Identificação de espécies e estimativa de parâmetros dendrométricos na Floresta Amazônica
Resumo
Resumo: O Inventário florestal é um método de levantamento de campo indispensável para conhecer as características quantitativas e qualitativas de uma floresta. Esta atividade em florestas nativas é demorada, onerosa, risco a saúda do trabalhador, além da possibilidade de ocorrer possíveis erros de humanos voltados a identificação e medição das árvores. O uso de geotecnologia remotas através das aeronaves para levantamento e mapeamento da vegetação vem sendo um verdadeiro aliado aos gestores ambientais na execução de projetos de manejo e conservação florestal. Considerado o potencial de uso desta ferramenta, o objetivo deste estudo foi buscar avanços técnicos e científicos para apoiar o inventário florestal remoto na floresta Amazônica, apresentado em dois capítulos. O Capítulo I corresponde a fusão de imagens aéreas adquiridas por meio de um sistema de aeronave não tripulada visando a identificação automática de espécies arbóreas utilizando rede neural convulacional (CNN) na floresta Amazônica. O estudo foi realizado em uma área de floresta nativa com 260 ha em Rio Branco, na sede da Embrapa Acre, em que foram identificadas e segmentado oito espécies florestais com potencial econômico, totalizando 406 indivíduos, com DAP a partir de 50 cm. A classificação automática das espécies foi feita através da CNN ResNet-18 combinada com a arquitetura DeepLabv3+, modelo de aprendizado de máquina profundo, dividiu-se o banco de dados em 60% treinamento e 40% para teste. O melhor resultado foi na fusão das imagens aéreas de diferentes meses (período seco e chuvoso), a acurácia média foi de 90,5%, onde para seis espécies a acurácia ultrapassou 90%, a espécie Phyllocarpus riedelii obteve uma acurácia de 100%. O mês de maio teve a menor acurácia, 69,3%, intensa mudança de fenofases das espécies, mais difícil para o algorítmico discriminar as espécies florestais enquanto que o melhor mês foi o de novembro, acurácia de 83,5%. O capítulo II abordou sobre a estimativa do volume de árvores a partir da morfometria da copa obtidas por UAS na mesma área de estudo. Utilizou-se o inventário florestal de campo, onde todas as árvores com diâmetro a altura do peito - DAP >= 50 cm foram mensuradas. A localização das árvores foram combinadas com as imagens aéreas e, após, filtrados 388 indivíduos distribuídos em 55 espécies. A morfometria da copa foi identificada através do diâmetro médio e da área da copa, posteriormente realizou um teste de correlação para analisar o tamanho do relacionado destas variáveis para posterior inserção nos modelos de volume, área basal e diâmetro das árvores distribuído por classe diamétrica. O banco de dados foi dividido em 70% para o treinamento e 30% para o teste. O teste de correlação identificou uma forte interação entre as variáveis da copa com volume e a inexistência da altura, logo foi descartada da análise. O volume predito teve métricas para o ajuste SEE 21,97% e para o teste RMSE 19,13%, ambos com R² superior a 0,90. Para área basal e diâmetro o SEE foi inferior a 6,5% e R² maior que 0,99, os resíduos dos modelos foram heterocedáticos, o volume teve a maior dispersão, mas concentrou-se até ± 25% demonstrando a acurácia do modelo e o potencial em utilizar variáveis da copa das árvores. Abstract: Performing a forest inventory is an indispensable field survey method for knowing the quantitative and qualitative characteristics of a forest. This activity in native forests is timeconsuming, costly, poses a risk to the worker's health, in addition to the possibility of possible human errors in identifying and measuring trees. The use of remote geotechnology through aircraft for surveying and mapping vegetation has been a true ally to environmental managers in executing forest management and conservation projects. Considering the potential use of this tool, the objective of this study was to seek technical and scientific advances to support the remote forest inventory in the Amazon rainforest, presented in two chapters. Chapter I corresponds to a fusion of aerial images acquired through an unmanned aircraft system aiming to automatically identify tree species using convulational neural network (CNN) in the Amazon rainforest. The study was conducted in an area of native forest with 260 ha in Rio Branco, Acre, Brazil, at the headquarters of Embrapa Acre, where eight forest species with economic potential were identified and segmented, totaling 406 individuals, with DBH from 50 cm. The automatic classification of species was done through CNN ResNet-18 combined with the DeepLabv3+ architecture, a deep machine learning model, and the database was divided into 60% training and 40% for testing. The best result was in the fusion of aerial images from different months (dry and rainy seasons); the average accuracy was 90.5%, it exceeded 90% for six species, and the Phyllocarpus riedelii species obtained an accuracy of 100%. The month of May had the lowest accuracy (69.3%), with intense change in the phenophases of the species, making it more difficult for the algorithm to discriminate forest species, while the best month was November, presenting an accuracy of 83.5%. Chapter II addressed the volume estimation of trees from crown morphometry obtained by UAS in the same study area. A field forest inventory was used, where all trees with diameter at breast height - DBH >= 50 cm were measured. The locations of the trees were combined with the aerial images, and then 388 individuals distributed in 55 species were filtered. The canopy morphometry was identified through the average diameter and the canopy area, and a correlation test was subsequently performed to analyze the related size of these variables to later insert the volume, basal area and diameter of the trees distributed by diametric class in the models. The database was divided into 70% for training and 30% for testing. The correlation test identified a strong interaction between the canopy variables with volume and the lack of height, which was therefore discarded from the analysis. The predicted volume had metrics of 21.97% for the SEE fit and 19.13% for the RMSE test, both with R² greater than 0.90. The SEE for basal area and diameter was less than 6.5% and R² greater than 0.99; the residuals of the models were heteroscedastic, the volume had the greatest dispersion, but was concentrated up to ± 25%, thereby demonstrating the model accuracy and the potential to use treetop variables.
Collections
- Teses [359]