影像處理技術應用於河床粒徑分析及魚類數量調查

Abstract

本文目的為應用影像處理技術於河川生態環境之調查。設法將 (1) 河床質調查和 (2) 溪魚監測問題自動化，鑒於傳統調查方法往往費時費力，如河床質調查最常採用的體積法需要60-120分鐘於野外進行採樣與初步篩分析，接著超過30分鐘於研究室進行複篩才可完成一個樣本；電魚法為溪魚生態調查最常採用的方法，缺點為個人努力量或是選取採樣點的不同，會造成不同的調查結果，且調查結束後常會伴隨魚類死亡，兩種傳統調查方法的共同點皆是費時費力和侵擾原有的生態環境，如河床表面破壞與溪魚死傷，近年來數位攝影設備與軟體分析進步神速，造就數位影像處理(Digital image processing)迅速發展成一門進步的數位技術，無論在量測、辨識與移動偵測方面，其優越的表現似有極大潛力發展自動化河床質量測與溪魚監測系統；自然河床的變異如礫石大小、顏色，甚至於細小泥砂所帶來的雜訊，與現場拍照時光源控制不易等許多因素，為目前研究所遇到的瓶頸，故研究鎖定河床質相關文獻中常被應用的AGS(Automated grain sizing)法，簡化需要調整的參數，結合反傳遞模糊類神經網路(Counterpropagation fuzzy neural network, CFNN)，提出一套可依輸入影像，輸出適宜參數的R-AGS (Refined automated grain sizing)模式，研究以蘭陽溪流域總計130張照片為研究材料，先利用人工影像分析法，量測石頭最大短軸，並搜尋理想二元門檻值協助反傳遞模糊類神經網路學習，以目前文獻中兩種自動推估方法做比較，其顯著的改善率(約減少20%的均方根差誤差)與80%的篩分析準確率，可供後續研究與工程應用之參考。另一方面，目前並無太多文獻著墨於自動化溪魚監測之研究，故本研究利用水底攝影機，以背景相減法為基礎，提出一套自動化方法推估魚群數目，方法共分為3個部分: (1) 影像前處理提高影像判識品質、(2) 背景模型提供乾淨背景和 (3) 前景偵測找出會游動的魚；本案例以新北市金瓜寮溪為實驗場址，依照野外拍攝的經驗，提供儀器裝設的建議，由於夜間水域對於調查者的威脅遠超過白天，故本研究更進一步地嘗試儀器的夜視功能，探討夜間調查的可行性，實驗共進行兩天，錄製27筆，每筆約5分鐘的動態影像樣本進行分析，與人工判識相比的結果顯示，約80%的準確率證明自動化方法可提供長期且穩定的觀測資料、夜視功能具有夜間調查的實用性、溪魚較為喜好緩流流域型態和經由自動推估所繪出的溪魚足跡圖，可協助生態領域專家了解溪魚游動的行為；總結兩個案例的表現，本研究的自動化模式可協助水文、水利工程與生態領域專家後續研究之用。

The major purpose of this dissertation is to investigate river of ecological environment by using digital image processing techniques. To achieve this goal, two different cases which are (1) river material and (2) fish monitor have shown their accuracy and applicability. Traditional methods for measuring grain size distribution and investigating fish number are time-consuming and lab-intensive. Our experiences indicated volume-by-number would take approximeately an hour in field to collect a representative sample and an hour to conduct a grain-sieving process in laboratory. A lot of fishes would be damaged when shocking or fishing them. Those methods would influen the results and harm the river ecosystem. Recent advances in image processing techniques facilitate automated grain indenitication and fish monitor through digitial images. This study introduces a refined automated grain sizing method (R-AGS) incorporating a neural fuzzy network for automatically estimating the grain size distribution, specifically designed for digital images. A total of 130 images captured from Lanyang River bed, Taiwan, are used to assess R-AGS performance. Two prevalent image processing methods are implemented for comparative purpose. The results indicate that the proposed R-AGS significantly outperforms the other two comparative methods (i.e. reduced 20% of the root-mean square error). The second case is to count automatically the number of fish in pristine rivers. In this case, author presents a background termed as subtraction-based method in order to count the fish automatically. The experiment is performed in Jin-Gua-Liao River situated in the North of Taiwan and the underwater camera records the fish on the river during daytime and also during the night. To date there are little references on monitoring river fish. Author details the monitoring method of river fish and provide recommendations to implement this method. Our results show that: (1) our approach is approximate 80% accurate, (2) the studied river fish habitat is mainly in the glide environment, (3) the monitoring of fish at night is available, (4) the developed methodology provides not only long-term statistics but also delivers useful information on the fish behavior which is to assist hydrologists and ecological researcher in preserving the ecosystem. Summary, the results clearly indicate that our approach can be reliably used for automated grain-size measurement and fish number count with high accuracy and much less labor-intensive.