Retrieving boreal forest structure from remote sensing data using reflectance modelling and machine learning

Boreal forests, which are increasingly affected by climate change, hold significant ecological value and are central to the global carbon cycle. Therefore, the effective monitoring of these forests is of great importance. This thesis aims to advance methodologies for monitoring the northern European boreal forests, with a primary focus on Finnish forests. Specifically, the thesis develops a new hybrid method that combines a forest reflectance model with a machine learning algorithm to retrieve forest variables from passive optical remote sensing data.

To develop this method, the research began by assessing the suitability of machine learning algorithms for retrieving forest variables in Finnish forests and determining the spectral resolution of the remote sensing data required for accurate forest variable retrievals. Subsequently, the accuracy of a forest reflectance model was improved for the determined spectral resolution. Finally, a new hybrid method was developed by integrating the forest reflectance model with a machine learning algorithm. In addition, very-high-resolution images were used to enhance the robustness and accuracy of the hybrid method.

This thesis makes several contributions to the advancement of boreal forest monitoring methods. First, it demonstrates that the added value of hyperspectral imaging is primarily linked to forest variables that include species-specific information, whereas traditional spaceborne multispectral remote sensing data is sufficient for accurately retrieving common forest structural variables. Second, the spatial pattern of trees and the ratio of branch area to leaf area were found to significantly influence forest reflectance modelling accuracy. Third, the results show that hybrid methods hold great promise for retrieving forest structure. Despite the promising results, challenges remain. The main challenge of the developed hybrid method was the disparity between the datasets used for training and testing, which highlighted the importance of spectral and structural representativeness in machine learning-based applications.

While the thesis presents relevant findings for advancing methodologies in boreal forest monitoring, it also highlights several challenges and limitations that need to be addressed. Overcoming these challenges requires further investigation, emphasising the need for additional research. Overall, this thesis provides a solid foundation for developing hybrid methods that use forest reflectance models, contributing to the evolving field of hybrid approaches in remote sensing research.