Monitoramento inteligente de peixes amazônicos: detecção e classificação com aprendizado profundo em passagens de peixes

Abstract

The Amazon Basin has one of the largest hydroelectric potentials in the world and is responsible for a significant portion of energy generation in Brazil. The construction of hydroelectric projects in the region, such as the Belo Monte Hydroelectric Complex, aims to meet the growing energy demand but may also impact migratory dynamics and the conservation of Amazonian fish fauna. Therefore, the development of automated monitoring systems to assess the effectiveness of mitigation structures, such as fish passages, becomes essential. This study presents the development of an automated monitoring system for the detection and classification of fish species in the fish passage at the Pimental dam, which is part of the Belo Monte Hydroelectric Complex. The research was conducted in the fish transposition system (FTS) of the Pimental dam, using computer vision techniques. To build the dataset, frames were extracted from underwater videos captured by the FTS and subsequently manually annotated on the V7 platform. The resulting database consisted of 1000 images, divided into training (700), validation (150), and test (150) sets. Species were selected based on their frequency of occurrence and migratory importance, with emphasis on Phractocephalus hemioliopterus and Cichla melaniae, among others. Modeling was performed using Convolutional Neural Networks (CNNs), implemented in the YOLO v8 model, known for its efficiency in image detection tasks. Data augmentation techniques were applied to expand the diversity of the training set, introducing transformations such as rotations, translations, scaling, and brightness adjustments. Training was conducted on the Google Colab PRO platform using an NVIDIA A100 GPU, ensuring high performance in image processing. During the process, parameters such as learning rate (0.01), momentum (0.937), and weight decay (0.0005) were adjusted to minimize overfitting and improve model generalization. The model was evaluated using metrics such as precision, recall, F1-score, and mean Average Precision (mAP). The results indicated superior performance for species more represented in the dataset, such as Phractocephalus hemioliopterus (F1-score of 91%) and Cichla melaniae (87%). Less frequent species showed lower classification accuracy, such as Leporinus friderici (52%) and Leporinus sp2 (55%).Learning curves showed a progressive reduction in training and validation losses, demonstrating the model’s ability to recognize visual patterns of the species. The model maintained consistent performance under different environmental conditions, including high turbidity and artificial lighting reflections, reinforcing its potential for continuous monitoring of aquatic biodiversity. However, some limitations were identified, such as seasonal variability in image quality and the low representativeness of certain species, which may compromise model generalization. Additionally, processing time and the need for robust computational infrastructure are factors to be considered. The implementation of this system at the Pimental dam, within the Belo Monte Hydroelectric Complex, represents an advancement in the evaluation of environmental impact mitigation structures, providing essential information for the sustainable management of aquatic fauna in large hydroelectric projects.