ESTRATÉGIAS PARA A MODELAGEM DE VOLUMES INDIVIDUAIS DE ESPÉCIES NATIVAS COMERCIAIS NA AMAZÔNIA BRASILEIRA

Abstract

Sustainable forest management is one of the main activities carried out in the Amazon. However, to efficiently manage natural forests, it is crucial to understand their structure and accurately estimate their timber stock. In this context, the aim of this study was to test and compare different strategies, modeling methods, and downgrade coefficients to estimate the individual commercial volume of native Amazonian species, identifying the most appropriate one for estimating commercial volume. The data on diameter at breast height (dbh), commercial height, and commercial volume used in the estimates were obtained through a complete (100%) forest inventory, trees cubage, and log scaling at Annual Production Unit number two (APU-2), located in the Tapajós National Forest, Belterra, Pará, Brazil. The methods used for volume estimates included form factor, traditional models without stratification (generic equation), stratified models (speciesspecific equations and diameter class-based equations), as well as mixed-effects models and artificial neural networks (ANN). These methods were applied to two data sources: whole stem cubage (scenario 1), considering 13 relative positions along the commercial height, and data derived from log scaling (scenario 2), a phase of forest management in which cubage is performed on logs that typically range between 4 and 7 meters in length. The species studied were Dipteryx odorata (Aubl.) Forsyth f., Hymenaea parvifolia Huber, Lecythis lurida (Miers) S.A. Mori, Manilkara elata (Allemão ex Miq.) Monach and Mezilaurus itauba (Meisn.) Taub. ex Mez, which account for 54% of the total harvested volume in APU-2. Diameter class intervals of 15, 20 and 25 cm were used in the adjustments. To compare and select the most accurate method, Pearson's correlation coefficient (rŷy), root mean square error (RMSE %), bias, and residual graphical analysis were considered. In addition to volumetric estimates, downgrade coefficients between inventoried and harvested commercial volumes were also generated. The results obtained across the evaluated scenarios were similar; however, the methods demonstrated better fit statistics and precision with the cubage data. In this scenario, errors ranged from 8.86% to 13.56% for the best adjustments of each method. For the log scaling data, the errors were higher, ranging from 15.13% to 34.90%. Volume estimates using form factors resulted in the highest errors compared to other methods in both scenarios and are therefore not recommended. The generic models also showed relatively low accuracy, being only slightly more accurate than the form factor. However, the implementation of stratification in traditional regression models improved the accuracy of the volume estimates. In both generic and stratified equations, the double-entry models were more accurate compared to Husch's single-entry model, with the Schumacher and Hall model standing out as the most efficient in all situations. Mixed-effects modeling, while promising, was slightly less accurate than the traditional stratified models by species. Nevertheless, the estimates obtained using Artificial Neural Networks (ANN) showed the best statistical performance in both analyzed scenarios, emerging as the most recommended method for commercial volume estimation. Finally, downgrade coefficients proved to be effective tools for increasing the accuracy of harvest volume estimates, providing more realistic predictions and contributing to the optimization and planning of forest operations.