| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 1.61 MB | Adobe PDF |
Authors
Abstract(s)
Esta dissertação tem como objetivo principal o desenvolvimento e validação de uma arquitetura baseada em Redes Neuronais Convolucionais (CNN, do inglês Convolutional Neural Network) para o reconhecimento automático da espécie invasora Acacia dealbata em ambientes florestais, recorrendo a imagens capturadas por Veículos Aéreos Não Tripulados (VANTs). A Acacia dealbata é uma árvore que possui um crescimento rápido, originária do sudeste da Austrália, que pode atingir cerca de 30 metros de altura e distingue-se pelas suas flores amarelas. Para tal, foram definidos objetivos específicos, como a automação do pré-processamento de dados através de scripts dedicados à extração, segmentação e rotulagem de imagens de vídeos de VANTs.
Para o processo de desenvolvimento da CNN, foram exploradas diversas arquiteturas. Os testes iniciais com VGG16 revelaram limitações, especialmente devido ao elevado tamanho e quantidade de imagens semelhantes do dataset, que resultava em valores de Accuracy e F1-Score baixos. A transição para a arquitetura ResNet50V2 trouxe algumas melhorias com um desempenho geral superior. No entanto, o avanço mais significativo foi alcançado com o desenvolvimento de uma CNN criada de raiz, aliado a um dataset mais otimizado.
A avaliação do desempenho do modelo demonstrou a sua eficácia e robustez. O modelo final, obtido dos testes com a CNN de raiz e o segundo dataset, apresentou uma capacidade de generalização na deteção da Acacia dealbata que foi confirmada pelos resultados da Accuracy de 0.8178 e F1-Score de 0.8459. Tais conclusões validam o potencial das CNNs na classificação de espécies em ambientes florestais e são uma contribuição significativa para a gestão ambiental.
No seguimento do estudo efetuado pode proceder-se à expansão e diversificação do dataset com novas gravações e em diferentes localizações geográficas e condições atmosféricas, visando também uma maior variedade de espécies. Poderá ainda ser explorada a adaptação do modelo para uma classificação multi-classe, permitindo a identificação de múltiplas espécies.
The main objective of this dissertation is the development and validation of an architecture based on Convolutional Neural Networks (CNNs) for the automatic recognition of the invasive species Acacia dealbata in forest environments, using images captured by Unmanned Aerial Vehicles (UAVs). The Acacia dealbata is a fast-growing tree, native to southeastern Australia, which can reach up to 30 meters in height and is distinguished by its yellow flowers. To this end, specific goals were defined, such as automating data preprocessing through scripts dedicated to the extraction, segmentation, and labeling of images from UAV videos. For the CNN development process, several architectures were explored. Initial tests with VGG16 revealed limitations, particularly due to the large size and number of similar images in the dataset, which resulted in low Accuracy and F1-Score values. The transition to the ResNet50V2 architecture brought some improvements, with overall better performance. However, the most significant progress was achieved with the development of a CNN built from scratch, combined with a more optimized dataset. The performance evaluation of the model demonstrated its effectiveness and robustness. The final model, obtained from the tests with the custom CNN and the second dataset, presented a strong generalization ability in detecting Acacia dealbata, confirmed by an Accuracy of 0.8178 and an F1-Score of 0.8459. These results validate the potential of CNNs in species classification in forest environments and represent a significant contribution to environmental management. Following this study, future work may include expanding and diversifying the dataset with new recordings in different geographic locations and under various atmospheric conditions, also aiming at a wider variety of species. Moreover, adapting the model for multi-class classification could be explored, enabling the identification of multiple species.
The main objective of this dissertation is the development and validation of an architecture based on Convolutional Neural Networks (CNNs) for the automatic recognition of the invasive species Acacia dealbata in forest environments, using images captured by Unmanned Aerial Vehicles (UAVs). The Acacia dealbata is a fast-growing tree, native to southeastern Australia, which can reach up to 30 meters in height and is distinguished by its yellow flowers. To this end, specific goals were defined, such as automating data preprocessing through scripts dedicated to the extraction, segmentation, and labeling of images from UAV videos. For the CNN development process, several architectures were explored. Initial tests with VGG16 revealed limitations, particularly due to the large size and number of similar images in the dataset, which resulted in low Accuracy and F1-Score values. The transition to the ResNet50V2 architecture brought some improvements, with overall better performance. However, the most significant progress was achieved with the development of a CNN built from scratch, combined with a more optimized dataset. The performance evaluation of the model demonstrated its effectiveness and robustness. The final model, obtained from the tests with the custom CNN and the second dataset, presented a strong generalization ability in detecting Acacia dealbata, confirmed by an Accuracy of 0.8178 and an F1-Score of 0.8459. These results validate the potential of CNNs in species classification in forest environments and represent a significant contribution to environmental management. Following this study, future work may include expanding and diversifying the dataset with new recordings in different geographic locations and under various atmospheric conditions, also aiming at a wider variety of species. Moreover, adapting the model for multi-class classification could be explored, enabling the identification of multiple species.
Description
Keywords
Redes neuronais convolucionais classificação binária Acacia dealbata VANTs Machine learning Pré-processamento de dados