计算机科学
分割
注释
人工智能
土地覆盖
特征(语言学)
钥匙(锁)
编码器
匹配(统计)
交叉口(航空)
封面(代数)
机器学习
图像分割
语义学(计算机科学)
模式识别(心理学)
数据挖掘
边距(机器学习)
特征提取
监督学习
特征匹配
特征向量
遥感
像素
尺度空间分割
作者
Yuanhao Zhao,Genyun Sun,Aizhu Zhang,Jinchang Ren
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-08-03
卷期号:: 1394-1397
标识
DOI:10.1109/igarss55030.2025.11243135
摘要
Semantic segmentation of remote sensing images is essential for applications in environmental monitoring, urban planning, and land cover classification. However, creating large-scale pixel-level annotations for these images is both labor-intensive and costly, which posing a significant challenge for developing high-accuracy segmentation models. In this paper, we propose a novel architecture, Multi-Scale Spectral-Detail-Global Network (MSDGNet), which leverages point-level annotations to minimize annotation requirements while preserving segmentation accuracy. MSDGNet comprises three primary modules: the Spectral-Detail-Global (SDG) encoder, the Multi-Scale Attention Feature Fusion (MSAFF) module, and a simple decoder. The SDG encoder captures comprehensive features by focusing on spectral, detail, and global contextual information, enhancing the model’s ability to distinguish complex land cover types and resist noise. The MSAFF module applies a multi-scale attention mechanism to dynamically weight key features across scales, refining the segmentation performance with limited annotation. Experimental results on the ISPRS Vaihingen and Zurich Summer datasets validate the effectiveness of MSDGNet, showing it outperforms other state-of-the-art models in overall accuracy (OA) and mean intersection over union (mIoU). This model offers a practical solution for large-scale remote sensing applications where minimizing annotation costs is essential.
科研通智能强力驱动
Strongly Powered by AbleSci AI