Emergent tree species identification in highly diverse Brazilian Atlantic forest using hyperspectral images acquired with UAV

Carregando...
Imagem de Miniatura

Data

2020-03-11

Autores

Miyoshi, Gabriela Takahashi [UNESP]

Título da Revista

ISSN da Revista

Título de Volume

Editor

Universidade Estadual Paulista (Unesp)

Resumo

O objetivo desse doutorado é propor uma nova metodologia para identificar oito espécies arbóreas emergentes (i.e., que se sobressaem do dossel florestal), em diferentes idades e estágios de desenvolvimento e pertencentes à Mata Atlântica brasileira. Para tal, imagens hiperespectrais foram adquiridas em Julho/2017, em Junho/2018, e em Julho/2019 em um transecto localizado no fragmento florestal Ponte Branca, localizado a Oeste do Estado de São Paulo, onde a floresta é considerada estacional semidecidual e submontana. As imagens com resolução espacial de 10 cm foram adquiridas com câmara hiperespectral (500–900 nm) acoplada em veículo aéreo não tripulado (VANT ou UAV, do inglês Unmanned aerial vehicle) e, posteriormente corrigidas geometricamente e radiometricamente. Em seguida, as copas arbóreas individuais (ITCs, do inglês Individual tree crows) foram delineadas manualmente em cada conjunto de dados para serem utilizadas como referência para os experimentos. Dentre os experimentos realizados, destaca-se o uso do espectro normalizado para redução da variabilidade espectral intra-espécies, o uso da classificação baseada em regiões utilizando o algoritmo Random Forest e o uso de superpixexls para delineamento automático das ITCs em cada conjunto de imagens. Além disso, avaliou-se o uso dos superpixels multitemporais com diferentes atributos multitemporais (espectro normalizado, textura e índices de vegetação) e estruturais (derivados do modelo de altura das copas), sozinhos ou combinados, para identificação das espécies arbóreas. O melhor resultado foi obtido a partir do uso combinado do espectro normalizado com os índices de vegetação, onde o valor da área sobre a curva característica de operação do receptor (AUCROC, do inglês Area under the receiver operating characteristics curve) atingiu valores de até 0.964. A partir dos resultados obtidos destaca-se o desafio ao trabalhar com esse tipo de floresta, devido à falta de árvores emergentes (que se sobressaem do dossel florestal), e a existência de árvores com diferentes idades e estágios de desenvolvimento, resultando em alta variabilidade espectral e estrutural para uma mesma espécie. Adicionalmente, foi verificado que o uso dos atributos estruturais e texturais não auxiliaram a tarefa de identificação de espécies e, que a alta resolução espacial das imagens mostrou as sutis diferenças de posição espacial das copas nas imagens dos diferentes anos. Por fim, apesar dos desafios, tem-se que os resultados são promissores e mostraram ser possível identificar espécies de árvores utilizando a informação multitemporal.
The objective of this doctoral dissertation is to propose a new methodology to identify eight emergent tree species (i.e., that stood out from the canopy) belonging to highly diverse Brazilian Atlantic forest and with different ages and development stages. To achieve the objective, hyperspectral images were acquired in July/2017, June/208, and July/2019 in a transect area located in the western part of São Paulo State. The area is in Ponte Branca ecological station, where the forest is classified as submontane semideciduous seasonal with different stages of succession. Images with a spatial resolution of 10 cm were acquired with a hyperspectral camera (500–900 nm) onboard unmanned aerial vehicle (UAV) and geometrically and radiometrically post-processed. In sequence, the individual tree crowns (ITCs) were manually delineated in each dataset to be used as reference in the experiments. From the performed experiments, it is highlighted the use of mean normalized spectra to reduce the within-species spectral variability, the use of region-based classification with the Random Forest algorithm, and the use of superpixels to automatically delineate the ITCs in each dataset. Additionally, the multitemporal superpixels with different multitemporal features (normalized spectra, texture and vegetation indexes) and structural features derived from the canopy height model, combined or not, were assessed to the tree species classification. The best result was achieved merging normalized spectra and vegetation indexes, where the value of area under the receiver operating characteristics curve (AUCROC) achieved values up to 0.964. From the obtained results it is pointed out the challenge when working with this type of forest due to the lack of emergent trees, which restrict the number of samples recognized in the field, and the existence of different ages and stages of development to the same tree species. Besides, the use of structural and textural features did not improve the tree species identification. Besides, the high spatial resolution of the images showed the slight differences in the spatial position of the tree crowns between the datasets. Finally, despite the challenges the results are promising and showed the feasibility to identify the tree species using multitemporal information.

Descrição

Palavras-chave

Tree species identification, Atlantic forest, Multitemporal spectral information, Superpixels, UAV, Identificação de espécies arbóreas, Mata Atlântica, Informação espectral multitemporal, VANT

Como citar