Detecção de espécies arbóreas utilizando redes neurais convolucionais
Resumo
Resumo : Avaliações precisas da composição das espécies florestais podem ser utilizadas para analisar padrões de biodiversidade, estimar estoques de madeira e melhorar as estimativas de carbono estocado. Portanto, o objetivo deste trabalho foi a avaliar a utilização de imagens de alta resolução obtidas por Unmanned Aerial Vehicle (UAV) na identificação de espécies florestais em áreas de regeneração florestal na Amazônia. As redes neurais convolucionais (CNN) foram treinadas utilizando o pacote Keras-Tensorflow com o modelo faster_rcnn_inception_v2_pets. Amostras de seis espécies florestais foram utilizados para treinar a CNN, a partir destas, foram realizados testes com o número de thresholds e identificação das espécies confrontando os resultados obtidos com dados obtidos em campo. Os resultados demostraram que a redução do threshold diminui a exatidão na identificação, assim como a sobreposição dos polígonos de identificação das espécies, no entanto, em comparação com os dados coletados em campo, observa-se que há uma alta correlação entre as árvores identificadas pela CNN e às observadas nas parcelas. As métricas estatísticas utilizadas para validar as classificações evidenciaram que a CNN é capaz de identificar as espécies com acurácia acima de 90%. Portanto, as redes neurais convolucionais é uma ferramenta eficaz na classificação de objetos em imagens UAV, demonstrando boa acurácia e precisão na identificação das espécies em estudo. Abstract : Precise assessments of forest species composition can be used to analyze biodiversity patterns, estimate wood stocks and improve stock carbon estimates. Therefore, the objective of this work was to evaluate the use of high resolution images obtained by Unmanned Aerial Vehicle (UAV) in the identification of forest species in areas of forest regeneration in the Amazon. Convolutional neural networks (CNN) were trained using the Keras-Tensorflow package with the faster_rcnn_inception_v2_pets model. Samples of six forest species were used to train CNN, from these, tests were performed with the number of thresholds and species identification comparing the results obtained with field data. The results showed that the reduction of the threshold decreases the accuracy of identification, as well as the overlap of the polygons of species identification, however, in comparison with the data collected in the field, it is observed that there is a high correlation between the trees identified by the CNN and those observed in the plots. The statistical metrics used to validate the classifications showed that the CNN is able to identify species with accuracy above 90%. Therefore, convolutional neural networks are an effective tool in classifying objects in UAV images, demonstrating good accuracy and precision in the identification of species under study.
Collections
- Data Science & Big Data [107]