Jaime Candelas Bielza defended his PhD thesis “Advancing tree species composition prediction in boreal forests with remote sensing” April 24th, 2025.
The topic of the trial lecture was “The need and value of tree species information for forest biodiversity, resilience, and management”.
Thesis abstract
Accurate tree species composition information is crucial for forest management,
influencing harvest scheduling, regeneration choices, and silvicultural treatments. In
operational forest management inventories (FMIs), species composition is typically
estimated through manual photo interpretation, a costly, subjective method prone to
systematic errors. While efforts to improve species data have focused on species
classification (identifying tree species), species composition (quantifying relative
amount) provides a more detailed forest description, particularly relevant in Nordic
countries where forest attributes are reported by species. Most tree species classification studies have been limited to small experimental areas and individual trees, offering limited insights for operational applications.
To address these gaps, the first two studies in this thesis employed an area-based
approach, which linked remotely sensed data with field inventory plots to predict forest
attributes over larger areas. Conducted across eight study areas, these studies provided findings that are more generalizable to a broader range of forest conditions andoperational management contexts.
The first study evaluated combinations of remote sensing (RS) data (ALS, aerial imagery and Sentinel-2 imagery) to determine the most accurate predictors of species
composition. Results demonstrated that combining ALS data with spectral Sentinel-2
imagery, particularly when multi-season Sentinel-2 imagery was used, capturing
phenological variations. The findings highlight the potential of such data combinations
to reduce reliance on manual photo interpretation.
The second study compared parametric (Dirichlet regression and multinomial logistic
regression) and non-parametric (random forest (RF), k-nearest neighbors (k-NN),
extreme gradient boosting and multilayer perceptron) modeling techniques for
predicting species composition. While Dirichlet regression, RF and k-NN showed
similar predictive accuracy without significant differences in performance, Dirichlet
regression offers advantages such as simplicity and lower data requirements, making it
more suitable for FMIs with limited sample plots. On the other hand, non-parametric
RF and k-NN are better suited for modeling more complex relationships in the data,
emphasizing the importance of choosing modeling techniques based on data
availability and application needs.
The third study examined the separate and combined effects of species composition
and site index (SI) prediction uncertainties on net present value (NPV) calculations, a
common measure of forest economic value. The findings revealed that site index
uncertainty had a greater impact on NPV than species composition uncertainty and
that their interaction amplified errors in economic valuation. The study also
demonstrated that calibration of both species composition and SI predictions
systematically reduced NPV errors, highlighting the value of calibrating remote sensing
predictions to minimize uncertainty in forest inventories.
Overall, this thesis demonstrates the potential of integrating remotely sensed data and
statistical modeling to improve forest inventory practices. By combining ALS and
multi-season Sentinel-2 data and selecting appropriate modeling techniques, tree
species composition can be predicted with improved accuracy and reliability than
current operational methods. These advancements provide scalable and cost-efficient
alternatives to conventional methods, reducing reliance on manual aerial photo
interpretation to support forest management decisions. Furthermore, the demonstrated
impact of prediction uncertainty on economic valuation emphasizes the need for
accurate species composition and SI predictions in forestry planning. Ultimately, this
research contributed to the development of objective, data-driven methods that can
improve the precision, efficiency and adaptability of operational FMIs.
Supervisors:
Hans Ole Ørka, Norwegian University of Life Sciences (NMBU)
Lennart Noordermeer, Norwegian University of Life Sciences (NMBU)
Erik Næsset, Norwegian University of Life Sciences (NMBU)
Evaluation committee
Joanne C. White, Canadian Forest Service
Lauri Korhonen, University of Eastern Finland (UEF)
Hans Fredrik Hoen, Norwegian University of Life Sciences (NMBU)