Image classification using two neural networks and activation functions: a case study on fish species

Abstract

Lake Toba is utilized for aquaculture fishing as a clear example of how this technology can be applied. One of the species presents is the red devil fish (Amphilophus labiatus), which is known to have started appearing in the last 10 years. This species is known to be very aggressive and damage the ecosystem. When their populations go unchecked, red-devils can cause a decline in local fish populations, potentially destroying the balance of the food chain in those waters. This study used artificial neural network (ANN) and convolutional neural network (CNN) algorithms to successfully create two classification models for fish species from Lake Toba: red devil fish (Amphilophus labiatus), mujahir fish (Oreochromis mossambicus), sepat fish (Trichogaster trichopterus). The purpose of this model is to automatically identify fish species by using image-based classification techniques. According to the study's findings, both models performed exceptionally well and had a high degree of accuracy. This study addresses the lack of effective automated fish classification systems for ecosystems like Lake Toba, Indonesia, which are threatened by invasive species such as the red devil fish. By comparing CNN and ANN models with different activation functions and optimizers, we found that CNN with rectified linear unit (ReLU) activation and Adam optimizer provides the most accurate and stable results. The findings offer practical implications for fisheries management and biodiversity conservation.