The Comparison of Artificial Neural Network Models And Statistical Analysis In Determining The Types Of Freshwater Fishes Using Acoustic Descriptors

Abstract

Fisheries acoustic survey was one of holistic methods used to estimated the abundance of fish stocks to provide data and information for the fisheries management. Limitations of fisheries acoustic survey application that was in classifying the target backscattered acoustic energy (echo trace) into the classification of the target fishes in the species ranks. Therefore, it has developed a method of identification of fish species utilizing acoustic descriptors that can efficiently distinguish the structure of fish shoal. In this thesis, Hydroacoustic descriptor approach categorized as Volume Backscattering (Sv), Target Strength (Ts), Area Backscattering Strength (Sa), Skewness, Kurtosis, Height, Depth And Height Relative of Fish were used to classify Mas (Cyprinus carpio), Nila (Oreochromis niloticus), and Patin (Pangasius hypothalamus). Model of artificial neural network were developed utilized architecture Backpropagation and Multi Layer Perceptron compared with Statistical method. Results of Cluster analysis showed that the identification and classification of the carp was determined by the descriptors Height, Relative Height, Skewness and Kurtosis. Tilapia could be identified only by depth, whereas catfish classification determined by all parameters except depth. Discriminant analysis showed the results of the identification accuracy of 68.3% carp, tilapia of 79.4% and catfish could be identified with accuracy of 87.4%. Overall, discriminant analysis could distinguish three types of freshwater fish with a precision of 77.5%. Application of ANN with Backpropagation neural network model (8-30-1) obtained the optimum level of accuracy of the identification of three types of fishes at 84.8%. While the development of the Multi Layer Perceptron with ANN model (8-3-6-5-1) obtained the degree of accuracy of identification and classification of carp, tilapia and catfish at 87.5%. In this thesis concluded that the application and development of the Multi Layer Perceptron ANN gives the best accuracy rate compared with ANN Backpropagation and Statistical Analysis.

Survey akustik perikanan merupakan salah satu metode holistik yang digunakan untuk menduga kelimpahan stok ikan untuk menyediakan data dan informasi bagi pengelolaan sumberdaya perikanan. Keterbatasan aplikasi survey akustik perikanan yaitu dalam mengklasifikasi backscattered energy target akustik (echo trace) menjadi klasifikasi target ikan dalam tingkatan spesies. Oleh karena itu telah dikembangkan metode identifikasi spesies kawanan ikan dengan menggunakan parameter deskriptor akustik sehingga dapat membedakan secara efisien struktur dari kawanan ikan yang berbeda. Dalam tesis ini dilakukan identifikasi dan klasifikasi ikan menggunakan ikan uji yaitu ikan Mas (Cyprinus carpio), Nila (Oreochromis niloticus), dan Patin (Pangasius hypothalamus). Parameter deskriptor akustik yang diperoleh yaitu backscattering volume (Sv), target strength (TS), backscattering area (Sa), Skewness, Kurtosis, Tinggi, Kedalaman dan Ketinggian Relatif ikan. Permodelan Jaringan Saraf Tiruan dilakukan dengan mengembangkan arsitektur JST Backpropagation dan Multi Layer Perceptron yang dibandingkan dengan hasil Analisis Statistik menggunakan parameter masukan deskriptor akustik.