Analysis of the Efficiency of the Taxonomic Identification of Small Fishes Using Artificial Neural Networks: A CASE Study of the Ichthyofauna of the Carajás Mountain (Pará—Northern Brazil) ^†

Abstract

The development of techniques that assist in the processes of taxonomic identification is of utmost importance, considering the scarcity of specialists and literature available in remote and diverse areas. Environmental studies such as the Biodiversity Monitoring Program of the Carajás National Forest in northern Brazil (FLONA de Carajás—PA, 6°6′29′′ S, 50°18′16′′ W) face challenges in this regard. These challenges include the particularity of the morphological and evolutionary characteristics of the fauna, present in a very diverse area of intense anthropogenic intervention by the use of resources of economic interest. Thus, this work sought to analyze the efficiency of using Artificial Neural Networks (ANN), more specifically the “XCeption” algorithm, configured for the taxonomic identification of samples captured during this monitoring program. These samples were previously identified using traditional taxonomic identification keys. The taxa Aequidens tetramerus, Astyanax abramis, Bryconops spp., Knodus spp., and Moenkhausia spp. were used. After capturing the images, the content was assigned to different folders, named “Training” and “Test”. This procedure seeks to quantify the model’s ability to classify data characteristically different from that presented in the training base. The accuracy results obtained during the training phase of the algorithm used, executed in about 16 hours, were 98% for the Training phase and 92% for the Validation phase, with some categories presenting better prediction results, such as classes 4 (100%) and 2 (85%). The testing phase, executed in about 1 hour, obtained an accuracy value of 78.26%, with a 95% confidence interval (63.64–89.05%) and a kappa of 70%. The applied methodology presented high accuracy, configuring itself as an important tool for identifying fish species in extremely diverse and remote environments.