APRENDIZAGEM DE MÁQUINA PARA PREDIÇÃO DE VARIÁVEIS DENDROMÉTRICAS EM ESPÉCIES NATIVAS POR VARIÁVEIS HIPERESPECTRAIS

• Autor(a) principal: Enio Antonio Manfroi Filho
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFMS
• Texto Completo: https://repositorio.ufms.br/handle/123456789/8012
• topic: Aprendizagem de máquina; predição; variáveis dendrométricas; variáveis hiperespectrais

The use of remote sensing combined with machine learning (ML) techniques is a promising approach for estimating tree growth and productivity. Many studies show improvements in estimation accuracy when ML models are implemented compared to traditional methods. The objective of this study was to investigate the performance of AM techniques and spectral readings to predict diameter at breast height (DBH) and height (Ht) of native forest species using hyperspectral variables. The study area comprised a mixed forest stand, in which 195 trees were randomly sampled. To acquire hyperspectral data, leaf reflectance readings were carried out with the ASD FieldSpec® 4 equipment. The wavelengths obtained by spectral readings, whose range varied from 350 to 2500 nm, were used as input variables for the models. The ML techniques tested were, Artificial Neural Networks (ANN), REPTree (DT), Decision Tree (M5P), Zero R (R), Random Forest (RF) and Support Vector Machine (SVM). Two input configurations were tested: 1) using only wavelengths in the prediction of dendrometric variables, without including the qualitative variable species (SE) and, 2) using wavelengths together with the qualitative variable species (CE). The performance of each algorithm was checked using correlation coefficient (r), the mean absolute error MAE and root mean square error RMSE. There was a significant interaction between inputs and AM algorithms for the variables r, MAE and RMSE. The best ranked algorithm was DT, both for DAP and Ht in all tests. There was a significant difference when using the input variable CE, demonstrating improvement in prediction results. Therefore, the study demonstrates that it is possible to predict DAP and Ht with relative accuracy using spectral bands as input in the AM models tested. When the qualitative variable species was included, the DT, M5P and SVM algorithms performed better. When this information was not included as input, it was found that the RF algorithm obtained the best results due to its prediction capacity and high stability. Keywords: Computational intelligence. Spectral bands. Forest Measurement. Hyperspectral sensing.