Classificação da capacidade produtiva com alturas de árvores dominantes estimadas por RNA

• Autor(a) principal: Marcos Vinicius Santana Leite
• Data de Publicação: 2022
• Outros Autores: Adriana Leandra de Assis, Christian Dias Cabacinha
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: https://doi.org/10.5902/1980509867120; http://hdl.handle.net/1843/59448; https://orcid.org/0000-0002-8148-083X; https://orcid.org/0000-0001-7777-9276; https://orcid.org/0000-0002-0723-6935
• topic: Inteligência artificial; Sitios florestais - Qualidade

The generation of site curves built from modeling the height of dominant trees measured in permanent plots at different ages considering a reference age constitutes the most practical and widespread method to classify the local productive capacity. Within a Forest Planning process, the assertiveness of the generated curves has quantitative and qualitative consequences in the allocation of resources, so that the continuous improvement of classification methods is of paramount importance. This study aimed to propose models of Artificial Neural Networks (ANN) to estimate the height of dominant eucalyptus trees, and apply them in the generation of site curves using the guide curve method, as an alternative to non-linear and assess the accuracy of estimates and stability of the classification of local productive capacity generated by these approaches. The data used are from measurements of 8,819 permanent plots installed in clonal stands of Eucalyptus urophylla × Eucalyptus grandis. Five classical non-linear models were fitted and the ANN were trained with two algorithms: Feed Forward Back Propagation Network (FFBP) and Cascade Forward Back Propagation Network (CFBP). In general, when only the age of the plots was used to estimate the dominant height, there was no difference in the results between the ANN trained with the two algorithms and the non-linear models. However, with the addition of new stand variables during ANN training, there was an improvement in estimates of dominant heights and generated a 13% more stable productive capacity classification compared to non-linear regression models.