Uso de redes neurais artificiais como alternativa para estimativa volumétrica na floresta estadual do Amapá.

Abstract

Good forest management practices require estimates of available forest volume to be efficiently and accurately obtained to assist in making forest use decisions. The conventional methodology applied consists of the use of the form factor to correct the cylindrical volume and the adjustment of regression equations by least squares method. Recently a new altemative has been used that uses computational techniques of artificial intelligence or the so-called artificial neural networks (ANNs) that also search for precise estimates of the volumes of the inventoried standing trees. Therefore, this study aims to apply this new technology to estimate the volume of standing trees in the Amapá State Forest. Volumes of 1028 fallen trees in good condition with DAP >10cm were calculated by the Smalian method combined with the Hohenald method in 30 sample units distributed in three phytophysiognomies: Low Plateau Terra Firme Forest (FTFBP), Submontana Terra Firme Forest (FTFSM) and Forest-Savanna Transition (TCF). Fifty artificial neural networks of the Multilayer Perceptron type (MPL) were trained for each phytophysiognomy using DAP and HC (commercial height) in the inlet layer and Volume (m3) in the outlet layer. The type of training applied was Resilient Propagation with Sigmoidal function and stop criteria based on average error and number of cycles. The criteria used to evaluate the best RNA were: graphical analysis of residues, correlation coefficient (ryy), bias and square root mean square error (RQME). Five mathematical volume models were also tested via regression analysis, with the independent variables being DAP (diameter at breast height) and HC (commercial height). According to the results obtained, ANNs and mathematical models presented similar results. ANNs with 2-6-1 architecture with DAP and HC input variables, training algorithm (MLP), with 30,000 cycles and sigmoidal activation function, produced the best results in the trained networks. The proposed regression model of Schumacher and Hall was the one that best fit the data, and can be used to estimate the volume of standing trees in the Amapá State Forest in its three phytophysiognomies FTFDBP, FTFDSM and TCF.