Emprego de redes neurais artificiais com skip-layer connections na mensura??o florestal

Submitted by M?rden L?les (marden.inacio@ufvjm.edu.br) on 2016-07-12T23:59:50Z No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) paula_ventura_silva.pdf: 2872431 bytes, checksum: 13667045a80e3b80429055fd7e4a5e15 (MD5) === Approved for entry into archive b...

Full description

Bibliographic Details
Main Author: Silva, Paula Ventura da
Other Authors: Binoti, Mayra Luiza M. da Silva
Language:Portuguese
Published: UFVJM 2016
Subjects:
Online Access:http://acervo.ufvjm.edu.br/jspui/handle/1/999
Description
Summary:Submitted by M?rden L?les (marden.inacio@ufvjm.edu.br) on 2016-07-12T23:59:50Z No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) paula_ventura_silva.pdf: 2872431 bytes, checksum: 13667045a80e3b80429055fd7e4a5e15 (MD5) === Approved for entry into archive by Rodrigo Martins Cruz (rodrigo.cruz@ufvjm.edu.br) on 2016-07-18T14:50:22Z (GMT) No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) paula_ventura_silva.pdf: 2872431 bytes, checksum: 13667045a80e3b80429055fd7e4a5e15 (MD5) === Made available in DSpace on 2016-07-18T14:50:22Z (GMT). No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) paula_ventura_silva.pdf: 2872431 bytes, checksum: 13667045a80e3b80429055fd7e4a5e15 (MD5) Previous issue date: 2015 === Coordena??o de Aperfei?oamento de Pessoal de N?vel Superior (CAPES) === RESUMO SILVA, Paula Ventura da, M.Sc., Emprego de redes neurais artificiais com Skip-Layer Connections na mensura??o florestal. 2015. 46 f. Disserta??o (Mestrado em Ci?ncia Florestal) ? Universidade Federal dos Vales do Jequitinhonha e Mucuri, Diamantina, 2015. O objetivo principal deste estudo foi avaliar a aplica??o de Redes Neurais Artificiais (RNA) utilizando a t?cnica Skip-layer connections, com e sem recorr?ncia, para estima??o do volume individual e da altura total de ?rvores de eucalipto. Os objetivos espec?ficos foram testar e avaliar as redu??es no tamanho da base de dados do conjunto de ajuste (treinamento) para estima??o dessas vari?veis. Os dados utilizados foram provenientes de ?rvores abatidas para cubagem (estima??o do volume individual) e de medi??es de parcelas permanentes de invent?rios florestais cont?nuos (estima??o da altura total), em ?rea de povoamentos de eucalipto localizados no sul da Bahia, Brasil. Foram treinadas redes do tipo Multilayer Perceptron (MLP), utilizando a fun??o de ativa??o log?stica nas camadas intermedi?ria e de sa?da e oito neur?nios na camada oculta. O n?mero de neur?nios na camada de entrada variou conforme o n?mero e o tipo de vari?vel (qualitativa ou quantitativa) em cada estudo. Os crit?rios de parada foram o erro m?dio quadr?tico de 0,0001 ou 3.000 ciclos (?pocas). Em seguida, as RNA selecionadas foram aplicadas em parte dos dados separados, para generaliza??o (valida??o). O software utilizado para o treinamento e a generaliza??o das RNA foi o NeuroForest 3.3. Para compara??o dos resultados obtidos pelas RNA, foram ajustados os modelos tradicionais de regress?o tanto para volume, quanto para altura, e tamb?m foram treinadas e aplicadas RNA usando o algoritmo Resilient Propagation, comumente utilizado em aplica??es da mensura??o florestal. A avalia??o dos resultados gerados pelas RNA e pelos modelos de regress?o foi feita por meio do coeficiente de correla??o entre os valores observados e estimados, de gr?ficos de dispers?o e de histogramas de frequ?ncia percentual dos erros percentuais. As Redes Neurais Artificiais utilizando Skip-layer connections apresentaram resultados satisfat?rios para estima??o de volume e de altura de ?rvores de eucalipto, o que evidencia a possibilidade de aplicar a t?cnica em mensura??o e manejo florestal e uma expressiva redu??o das bases de dados para treinamento das RNA. === Disserta??o (Mestrado) ? Programa de P?s-Gradua??o em Ci?ncia Florestal, Universidade Federal dos Vales do Jequitinhonha e Mucuri, 2015. === ABSTRACT SILVA, Paula Ventura da, M.Sc. Artificial neural networks with skip-layer connections in forest measurement. 2015. 46 f. Dissertation (Master in Forest Science) ? Universidade Federal dos Vales do Jequitinhonha e Mucuri, Diamantina, 2015. The aim of this study was to evaluate the application of Artificial Neural Networks (ANN) using the Skip-layer connections technique, with and without recurrence, to estimate the individual volume and total eucalyptus trees height. Its specific objectives were to test and evaluate reduction in the size of the adjustment assembly database (training) for estimating these variables. The data came from trees felled for scaling (estimation of individual volume) and measurements of permanent plots of continuous forest inventories (estimation of the total height), in eucalypt plantation area located in the south of Bahia, Brazil. Multilayer Perceptron (MLP) network type, using the logistic activation function in the intermediate and output layers and eight neurons in the hidden layer, were trained. The neurons number in the input layer varied according to the number and type of the variable (qualitative or quantitative) in each study. The stopping criteria were the root mean square error 0.0001 or 3,000 cycles (seasons). The software used for the RNA training and generalization was the NeuroForest 3.3. To compare the results obtained by RNA, traditional regression models were set for both the volume and the height, as well as RNA were trained and applied using the Resilient Propagation algorithm, commonly used in forest measurement applications. The evaluation of the results generated by the RNA and by the regression models was made via the correlation coefficient between observed and estimated values, scatter plots and histograms percentage frequency of the percentage errors. Artificial Neural Networks using Skip-layer connections showed satisfactory results for the estimation of volume and eucalyptus trees height, demonstrating the possibility of applying the technique in measuring and forest management and a significant reduction of databases for RNA training.