Utilisation of remote sensing, machine learning, and agent-based simulation for biophysical assessment of young boreal forest

Abstract

Finland is transitioning to carbon neutrality, as demonstrated by the fact that in 2024, 95% of electricity production and 78% of heat production were carbon neutral. The biomass byproducts from first and pre-commercial thinning in young forests are recognised as a crucial bioenergy resource in Finland. Accurate forest monitoring in young forests is critical for sustainable biomass utilisation, carbon stock assessment, and effective silvicultural management. Remote sensing and simulation techniques offer efficient alternatives to labour-intensive field measurements for assessing key forest parameters such as tree height, canopy cover, species classification, and harvesting yield. This dissertation seeks to evaluate the performance of light detection and ranging (LiDAR) and digital aerial photogrammetry (DAP) in estimating individual tree height in young forests. It also examines boom corridor thinning as a cost-efficient management strategy, employing simulations to enhance corridor dimensions for harvester efficiency. Moreover, the estimation of forest canopy cover is assessed utilising vegetation indices derived from PlanetScope multispectral imagery, examined across six machine learning regression models. The final objectives include tree species identification using the latest YOLOv12 algorithm and semantic segmentation using random forest, CatBoost, and convolutional neural networks (CNN) models. A 30-hectare young forest in Pieksämäki, Finland, was surveyed using field measurement, UAV LiDAR, and DAP. Multispectral data from PlanetScope were collected for canopy cover modelling (regression) and tree species classification. The results show LiDAR outperformed DAP in estimating pine and birch height (R² = 0.81–0.86, RMSE = 1.44–1.56), while both performed similarly for spruce (R² = 0.83). Simulation results indicate that the optimal ratio of corridor width to unclear area length is between 0.5 and 1.0 to achieve a removal of 40% to 60%. PlanetScope-based canopy cover estimation using LightGBM yielded the highest accuracy (R² = 0.64–0.69, RMSE = 0.16). For species classification, YOLOv12 reached 79% detection accuracy, and CatBoost segmentation achieved 85% accuracy and 81% MCC, surpassing CNN when combining RGB, CHM, and vegetation indices. The results show the potential of remote sensing, simulation, and machine learning methods for forest inventory, thinning strategies, and tree species classification in boreal forests.