利用更快速區域類神經網路於河川流速量測之研究

Abstract

從工程規劃設計的面向來看，水資源工程規劃所需要的設計資料都是極端事件下的水文資料，使用傳統方式量測極端事件的水文資料就現實面來看還存在許多問題需要被解決。近年來利用非接觸式的影像方法測量水流表面流速，已經是發展的主流研究方法。本研究以大尺度粒子影像測速(Large Scale Particle Image Velocimetry, LSPIV)的概念為基礎，發展一套以深度學習架構為核心的影像測速法，同時提出一個全新的分析流程與概念進行流速量測。前期以人工水槽進行實驗來擷取大量的影像特徵，然後建置水流特性的影像資料庫來藉此訓練區域卷積神經網路(Faster Region-Convolutional Neural Network, Faster R-CNN)，同時使用都卜勒聲學儀器量測水槽流速來做為率定驗證的資料。透過Faster R-CNN網路辨識水流表面粒子並且加以定位，並且進行流速分析，將結果與LSPIV做比較，研究顯示利用Faster R-CNN得到的速度平均值高於LSPIV的速度平均值4%，且更加的接近透過聲學儀器量測的結果。本研究提出以深度學習架構為核心的方法成功地避免了LSPIV低估流速的問題，也解決CNN網路在流速定位上的不確定性，因此以Faster R-CNN網路為核心的影像測速可以有效避免人為介入所造成的不確定性，大大提高影像分析在水利應用的可能性。

There are lots of methods that can measure river discharge. It is a really important thing when allocating water resources. First of all, we need to obtain the depth-averaged velocity in the flow. Another common way is using LSPIV, which is image-based measurement. The advantage of this method is non-intrusive, that is, it doesn’t affect the surface velocity. It saves lots of time and have a high temporal resolution. However, many literatures mentioned that the velocity will be underestimated by using LSPIV. Because of the seeding density, illumination and set of interrogation area, these will affect the result. In recent years, the development of deep learning has made computer vision more powerful. Due to this reason, we apply the conception of convolution neural network for the surface flow measurement. Using Faster R-CNN (Faster Region-Convolutional Neural Networks) to detect and locate particles on water surface and caculate the water surface velocity. This method won’t be affected by illumination and no need to set interrogation area. In this paper, we compare the results of LSPIV and Faster R-CNN, and find out that the result of using Faster R-CNN is more close to ADV data. As mentioned above, we demonstrate that object detection with using deep learning in streamflow velocimetry is feasible.