融合残差注意力机制的深度可分离UNet泥石流堆积扇分割
Segmentation of debris flow fan by a depth-separable UNet incorporating a residual attention mechanism
宋昕 1王保云 2朱双颖3
作者信息
- 1. 云南师范大学 数学学院,云南 昆明 650500
- 2. 云南师范大学 数学学院,云南 昆明 650500;云南师范大学 云南省现代分析数学及其应用重点实验室,云南 昆明 650500
- 3. 昆明学院 建筑工程学院,云南 昆明 650500
- 折叠
摘要
为解决传统机器学习方法在泥石流堆积扇识别中的精度低、效果差的问题,提出一种基于深度学习的残差注意力可分离UNet算法(RAMS-UNet).该算法在编码部分采用VGG16 主干网络进行特征提取,加深网络层次;在跳跃连接部分引入改进的注意力机制,强化信息传递;在解码部分使用深度可分离卷积和密集连接块,进一步增强空间和通道上的信息表达能力.研究结果表明:与其他算法相比,RAMS-UNet算法对泥石流堆积扇的分割精度更高,mIoU、mPA、PA和F1 指数等评价指标均显著提升.RAMS-UNet算法突破了传统方法在泥石流堆积扇识别中的局限性,为泥石流灾害评估提供了更加精准的信息支持.
Abstract
In order to solve the problem of low accuracy and poor effect of traditional machine learning methods in debris flow accumulation fan recognition,a residual attention separable UNet algorithm based on deep learning(RAMS-UNet)is proposed.The algorithm uses VGG16 backbone network for feature extraction in the coding part to deepen the network level;an improved attention mechanism is introduced in the jump connection part to strengthen information transmission.In the decoding part,deep separable convolution and dense connection blocks are used to further enhance the information expression ability on space and channels.The results show that compared with other algorithms,the RAMS-UNet algorithm has higher segmentation accuracy for debris flow accumulation fans,and the evaluation indexes such as mIoU,mPA,PA and F1 index are significantly improved.The RAMS-UNet algorithm breaks through the limitations of traditional methods in the identification of debris flow fans,and provides more accurate information support for debris flow disaster assessment.
关键词
泥石流堆积扇/沟谷型泥石流/语义分割/UNet算法/注意力机制/深度可分离卷积Key words
debris flow fan/valley-type debris flow/semantic segmentation/UNet algorithm/attention mechanism/depth-wise separable convolution引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据