利用無人機攝影測量進行河床表面粒徑調查

Abstract

河道沉積物粒徑是了解河流作用的重要元素之一，除了反映侵蝕、搬運、沉積的狀態之外，在水文模型的計算中也有重大的影響，例如曼寧係數、河流水力侵蝕模型等。此外，在颱風、土石流等事件過後，河道沉積物粒徑的變化也有助於我們評估事件帶來的影響。作為一個重要的地形特徵，如何有效地蒐集準確的粒徑資料也成為值得探討的議題。過去粒徑調查多以現地採樣為主，然而人為操作下的誤差可能會使採樣結果略大於實際粒徑分布，且需要在野外工作較長時間。隨著遙測技術與電腦科學的發展，近年來不乏有使用無人機空拍，並結合機器學習以達到影像自動判釋粒徑的研究，惟前人研究多以提出新的粒徑判釋方法為主，較少針對現有的方法進行比較，尤其是以點雲、正射影像作為判釋的對象時，在資料本質上便有二維與三維的差異。本研究中首先使用無人機在河道低空飛行拍攝毫米級解析度影像，透過運動回復結構與多視立體產出點雲，並且得到相機參數後將原始影像校正為正射影像，以不同的程式由這兩種資料判釋粒徑，再與現地測量的粒徑資料進行比較。以使用5公尺航高拍攝、解析度1.88 mm/pix的影像產製出的點雲與正射影像進行分析，在D50 < 15公分的區塊，影像辨識可以獲得與現地測量相近的結果，RMSE約為D50的10-15%，最小可以辨識的粒徑約為8 mm；點雲則是僅在D50 ~ 10-15公分、礫石緊密排列下有較好的分割結果，在細顆粒與粗顆粒中分別有欠分割（under-segmentation）和過度分割（over-segmentation）的現象。而在數十公分的大顆粒當中，透過影像和點雲測得的結果皆與現地測得的粒徑有較大差異。除了礫石本身大小之外，彼此之間的排列情形也是影響分割的因素。此外，在不同時期的資料中，透過大範圍的影像辨識和傳統採樣方式獲得的粒徑分布相比，前者可以觀察到更細微的變化。這些不同方法之間的誤差與限制可以作為後續粒徑調查研究之參考。

Sediment grain size in stream beds is an important factor for understanding fluvial process. Grain size is one geomorphological feature of a river reflecting river dynamics, such as transport and deposition processes, stream power, flow history and the origin of sediments. Besides, the change of grain size after natural disaster, such as typhoon and landslide, is another issue worth to concern. Therefore, how to acquire grain size data efficiently and accurately is an important work. In previous studies, high resolution image-based grain size analyses are limited in patch scale, while site scale studies are limited in resolution to few centimeters. On the other hand, grain size detected in images is the “2D” grain size, and the difference between 2D grain size and real 3D grain size are rarely discussed. In this research, UAV (Unmanned Aerial Vehicle) is used flying at a low altitude to capture high resolution images in gravel- and cobble-bed rivers. During the process of structure-from-motion photogrammetry and multi-view stereo algorithm, point cloud and orthoimages are generated firstly, then grain size is detected from these two outputs respectively. Using point cloud and orthoimage generated from images captured at 5-m flight height, with resolution = 1.88, as input data to analysis, in patches with D50 < 50 cm, grain size detected from image is close to grain sized measured in the field. RMSE is 10-15% of D50 approximately, and the minimum grain axis length can be identified is about 8 mm. As for point cloud, a preferred detection results only showed in patch with D50 ~ 10-15 cm, and grains are tightly arranged. In patches with finer or courser grains, under-segmentation and over-segmentation are happened respectively. However, both image and point cloud detected grain sizes have a larger difference for grains in few decimeters large. In addition to the size of grains itself, the arrangement among grains will affect the result of segmentation in point cloud. Besides, in different period of data, grain size distribution derived from image detection is able to catch more detailed changes comparing with traditional sampling method, which is able to serve as a reference for future studies on particle size analysis using imagery and point cloud.