基于神经网络的图像水位自动测量方法研究

Automatic image water level measuring based on neural network

基于图像的测量站可以有效加密河流水位监测网络,但由于河流在一年中可能会呈现出不同的外观,所以利用图像无法得到精确的水位测量结果。为此,本文介绍了一种结合深度学习和摄影测量技术的方法,通过图像自动获取准确水位。首先,本文利用深度学习中卷积神经网络(CNN)的全卷积神经网络(FCN)和语义分割(SegNet),对树莓派像机采集的图像进行水域分割,两种卷积神经网络的水域分割精度均高于98%;其次,将分割结果生成的水边线与无人机倾斜摄影测量得到的数字高程模型(DEM)相交,将图像信息转化为米制水位值。标准水位计得到的参考水位与基于图像方法测量的水位之间的相关性最高可以达到0.947,平均偏差最小仅为1.2 cm。本文方法实现了对河流水位监测网络的加密,并提供了准确的水位测量结果。

Image-based gauging stations can effectively densify river level monitoring network.However,accurate water level measurements cannot be obtained by using images due to the reasons that the river may have different appearances during the year.Therefore,a method combining deep learning and photogrammetry is introduced in this paper to realize automatic and reliable water level measuring.Firstly,SegNet and FCN in the convolutional neural network in deep learning were used to segment the water areas collected by Raspberry Pi camera,and the accuracy from the two convolutional neural networks is better than 98%.Secondly,the water boundary line generated by the segmentation is intersected with the digital elevation model obtained by UAV tilt photogrammetry,and the image information is converted into metric water level.The correlation between the reference water level from the standard water gauge and that measured by the image-based method can be up to 0.947,with a minimum average deviation of only 1.2 cm.The method introduced in this paper realizes the densification of river level monitoring network by using cameras,and it can provide accurate water level measurements.