Detecção de invasões biológicas no cerrado utilizando deep learning

• Ano de defesa: 2019
• Autor(a) principal: Santos, Igor Araujo Dias
• Orientador(a): Vivaldini, Kelen Cristiane Teixeira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Aprendizagem de máquina; Cerrado; UAV; Segmentação semântica; Vegetação; U-Net
• Palavras-chave em Inglês: Deep learning; Drone; Fully convolutional networks; Autoencoders; Semantic segmentation; Data augmentation
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/12178

The Cerrado represents an important reserve of natural resources, with biodiversity representativity worldwide. On the other hand, biological invasions can threaten the balance and put in risk local species, in this way making it urgent to elaborate technological resources that may cooperate in the natural preservation and conservation process. The present study intends to use images from visual spectrum areas (RGB) collected by an UAV for autonomous detection of biological invasions in Cerrado, adopting techniques from Deep Learning. For getting the images, the UAV (Quadcopter) and the attached RGB sensor were chosen from their greatest accessibility and resulting reproducibility. The Convolutional AutoEncoder (CAE) and U-Net networks were adopted for being widely used in Dataset with a few samples, because of its capacity of generalizing, despite having few examples for the training. Therefore, an original Dataset was created from the study area using manual delineation and later the same basis was broadened with Data Augmentation technique. For analyzing the unchanged database, the Convolutional AutoEncoder network overcome the U-net one with an 88% F-score against 84%. With the second DataSet with Data Augmentation, the results were even better, with an 93% CAE F-score, compared with 84% from U-net and superior Precision on both scenarios (85.4% CAE and 82% U-net for original DataSet and 93% CAE and 84% with Data Augmentation). Those differences are relevant because of the necessity of precision in the results to correctly direct teams on their search tasks for biological invasions through the wide Cerrado territory. It also emphasizes CAE characteristics considering its smallest size, with a small number of layers and neurons, and with higher metrics for this application. Thus, it was possible to note that the predictive model generated by AutoEncoder Network can be used efficiently, with great potential for other databases. Finally, it is concluded that this paper represents the Machine Learning progress and its capacity of assisting daily life, expanding the possibilities of future works.