以類神經網路判別台灣沿近海漁船VDR資料庫之作業狀態

隨著漁船監控系統(Vessel monitoring system)普遍被應用於各種漁業之漁船，漁船船位資訊(vessel position information)被運用於估計努力量、漁獲狀況，並成為漁船管理之利器。我國沿近海漁船數量總數接近二萬艘，其動態不易掌控，自 2007 年農業委員會漁業署開始要求該等漁船安裝漁船航程記錄器(Voyage Data Recorder,VDR)，使得漁船作業位置得以被掌握。本研究目的希望將龐大的VDR資料，能夠正確轉換為有效努力量，作為未來漁業動態分析與漁業管理之用。本研究以 2011年台灣沿海60 艘漁船(包含延繩釣、拖網、扒網等漁法)VDR資料庫以及對應之漁獲報表資料，利用類神經網路(Artificial Neural Network,ANN)為分析工具，採用 MATLAB軟體，於樣本船中挑選確實作業之作業時間、日期相對應的VDR 資料庫作訓練資料庫、將速度、航向、加速度、漁業別等列為可能變數，調整神經網路參數(主要包含神經元數量、訓練次數、均方差極值)以達到性能最佳化，並分析各項漁業的作業模式。主要目標有二，一為判斷各VDR紀錄點的作業狀態，另一為判斷各航次的作業漁法為何。結果顯示在參數上神經元最佳為3個、均方差極值最佳值為0.02，訓練次數差異不大。而在判斷作業狀態的結果顯示，拖網作業狀態的判別率為77%-90%，扒網作業判別率為88%，延繩釣判別率為77%-90%。在漁法判別上拖網為69%-91%，扒網為72%，延繩釣為53%-90%差異最大。使用類神經網路在判別作業點的準確度有不錯的正確率，但在漁法判別上則出現較大差異。這三種漁業的作業模式分析顯示，延繩釣ct0、ct1實際作業時間比例為67%，ct2為60%，拖網ct2為80%、ct3為85%、ct4為74%，扒網ct3為50%。在不同時間間隔上，判斷作業狀態部分，CT2以上漁船可採用30分鐘一筆，可獲得足夠準確度，且可省去計算時間，CT2以下選擇3分鐘一筆較佳，可獲得較精準的結果。未來如能持續針對參數進行最佳化設置及取得更準確的資料來源，甚至獲得觀察員資料，可增加辦別的可信度。

關鍵字: 漁船監控系統，漁船船位資訊，船航程記錄器，類神經網路，延繩釣，拖網，扒網

ABSTRACT

Along with the VMS (vessel monitoring system) has been implemented in many fisheries, vessel position information were used in estimating effective effort and catch and become one of the best tools for vessel management. There are around twenty thousand fishing vessels operating in coastal and offshore waters of in Taiwan. It’s difficult to monitor their fishing activities. Since 2007, it is mandatory for the vessels to install VDR (voyage data recorder) under the request of Fisheries Agency, Council of Agriculture. VDR data provide real fishing positions for those fleets. This research aims to identify the efforts efficiently through VDR data for management purposes. Sixty vessels’ VDR data, including 28 longline,14 trawl vessels, and 8 Taiwanese purse seiners. Their logbooks and VDR data were used for analysis. The Matlab is applied for using ANN(Artificial Neural Network) for identify fishing effort. The fishing time, date and correspondent logbook were selected as training database. The speed, degrees, heading, acceleration, types of fisheries are selected as variables and the parameters (number of neurons, training frequency, mean squared error) would be adjusted to maximize the performance, and explored for the fishing pattern. There are two objectives of this study, to identify the vessel operating status and gear types used in each voyage. The results showed the best number of neurons is 3, mean squared error is 0.02, training frequency have no significant differences. The correct rates for longline fishery were 77%-90% ,77-90% for the trawl fishery have,and 88% for the Taiwanses seine fishery. For gear type identification, 52%-90% of longline fishery,69%-91% trawl fishery, and 72% Taiwanese seine fishery could be identificed correctly. The results showed the fishing time were 67%, 67%, 60% for CT0 to CT2 longline fishery vessels respectively, and 80%, 85%, 74% for CT2 to CT4 trawl vessels respectively. It is suggested to use 30 min frequency data to identifying fishing status for vessels larger than CT2. It would have good performance and time saving. As for vessels smaller than CT2, three min interval would be needed. In conclusion, there is good performance for predicting fishing operations with ANN, but in gear identify needs to be improved. Continue to test for best parameters and collect more information, such as observers data could be useful to increase the correct rate.

Keyword: vessel monitoring system, Voyage data Recorder, Artificial Neural Network, long-line, trawl, Taiwanses seine