基于Transformer模型的河湖疑似"四乱"图斑遥感智能识别技术研究
Research on intelligent remote sensing recognition technology of suspected"Four Illegal Behaviors"spots in rivers and lakes based on Transformer model
王越 1昝糈莉 2李兆永 3邸苏闯 2朱祺琪 4李佳伦 4王彦强 5李巧玲2
作者信息
- 1. 河海大学 江苏 南京 210098
- 2. 北京市水科学技术研究院 北京 100048
- 3. 菏泽市引黄灌溉工程管理服务中心 山东 菏泽 274000
- 4. 中国地质大学 湖北 武汉 430074
- 5. 北京市河湖流域管理事务中心 北京 102600
- 折叠
摘要
目前河道管理范围内的"四乱"监管对象的排查主要依赖于人工巡查和遥感影像目视解译,成本高、效率低且问题发现能力一般.采用智能识别技术成为清"四乱"常态化的关键,然而河湖"四乱"量大类杂、分散隐蔽,导致传统模型鲁棒性低,识别精度一般.为此,以北运河河道管理范围内的坑塘养殖、堆放问题、农业大棚及临河房屋等4种地类作为研究对象,提出基于Transformer模型的全局局部一体化多尺度细节智能遥感深度学习框架,实现了自动适应多类型、多尺度、多级别的大规模遥感影像的智能识别.同时通过空间叠加分析技术,高效聚焦"四乱"增量问题,为水务监管工作提供技术支持.结果表明,在北运河流域河湖"四乱"监管对象的预测识别中,该模型的MIoU和MPA分别达到了 80.42％和85.91％,满足水务工作需求,并可为相关问题提供借鉴价值.
Abstract
The inspection of the"Four Illegal Behaviors"within river management areas mainly relies on manual patrols and visual interpretation of remote sensing images.This approach results indicated with high costs,low efficiency,and limited problem detection capabilities.Therefore,the key approach to es-tablish regular and effective management of the"Four Illegal Behaviors"lies in the adoption of intelligent recognition technology.However,the"Four Illegal Behaviors"around rivers and lakes are diverse,wide-spread,and covert,which led to low robustness and moderate accuracy in traditional models.To address these challenges,this research focused on four land categories within the management scope of the North Canal River including pond breeding,stacking problems,agricultural greenhouses,and riverfront houses.A comprehensive and localized integrated multi-scale detail intelligent remote sensing deep learning framework was proposed based on the Transformer model.This framework enabled intelligent recognition of large-scale remote sensing images,adapting to multiple types,scales,and levels automatically.Addi-tionally,spatial superposition analysis technology was utilized to efficiently target on the incremental problems associated with the"Four Disruptions,"which provide technical support for water management supervision.These results demonstrated that the proposed model achieved an MIoU(Mean Intersection over Union)of 80.42％and an MPA(Mean Pixel Accuracy)of 85.91％in the prediction and identification of supervised objects related to the"Four Illegal Behaviors"in rivers and lakes within the North Canal Ba-sin.These findings will meet the requirements of water management tasks and offer valuable insights for related issues.
关键词
智能识别/河湖四乱/Transformer神经网络/北运河流域Key words
intelligent identification/"Four Illegal Behaviors"of river and lake/Transformer neural network/North Canal basin引用本文复制引用
基金项目
2023年度北京水务科技开放项目(SK-2022-001-DSJ)
出版年
2024
国家科技期刊平台
NSTL
万方数据