Mapping successional forest stages and tree species in subtropical areas integrating UAV-based photogrammetric point cloud and hyperspectral data: comparison of machine and deep learning algorithms

Detalhes bibliográficos
Ano de defesa: 2019
Autor(a) principal: Camile Sothe
Orientador(a): Cláudia Maria de Almeida, Marcos Benedito Schimalski
Banca de defesa: Hermann Johann Heinrich Kux
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: eng
Instituição de defesa: Instituto Nacional de Pesquisas Espaciais (INPE)
Programa de Pós-Graduação: Programa de Pós-Graduação do INPE em Sensoriamento Remoto
Departamento: Não Informado pela instituição
País: BR
Link de acesso: http://urlib.net/sid.inpe.br/mtc-m21c/2019/11.21.08.24
Resumo: The use of Remote Sensing for successional stages and tree species mapping in (sub)tropical forests is a challenging task, due to high floristic and spectral diversity in these environments. Fortunately, in the latest decades, mankind has witnessed a remarkable advancement of space technologies targeted to monitoring forest resources, such as the availability of high spatial and spectral data and advanced classification methods. Besides providing high spatial and spectral resolution images, unmanned aerial vehicle (UAV)- hyperspectral cameras operating in frame format enable to produce tridimensional (3D) hyperspectral point clouds. This study investigated two major topics concerning the successional stages and tree species mapping in a subtropical forest environment in Southern Brazil: a) the use of UAVacquired hyperspectral images and UAV-photogrammetric point cloud (PPC) for the classification of successional stages, comparing these data with classifications using multispectral images acquired by the WorldView-2 (WV- 2) satellite and Light Detection and Ranging (LiDAR) data and; b) the use of UAV-acquired hyperspectral images and UAV-PPC for individual tree crown (ITC) delineation and semiautomatic classification of 16 major tree species in two subtropical forest fragments. For both goals, different datasets containing hyperspectral visible/near-infrared (VNIR) bands, PPC features, canopy height model (CHM), and other features extracted from hyperspectral or WV- 2 data (e.g., texture, vegetation indices-VIs, and minimum noise fraction- MNF) were tested. To classify the successional forest stages, an objectbased image analysis (OBIA) was conducted using two conventional machine learning classifiers, support vector machine (SVM) and random forest (RF). For tree species classification, two conventional machine learning, SVM and RF, and one deep learning classifier, the convolutional neural network (CNN), were tested in a pixel-based approach. Besides these classifiers, a new SVM approach focused on an imbalanced sample set was also tested, the weighted SVM (wSVM). For ITC delineation, three methods were tested: two using hyperspectral bands, the multiresolution region growing (MRG) and the itcIMG, and the other one using the PPC, named multiclass cut followed by recursive cut (MCRC). The best segmentation result was used in two classification approaches tested using the conventional machine learning methods: OBIA and the majority vote (MV) rule. The results showed that the successional forest stages were successfully classified with accuracies over 80% when the WV-2 data were applied, and over 90% with the UAVhyperspectral data. The best result reached an overall accuracy (OA) of 99.28% using the hyperspectral data associated with the CHM and RF classifier. The CHM and features derived from WV-2 and hyperspectral data increased between 5% and 13% the classification accuracies. Regarding the tree species classification, the CNN outperformed the RF and SVM for both areas, with an OA of 84.4% in Area 1, and 74.95% in Area 2, using only the VNIR bands. This method was 22% to 26% more accurate than the SVM and RF when considering the VNIR dataset. The inclusion of PPC features and the CHM provided a great increase in tree species classification results when machine learning methods were applied (SVM, wSVM and RF), between 13% and 17% depending on the selected classifier and the study area. However, a decrease was observed when these features were included in the CNN classification. The OBIA approach did not increase the OA for the SVM classifier, while a slightly increase was observed for the RF algorithm in comparison with the RF using the pixel-based classification. The MV rule approach, on the other hand, brought a marked increase in accuracy for both study areas (5% for Area 1 and 11% for Area 2). When using PPC features and the CHM, associated with the MV approach, the machine learning classifiers reached accuracies similar to the ones achieved by the CNN (82.52% for Area 1 and 75.45% for Area 2). The wSVM provided a slightly increase in accuracy not only for some lesser represented classes, but also for some major classes in Area 2. None of the three ITC delineation methods reached a suitable result for all reference ITCs. The MRG method tended to oversegment most ITCs, while the itcIMG and MCRC tended to undersegment or missed some suppressed ITCs. With the inclusion of the CHM in the MRG segmentation and merging homogenous segments with the Jeffries Matusita (JM) distance, visually and according to supervised evaluation metrics, a better delineation was reached. The results found in this study are relevant to favor the conservation of the Atlantic Rain Forest, a severely threatened biome, optimizing the mapping and monitoring of its forest remnants, and also to subsidize actions within the scope of the rural environmental register (Cadastro Ambiental Rural- CAR) in Brazil. In addition, the methodology can be used to map specific tree species, such as the endangered ones, in this case Araucaria angustifolia and Cedrela fissilis.