Selection of a machine learning algorithm for inference of the basic density of wood a from measurements on the standing tree

Abstract

Even with advances in genetic engineering and cloning, the properties of wood are still affected by different factors (climate, soil, altitude, etc.) related to the development of trees. These interferences cause the wood produced in the forest to have great variability, making it difficult to know important reference values for the production line of forest companies linked to cellulose and paper, wood-derived products (fiber and particle boards) and, also, to round or processed wood to be used in solid or engineered form (glued laminated wood or counterlaminated). For this reason, methods are sought all over the world to anticipate and monitor the inference of wood properties with tests on the standing tree, with non-destructive techniques being the most commonly used in the generation of inference models. In addition to allowing direct application to the tree, these techniques allow the tests to be repeated over time as many times as necessary, as the material under evaluation is not destroyed during the test. In general, they are also easy and quick to apply, allowing the generation of a lot of data. However, no technique is complete, and the association of techniques is the most suitable for the generation of more accurate models. In general, the inference models used in studies for the forest sector use conventional statistical techniques, which aim to formalize relationships between variables in the form of mathematical equations (statistical modeling). However, it is currently considered that machine learning can be a more efficient tool when you have a very large volume of data, which is often the case with non-destructive techniques, especially when you want to associate different techniques. With statistical modeling, you can find relationships between variables to predict an outcome, while machine learning allows you to build systems that can learn from the data. The company Valora Madeira operates in the provision of services for the forestry and civil construction sector with wood, one of the services being the prediction of wood properties from measurements in standing trees. Thus, the feasibility of creating an innovative modeling process will allow the company to offer a better quality service, with more assertive results. Considering the mentioned aspects, the objective of this project is to study different machine learning algorithms and propose the most appropriate and viable to learn from the data generated in the field with different non-destructive techniques and, with this learning, to predict specific properties of wood. Bearing in mind that a large-volume database is needed to study these algorithms, we adopted the basic density as a variable for this phase of the research. This variable was adopted because it is the property most studied by pulp and paper companies, the sector in which we intend to carry out the research. This sector was chosen because it allows us to have access to a large number of trees and with the necessary variability for the study, as they come from different clones and grow in regions with different edaphoclimatic conditions. (AU)