RECONHECIMENTO DE PADRÕES ESPAÇO-TEMPORAIS EM IMAGENS WorldView-2 E RPAs POR APRENDIZADO DE MÁQUINA EM FLORESTA OMBRÓFILA MISTA

• Autor(a) principal: Crisigiovanni, Enzo Luigi
• Data de Publicação: 2021
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do UNICENTRO
• Texto Completo: http://tede.unicentro.br:8080/jspui/handle/jspui/1725
• topic: Sensoriamento remoto; Aprendizado de máquinas; Ecologia da paisagem; Bioinvasão; Remote sensing; Machine learning; Landscape ecology; Bioinvasion; CIENCIAS AGRARIAS::RECURSOS FLORESTAIS E ENGENHARIA FLORESTAL; RECURSOS FLORESTAIS E ENGENHARIA FLORESTAL::MANEJO FLORESTAL

Habitat loss and bioinvasion are the main causes of species extinctions associated with anthropic processes. One of the main tools for the Spatio temporal analysis of these processes is remote sensing. With the focus on recognizing patterns in the Mixed Ombrophylous Rainforest (MOF) at different spatial and temporal scales, three chapters were sought: (1) Classify and measure elements of the landscape of a hydrographic sub-basin in the Mixed Rainforest in Center-South Paraná state, associating the landscape components to the phytosociological parameters of the tree community; (2) Quantify the recognition rate of Hovenia dulcis Thunb. (Japanese raisin tree), invasive alien species and Araucaria angustifolia (Bertol.) Kuntze (Araucária), native conifer threatened with extinction, by machine learning algorithms; and (3) Improve the extraction of attributes for the classification of H. dulcis, in a pixel-by-pixel classification applied in a One-class-Classification (OCC) model. To achieve the objectives, the methods employed were: (1) For two WolrdVew2 images from the years 2012 and 2016, using only the multispectral bands of the sensor (2m spatial resolution), multiresolution segmentation, and Support Vector Machine (SVM). Nine landscape metrics were extracted and submitted to a Principal Component Analysis (PCA) from the most accurate classifications. These components were associated with regression models generated by Generalized Estimation Equation (GEE) for nine phytosociological parameters calculated from nine permanent forest inventory plots, with three measurements 2011, 2014, and 2017 distributed in the area of the basin in question; (2) For the recognition of A. angustifolia and H. dulcis, the visual interpretation and clipping of the treetops validated by RPA images (Remoted piloted aircraft) was performed in the 2016 WorldView-2 image fused (0.5 m spatial resolution). The cutouts had textural attributes extracted by Edge Filter and Pyramid Histogram of Oriented Gradients (PHOG), the textural attributes, and the images' pixels' spectral values were submitted to data mining using a genetic algorithm. Finally, they were submitted to two classifiers, Artificial Neural Networks (ANN) and Random Forest (RF); and (3) In two 2018 RPA images, random pixels from H. dulcis and the negative class (non-H. dulcis) were selected in a balanced way. Then, OCC models classified pixel-by-pixel were created by the Random Forest algorithm, using only the pixels' spectral attributes for the classification. In general, approximately 90% accuracy was found for all classifications used, with the different classification techniques applied to different images and targets. Concerning Chapter 1, there was a reduction in the area of MOF coverage of 6.4% and an agricultural expansion of 8.07% in the four years of analysis, mainly due to the increase in the landscape fragmentation process. It was still possible to find associations of fragmentation in the tree community, which showed changes mainly related to the increase in biomass, indicative of regeneration processes. Chapter 2, on the other hand, species classification was performed with high accuracy, reaching 95% of correct answers for Cross-Validation. RF surpassed the ANN classification rates and still proved to be more stable and faster for training and later species classification. For chapter 3, the classification of RPA images was also highly accurate, with around 95% correct answers. With only three spectral bands: Red (R), Green (G) and Blue (B), the spectral attributes of the red and green range were more relevant for the classification, and the pixel-by-pixel classification model of only one class proved efficient in the detection of H. dulcis. It was concluded that the WorldView-2 multispectral sensor and images of RPAs, associated or not, have a high potential to provide information for recognizing MOF forest patterns, ranging from landscape elements, e.g. forest remnants, to tree species canopies, such as H. dulcis and A. angustifolia. Thus, it can be effective in contributing to the conservation and sustainable management of the MOF.