Estimativa da biomassa acima do solo em floresta de terra firme na Amazônia com dados LiDAR aerotransportado e upscaling com imagens orbitais

• Ano de defesa: 2023
• Autor(a) principal: Schuh, Mateus Sabadi
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/26339/0013000005t25
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Recursos Florestais e Engenharia Florestal; UFSM; Programa de Pós-Graduação em Engenharia Florestal; Centro de Ciências Rurais
• Palavras-chave em Português: Carbono florestal; Multiescala; Sensoriamento remoto; Inteligência artificial; Forest carbon; Multiscale; Remote sensing; Artificial intelligence; CNPQ::CIENCIAS AGRARIAS::RECURSOS FLORESTAIS E ENGENHARIA FLORESTAL
• Link de acesso: http://repositorio.ufsm.br/handle/1/30289

The high concentration of living biomass stored in its different vegetation formations gives the Amazon forest a leading role in discussions on carbon cycling and climate monitoring. The detailed study of the dynamics of forest carbon involves improving the measurement of stored biomass, since the traditionally employed methods are costly and with limited spatial range. In this sense, the study addressed the development of an upscaling protocol that associated data from an airborne LiDAR (Light Detection and Ranging) sensor, OLI/Landsat-8 images and field information, for modeling and mapping above ground biomass (AGB) in a region of Dense Ombrophylous Forest of the Amazon biome. The research was carried out considering three approaches: (1) Model via the AGB stock present in inventoried plots using LiDAR data and spatialize the estimates throughout the study area (Fazenda Cauaxi, municipality of Paragominas/PA); (2) Perform the same procedure using OLI/Landsat-8 image data as predictors; (3) Use the AGB map via LiDAR as a calibration reference in estimations with OLI/Landsat-8 images. The modeling was implemented using the Support Vector Machine (SVM) algorithm. AGB data derived from field observations were taken as reference for model validation. Finally, the AGB maps were submitted to an uncertainty analysis process associated with the pixel estimates. The results of the three approaches reveal maps with estimates within the reference AGB confidence interval, both in value per hectare (229,6 ± 18,2 Mg.ha-1), and for the total area (289.256,5 ± 22.851,2 Mg). At the plot level, the estimates were shown to be valid by the Wilcoxon Rank Sum Test. The LiDAR model showed the highest Spearman’s correlation (rho=0,89 and p-value<0,0001), lowest RMSE (32,8 Mg.ha-1), and lowest standard error (14,5%) compared to the others. On the other hand, the OLI/Landsat-8 approach showed a weak correlation between image-derived predictors and AGB in plots, which determined the worst performance (rho=0,13 and p-value=0,2642, RMSE=87,2 Mg.ha-1, standard error=38,4%). The upscaling method brought performance gains by combining the AGB map via LiDAR with OLI/Landsat-8 images in the modeling (rho=0,31 and p-value=0,0699, RMSE=79,2 Mg.ha-1, standard error=34,9%). The uncertainty analysis revealed the difficulty for models based on spectral variables to reproduce the full amplitude of the AGB present in the study area. Even with the performance gain, the upscaling approach presented an average uncertainty of 108 Mg.ha-1. The research results reinforce the potential use of the combination of remote sensors in estimating forest attributes. The calibration of spectral models with previous AGB maps via LiDAR data can help compensate for the optical data saturation and improve predictions, especially in regions with high AGB density and structural complexity, characteristics of the Amazon rainforest.