Predição de volume em florestas plantadas utilizando escaneamento com laser aerotransportado e aprendizagem de máquinas

• Ano de defesa: 2021
• Autor(a) principal: Ferreira, Flavio Carvalho
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Lavras; Programa de Pós-Graduação em Engenharia Florestal; UFLA; brasil; Departamento de Ciências Florestais
• Palavras-chave em Português: Escaneamento aéreo com laser; Manejo florestal; LiDAR; Inventário florestal; Aprendizagem de máquinas; Airborne laser scanning; Forest management; Forest inventory; Machine learning; Recursos Florestais e Engenharia Florestal
• Link de acesso: http://repositorio.ufla.br/jspui/handle/1/49056

The management of planted forest have been increasing the necessity for more and fast information, in quantity and quality, to base its decision-making. Due to it, the present work aiming compare the volumetric estimative obtained by the forest inventory using the conventional method, based on the mean and standard deviation of the sampled units, and a method that uses data from airborne laser scanning (ALS) of the stands of the forest to generate maps with the volume of wood. In the ALS inventory, three different modeling methods were tested: linear model adjusted by ordinary least squares (OLS), Random Forest (RF) and Support Vector Machines (SVM). The models were trained and validated by the holdout method for 200 times, evaluating in the end the mean of the coefficient of determination (R²) and the root mean square error (RMSE). The model that obtained the best result was the linear model with R² of 53.74% and RMSE of 20,19 m³/ha. When comparing conventional and ALS inventory methods, both had low inventory errors. It can be concluded that the use of ALS inventories is promising for the forest sector.