Comparing Performance of Linear Regression Models Trained on Systematic Forest Measurement Datasets to Predict Diameter at Breast Height

Natural resources in forests are usually estimated through systematic forest measurement programs such as national forest inventories. One of the most important parameters of a tree is diameter at breast height (DBH), which is the base of assessing growth and yield in forests both for stand-level or for regional metrics. However, as most of the forest inventory programs operate with DBH threshold - meaning that only above certain DBH value the trees are measured in a plot -, it leads to missing points in the data. If the sample tree is measured in a cycle of measurement, it might have not been measured in a previous cycle (5 year before) because it might have been below threshold then. The objective of the research is to develop a workflow for predicting these missing DBH values by using machine learning algorithms. First, we integrated the observed data into data warehouse with an ETL (extract-transform-load) process by filtering, data cleaning and transformations. Next, we selected a use case, which is the prediction of the 1st cycle’s DBH values based on 2nd cycle’s tree- and plot-level data. Afterwards we developed, tested and evaluated three different linear algorithm models for the chosen use case. The evaluation shows that supervised machine learning models trained on tree- and plot-level parameters can help missing data imputation in systematic forest observation datasets. Although the models have promising results but we are aiming to continue our research by exploring other algorithms and use cases. These studies can serve as a stable base for future analysis research as well.