Prognose do volume de madeira em florestas equiâneas por meio de modelos agrometeorológicos de redes neurais artificiais
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
O manejo florestal sustentável requer estimativas precisas de estoque de crescimento, uma vez que essas informações auxiliam a tomada de decisão na análise econômica dos projetos florestais. A proposta de inclusão de novas metodologias ou modelos que estimem o estoque volumétrico da madeira é necessária em função dos custos do inventário florestal, bem como a importância desta na exatidão das estimativas volumétricas. Diversas ferramentas computacionais e de modelagem matemática tem sido utilizadas com êxito em situações de tomada de decisão no setor florestal, destacando-se as Redes Neurais Artificiais (RNA). Diante do exposto, o trabalho visa elucidar as seguintes questões: i) A redução do número de variáveis por meio de métodos de ordenação pode otimizar a construção de modelos de redes neurais para estimação da prognose do volume de madeira? ii) Existe diferença significativa entre prognoses realizadas por meio de diferentes métodos? iii) Qual o modelo mais eficiente para a realização da prognose do inventário florestal para a região do leste de Minas Gerais? Utilizou-se duas metodologias, Correlação de Pearson e Método de Garson, para ordenação das 21 variáveis do input dos modelos agrometeorológicos estudados. Para a construção dos modelos agrometeorológicos foram utilizadas 3 metodologias: i) modelos construídos de acordo com a ordenação da Correlação de Pearson; ii) modelos construídos de acordo com a ordenação do Método de Garson; e i) modelos construídos de acordo com a fusão das metodologias da ordenação da Correlação de Pearson e Método de Garson, denominada de metodologia do Modelo Híbrido. A seleção dos modelos ocorreu por meio do menor valor da raiz do erro quadrático médio (RMSE) do teste (%) e gráficos do volume futuro estimado versus volume futuro observado. Todos os processamentos foram realizados no software Neuro AgroClimate, desenvolvido para esta tese. Na ordenação pela metodologia da Correlação de Pearson destacaram-se as variáveis de inventário florestal, já na metodologia do Método de Garson destacaram-se as variáveis ecofisiológicas. Os valores do RMSE do teste (%) entre as três metodologias analisadas foram próximos, sendo: 6,24 do modelo 12 da metodologia da Correlação de Pearson; 6,42 do modelo 16 da metodologia do Método de Garson; e 6,61 do modelo 9 da metodologia do Modelo Híbrido. Houve diferença significativa entre os três modelos analisado, e a rede que apresentou o menor valor do RMSE do teste (%) foi o modelo 12 da metodologia da Correlação de Pearson. A seleção das variáveis dependentes foi eficaz ao otimizar o tempo de processamento das redes por se conhece as variáveis do input.
Sustainable forest management requires accurate estimates of growth stock, once this information helps decision-making process in the economic analysis of forestry projects. The proposal on inclusion of new methodologies or models that estimate the volumetric stock of wood is necessary in function of the costs of the forest inventory, as well as is importance in the accuracy of the volumetric estimates. Several computational tools and mathematical modeling have been used successfully in decision-making situations in the forestry sector, with emphasis on Artificial Neural Networks (RNA). In view of the above, the paper aims at elucidating the following questions: i) Can the reduction of the number of variables by ordering methods optimize the construction of neural network models to estimate the prognosis of the wood volume? ii) Is there a significant difference between prognoses performed using different methods? iii) What is the most efficient model for the prognosis of the forest inventory for the eastern region of Minas Gerais? Two methodologies, Pearson Correlation and Garson Method, were used to order the 21 input variables of the studied agrometeorological models. For the construction of the agrometeorological models, three methodologies were used: i) models constructed according to the order of the Pearson Correlation; ii) models constructed according to the ordering of the Garson Method; and iii) models constructed according to the fusion of the methodologies of the ordering of the Pearson Correlation and Garson Method, denominated methodology of the Hybrid Model. The selection of the models was done by means of the lowest root mean square error (RMSE) of the test (%) and graph the estimated future volume versus future volume observed. All the processes were performed in the software Neuro AgroClimate, developed for this thesis. By ordering using person correlation, variables of forest inventory were highlighted whereas by using Garson Method, it was ecophysiological variables who come to the fore. The RMSE values of the test (%) among the three methodologies analyzed were close, being: 6.24 of the model 12 of the Pearson Correlation methodology; 6.42 of the model 16 the Garson Method methodology; and 6.61of the model 9 the Hybrid Model methodology. There was a significant difference between the three models analyzed, and the network that presented the lowest RMSE of the test value (%) was model 12 of the Pearson Correlation methodology. The selection of the dependent variables was effective in optimizing the processing time of the networks when knowing the input variables.
Sustainable forest management requires accurate estimates of growth stock, once this information helps decision-making process in the economic analysis of forestry projects. The proposal on inclusion of new methodologies or models that estimate the volumetric stock of wood is necessary in function of the costs of the forest inventory, as well as is importance in the accuracy of the volumetric estimates. Several computational tools and mathematical modeling have been used successfully in decision-making situations in the forestry sector, with emphasis on Artificial Neural Networks (RNA). In view of the above, the paper aims at elucidating the following questions: i) Can the reduction of the number of variables by ordering methods optimize the construction of neural network models to estimate the prognosis of the wood volume? ii) Is there a significant difference between prognoses performed using different methods? iii) What is the most efficient model for the prognosis of the forest inventory for the eastern region of Minas Gerais? Two methodologies, Pearson Correlation and Garson Method, were used to order the 21 input variables of the studied agrometeorological models. For the construction of the agrometeorological models, three methodologies were used: i) models constructed according to the order of the Pearson Correlation; ii) models constructed according to the ordering of the Garson Method; and iii) models constructed according to the fusion of the methodologies of the ordering of the Pearson Correlation and Garson Method, denominated methodology of the Hybrid Model. The selection of the models was done by means of the lowest root mean square error (RMSE) of the test (%) and graph the estimated future volume versus future volume observed. All the processes were performed in the software Neuro AgroClimate, developed for this thesis. By ordering using person correlation, variables of forest inventory were highlighted whereas by using Garson Method, it was ecophysiological variables who come to the fore. The RMSE values of the test (%) among the three methodologies analyzed were close, being: 6.24 of the model 12 of the Pearson Correlation methodology; 6.42 of the model 16 the Garson Method methodology; and 6.61of the model 9 the Hybrid Model methodology. There was a significant difference between the three models analyzed, and the network that presented the lowest RMSE of the test value (%) was model 12 of the Pearson Correlation methodology. The selection of the dependent variables was effective in optimizing the processing time of the networks when knowing the input variables.
Description
Citation
MAGALHÃES, Mariana Rodrigues. Prognose do volume de madeira em florestas equiâneas por meio de modelos agrometeorológicos de redes neurais artificiais. 2017. ix, 62 f. Tese (Doutorado em Engenharia Agrícola) - Universidade Federal de Viçosa, Viçosa. 2017.
