Electronic Theses and Dissertation

Kemiripan morfologi antarspesies seperti tuna (thunnus), tongkol (euthynnus), dencis (sardinella), kembung (rastrelliger), tenggiri (scomberomorus), dan kuwe rambut (alectis) sering menyebabkan kesalahan identifikasi oleh masyarakat. kondisi ini menimbulkan perbedaan harga dan menurunkan kepercayaan konsumen, sehingga dibutuhkan sistem klasifikasi ikan otomatis berbasis citra digital. penelitian ini mengembangkan model klasifikasi ikan menggunakan resnet-50 dan mobilenet-v2 dengan pendekatan transfer learning. dataset dikumpulkan langsung dari pelabuhan lampulo dan terdiri atas 2.790 citra enam jenis ikan setelah proses augmentasi (rotasi, flipping, dan penyesuaian kontras). model dilatih selama 100 epoch dengan menggunakan dua jenis optimizer, yaitu stochastic gradient descent (sgd) dan adaptive moment estimation (adam), serta dua variasi learning rate, yaitu 10−4 dan 10−3. hasil penelitian menunjukkan bahwa performa terbaik dicapai oleh kombinasi sgd dan learning rate 10−3. dengan adanya penelitian ini diharapkan dataset citra enam jenis ikan yang dikumpulkan dan model klasifikasi berbasis deep learning dapat menjadi dasar pengembangan sistem identifikasi ikan otomatis.

Morphological similarities among species such as Tuna (Thunnus), Mackerel Tuna (Euthynnus), Sardines (Sardinella), Indian Mackerel (Rastrelliger), Spanish Mackerel (Scomberomorus), and Threadfish (Alectis) often lead to identification errors among the public. These inaccuracies result in price discrepancies and diminish consumer trust, highlighting the need for an automated fish classification system based on digital imagery. This study develops a fish classification model using ResNet-50 and MobileNet- V2 architectures through a transfer learning approach. The dataset was collected directly from the Lampulo Fishing Port, comprising 2,790 images of six fish species after undergoing augmentation processes (rotation, flipping, and contrast adjustment). The models were trained for 100 epochs using two types of optimizers—Stochastic Gradient Descent (SGD) and Adaptive Moment Estimation (Adam)—along with two learning rate variations: $0.01$ and $0.001$. The results indicate that the best performance was achieved by the combination of SGD and a learning rate of $0.01$. This research provides a primary image dataset of these six fish species and a deep learning-based classification model, which are expected to serve as a foundation for developing automated fish identification systems.