Volumetric spatiality of wood in forests with artificial neural networks based on satellite imagery

Abstract

Sustainable agriculture of forest plantations requires the permanent estimation of the amount or volume of wood being produced at any given time, which is difficult in large forest areas if only manual procedures are used in the field. In the present research, multilayer perceptron artificial neural networks (ANNs) were modeled for the spatial estimation of wood volumes in a Eucalyptus sp. plantation located in the state of Mato Grosso do Sul, Central-West region of Brazil. For this purpose, spectral bands, band textures obtained with gray level cooccurrence matrix and vegetation index, which were derived from SPOT 6 digital satellite image, were used as prediction variables. The resulting ANN with the best performance presented an accuracy of 93.32% and a coefficient of determination of 0.9761, with respect to values obtained with field measurements; however, it presented a relative mean square error of 16.32% (RMSE of 7.85 m3/hm2), but the distribution of residuals was not biased, therefore, the model was promising for mapping timber volumes in large areas without overestimating or underestimating the prediction. The constructed network showed greater precision and accuracy when compared to other methods using similar estimation variables, including when compared to neural models using only spectral bands and vegetation indices.