Supervised machine learning for predicting and interpreting dynamic drivers of plantation forest productivity in northern Tasmania, Australia

The capacity of a plantation forest to grow and produce timber is locally constrained by topography, climate, soil conditions, and external factors such as fire and timber harvesting. Accurate estimation of forest productivity supports effective forest management. However, efficiently generating accurate models of forest productivity is hampered by the need to gather, process and integrate large volumes of disparate, high dimensional data that require computationally intensive data analysis processing methods. Recent developments in cloud-based ma-chine learning systems offer a means to address this problem.This research investigates the use of supervised machine learning to model and predict forest productivity across pine (Pinus radiata) plantations in northern Tasmania, Australia. Forest productivity models are generated by integrating 23 predictive features, including multi-temporal LiDAR (Light Detection and Ranging) derived topographic attributes, climate (rainfall and temperature) information, and edaphic conditions (geology and soil). Five machine learning (ML) regression algorithms are compared for this task: Linear Regression (LR), Polynomial Regression (PR), Decision Trees Regression (DT), Random Forests (RF) and Gradient Boosted Deci-sion Trees Regression (GBDT). The best performing algorithm, in terms of the optimal bias-variance trade-off, was RF (RMSE 2.08 and Bias-0.72) followed closely by GBDT (RMSE 2.13 and Bias-0.68) and DT (RMSE 2.94 and Bias-0.68). Tuning Model Complexity was used to provide a clear understanding of the relationship and interactions between input features and forest productivity, resulting in more accurate and interpretable models. In contrast, we conclude that GBDT results in a more reliable predictive performance than RF, and transferability model to unseen data by assessing the spatial autocorrelation. Across the top performing models, rainfall was the most important factor driving forest productivity, followed by geological class, topographic position index (TPI), landscape aspect and Digital Elevation Model (DEM). This work demonstrates the usefulness of supervised machine learning techniques to generate accurate and efficient forest productivity predictions from large and diverse spatial datasets. Moreover, users are afforded the ability to gain insight into changes that affect forest productivity through time, such as the increasing risks of wildfire and climate change by identifying the factors that contribute the most to the tree growth. By delivering a framework to understand complex and dynamic drivers of productivity through a machine learning pipeline enhanced by cloud-based machine learning systems, forest managers are provided with easily accessible tools to maximisation productivity.

Automated summary

This research investigates the use of supervised machine learning to model and predict forest productivity across pine plantations in northern Tasmania, Australia. The study found that rainfall is the most important factor driving forest productivity.