基于无人机航拍的河道施工水污染图像智能识别与定位

Intelligent image recognition and positioning of water pollution in river regulation projects based on UAV aerial photography

摘要

小流域河道整治中施工废水等会给河道水体带来污染，且工区往往沿长线性河道呈点状分布形态，交通不便，人工巡检难以保证施工水污染管控的时效性。目前通常采用智能巡检手段对水面漂浮物进行图像识别，但对于污水、水华等水污染的识别仍存在困难。针对上述问题提出基于无人机航拍与深度学习的河道施工水污染异常图像智能识别与定位方法。利用无人机巡检采集工区航拍图像，基于SENet注意力机制优化的MBConv模块，建立了施工水污染图像的智能快速识别分类EfficientNet-B0模型，并利用迁移学习方法进行模型训练，可以使模型提取的污染相关特征指向性更强，提升模型的训练速度及精度；同时，基于Grad-CAM(Class Activation Mapping,类激活热力图)方法和全连接条件随机场(Conditional Random Fields, CRF)方法得到的污染区域特征图，可修正错分区域，实现更为精细的污染区域快速标记定位。实例应用表明，水污染图像分类准确率可达98%,追踪标记评价指标ORI值可达96.91%。结果表明：研究成果能够为工程管理人员快速管控长线性河道整治工程的施工水污染提供了先进的技术手段。

Abstract

Construction wastewater of small watersheds regulation projects cause river pollution. As a long-line project, river regulation often shows a scattered distribution, with inconvenient transportation, manual inspection is difficult to guarantee the efficiency of construction water pollution control. At present, intelligent inspection is usually used to identify floating objects, which is difficult to identify water pollution such as sewage and water bloom. This paper proposed an intelligent recognition and positioning method for anomaly images of water pollution in river regulation projects based on UAV aerial photography and deep learning. The aerial images of working area are collected by UAV inspection, and an EfficientNet-B0 model of intelligent identification and classification of construction water pollution image is established based on the MBConv module optimized by the SENet attention mechanism. Transfer learning is used to train the model, which can make the pollution-related features extracted by the model more directional, and improve the training speed and accuracy of the model. Then, the feature map of pollution area based on Grad-CAM(Class Activation Mapping) and fully connected CRF(Conditional Random Fields) can modify the misclassification area that achieve more refined fast marking and positioning of pollution area.The practical application result show that the classification accuracy of water pollution images is 98% and the evaluation index ORI of the positioning method is 96.91%. This study can provide advanced technical means for project managers to quickly control the water pollution of long-line river regulation projects.

关键词

河道整治工程 / 施工水污染 / 图像识别 / 图像定位 / 无人机航拍 / 特征可视化

Key words

river regulation / construction-caused water pollution / image recognition / image positioning / UAV aerial photography / feature visualization

Author summay

侯建刚(1971—), 男, 正高级工程师, 学士, 研究方向为水利水电工程。E-mail: [e-mail]

基金资助

中国长江三峡集团有限公司企业科研项目(202103551)