependent nature. |
| Keywords: remote sensing agricultural land segmentation contrastive learning multi-level feature |
