Modelling Pinus pinaster Ait. resin yield using a "divide and conquer" approach and machine learning

In the context of climate change, society is facing new challenges that demand greater attention to the management of natural resources. To estimate these stocks more accurately, improved methods are necessary. Currently, machine learning techniques are considered state-of-the-art for developing models for complex processes. However, this field has only been vaguely explored in the context of the production of timber and non-timber forest products, which are fundamental to the aforementioned challenges. This paper presents a two-step process for modelling the resin production of individual trees both before and during the tapping season. The first phase involves unsupervised learning, where the internal structure of the data was extracted through a clustering process. In the second phase, the production of each cluster was modelled using machine learning techniques such as bagging, boosting, neural networks or ensemble models. The k-prototype algorithm was used for the clustering process, as it is able to work with mixed data (combinations of numerical and categorical variables) natively. The technique that obtained the optimal result in the largest number of models was the multi-layer perceptron neural network. In the modelling before the resin campaign, the most accurate model among all clusters obtained an RRMSE of 13.5 %. The combined RRMSE of all models was 26.5 %. When compared to the combined model, the non-clustered model had a 50 % higher RRMSE, and the reference model had a 67 % higher RRMSE. Similarly, when intermediate productions were incorporated into the models, the combined model outperformed the model without clustering and the baseline model. The RRMSE of the models decreased as the end of the season approached, with the accumulated production up to that date being the most influential variable. Therefore, it can be concluded that the two-stage process significantly improves the accuracy of the predictions. • The divide and conquer methodology improve the modelling process by 50–60 % in RRMSE. • Machine learning can effectively address limited datasets (409 observations). • K-prototype algorithm captures data's inherent information structure effectively. • MLP-NNs obtained the best results as a modelling technique for pine resin yield.