Integração de redes neurais convolucionais e modelos de distribuição de espécies via computação evolucionista para a classificação de animais no monitoramento da saúde da fauna silvestre

Abstract

Citizen science projects play an important role in monitoring biodiversity and zoonoses. Such projects allow citizens to report sightings of wild animals, along with images and contextual information such as location, date, and time. In Brazil, the Wild Health Information System (SISS-Geo) uses these records to support epidemiological surveillance, relying on them to issue alerts regarding potential zoonotic outbreaks. However, the high volume of data submitted to the system makes manual auditing a slow and costly process. Image-based classification models and Species Distribution Models (SDMs) allow for the exploration of the records’ images and metadata, respectively. Although they can be used individually, their integration allows these information sources to be explored in a complementary manner. Thus, this study proposes an approach for classifying SISS-Geo records by integrating classification models based on images and location metadata of the records. The work’s methodology is divided into two main fronts. The first consists of advancing preliminary studies, bringing the experimental design closer to real-world usage conditions of the system through record-level classification, applying Test-Time Augmentation (TTA) to image classification models, and utilizing other SDMs built from metadata, evaluating model combinations via Genetic Algorithms (GAs). The second consists of investigating the use of Genetic Programming (GP) as an alternative to GAs for combining individual model predictions, aiming to produce superior ensemble models capable of inferring interpretable classification rules. A convolutional neural network with a ResNet152 architecture was trained, along with species distribution models BIOCLIM, MaxEnt and Generalized Linear Model (GLM), as well as a model referred to as SDM Mean, obtained by averaging the predictions of the three aforementioned models. Experiments were conducted using real SISS-Geo records, comparing the performance of individual models, GA-based combination strategies, and the proposed GP-based approach. ResNet152 achieved an average accuracy of 53.84% and an F1-Score of 49.63%, which increased to 55.29% and 51.65%, respectively, with the application of TTA. Meanwhile, the SDMs achieved a maximum performance of 22.00% accuracy and a 13.42% F1-Score. The best GA-based ensemble model was between ResNet152 with TTA and the SDM Average, achieving an accuracy of 60.97% and an F1-Score of 55.93%, while GP reached 63.82% and 58.08%, respectively, for the same model pair. It is demonstrated that incorporating metadata into image-based classification models improves the classification performance of SISS-Geo records, and that GP outperforms the GA-based approach in combining individual models, producing ensemble models with higher accuracy and interpretable classification rules.