Classification of Nile Tilapia Freshness Levels Using the Convolutional Neural Network Method

Indonesia has significant potential in fisheries, with national production in 2011 reaching 12.39 million tons, consisting of 5.41 million tons of capture fisheries and 6.98 million tons of aquaculture. Freshwater fish production has increased by approximately 11% per year, with Nile tilapia being one of Indonesia's staple foods. Based on interviews with traders, the freshness of Nile tilapia is assessed through changes in the color of the gills and eyes. This study aims to evaluate the performance of a Convolutional Neural Network (CNN) model with AlexNet architecture in classifying the freshness levels of Nile tilapia based on post-mortem phases, using metrics such as Accuracy, Precision, Recall, and F1 Score, implemented in an Android application. The dataset comprises images of the eyes, gills, and body of Nile tilapia categorized into phases Fresh, 4 hours later, 8 hours later, and Rotten. The CNN model was developed with data splits of 70:30, 80:20, and 90:10, using hyperparameters including batch sizes of 16, 32, and 64, 100 epochs, and Adaptive Momentum optimization with learning rates of 10-2, 10-3, 10-4, and 10-5. The best results were obtained with an 80:20 data split, a learning rate of 10-4, and a batch size of 64, achieving an accuracy of 0.9686, validation accuracy of 0.9162, loss of 0.0892, and validation loss of 0.2733. Precision, Recall, and F1 Score were all above 0.86. The model was integrated into an Android application using TensorFlow Lite and performed well in inference tasks.