基于改进DeeplabV3+的滑坡识别方法优化
Optimization of landslide identification Method Based on Improved DeeplabV3+
万逸轩 1黄建华 1孙希延 1罗明明1
作者信息
- 1. 桂林电子科技大学信息与通信学院,广西 桂林 541004;桂林电子科技大学广西精密导航技术与应用重点实验室,广西 桂林 541004
- 折叠
摘要
对于滑坡的研究主要包括两个方面,滑坡识别和滑坡灾害的风险评估,而滑坡的识别是所有滑坡研究的基础.随着计算机视觉的发展,将深度学习应用于滑坡识别的研究也在相继展开.通过残差接入通道注意力机制以及设计一个多尺度特征融合模块以融合不同尺度的特征,提出了一种基于deeplabV3+的改进的滑坡识别方法.根据制作的包含 1227 张滑坡图像的数据集,在此数据集上进行了对比实验与消融实验,将数据集以比例9:1 分为训练集和验证集,结果表明在验证集上提出的优化方法效果最佳,其中平均交并比mIoU达到了 84.12％、PA达到了94.23％.相比其它网络中表现最好的网络HR-NetV2,评价指标mIoU也提高了2.54％、PA提高了1.01％.
Abstract
The study on landslides mainly includes two aspects,landslide identification and landslide hazard risk assessment,and landslide identification is the basis of all landslide research.With the development of computer vi-sion,the application of deep learning to landslide recognition has also been carried out one after another.In this pa-per,an improved landslide recognition method based on deeplabV3+is proposed by using the residual access channel attention mechanism and designing a multi-scale feature fusion module to fuse features at different scales.According to the dataset containing 1227 landslide images,comparative experiments and ablation experiments were carried out on this dataset,and the dataset was divided into training set and validation set with a ratio of 9:1.And the results show that the optimization method proposed in this paper has the best effect on the validation set,in which the average in-tersection union ratio of mIoU reaches 84.12％ and PA reaches 94.23％ .Compared with other best-performing net-works,the HRNetV2 evaluation index mIoU also increases by 2.54％ and PA by 1.01％ .
关键词
滑坡图像/神经网络/通道注意力机制/多尺度特征/特征融合Key words
Landslide image/Neural networks/Coordinate attention mechanical/Multi scale feature/Feature fusion引用本文复制引用
基金项目
国家自然科学基金(61861008)
国家自然科学基金(62061010)
国家自然科学基金(62161007)
桂林市科技局桂林市国家可持续发展议程创新示范区建设重点项目(20190219-1)
广西自然科学基金(2018GXNSFAA294054)
广西自然科学基金(2019GXNSFBA245072)
出版年
2024
NETL
NSTL
万方数据