Brent Murray
Accurate and spatially explicit information on tree species and species composition is fundamental for forest inventory and sustainable management, yet conventional approaches based on manual aerial photo interpretation are time-consuming, subjective, and limited in their ability to capture fine-scale variability within heterogenous forest stands. Although remotely sensed data such as airborne laser scanning (ALS) and multispectral imagery provide valuable structural and spectral information, existing methods often struggle to reliably distinguish species and quantify composition in complex forest environments, particularly when similarities exist among species. This dissertation investigates how deep learning approaches applied to remotely sensed data can improve the estimation of individual tree species and species composition across multiple spatial scales. To address this objective, a series of data-driven and deep learning approaches were developed leveraging forest inventory information, and multisource remote sensing data. Automated data generation strategies were developed to produce large, consistent, and representative datasets required for training deep learning models, reducing reliance on manual annotation and enabling scalable application. Various deep learning approaches, including point-, graph-, tabular, and image-based models, were then applied to learn complementary structural and radiometric feature representations, capturing complex spatial patterns and interactions within the remotely sensed data that are difficult to model using conventional methods. The integration of multi-source data further enhanced the ability to characterise forest structure and composition across varying conditions. Results demonstrate that deep learning methods can effectively learn complex relationships within remotely sensed data, leading to more reliable, consistent, and spatially explicit estimates of tree species and species composition across diverse forest conditions. These approaches enable improved representation of within-stand variability and support the generation of detailed tree species information at multiple scales. This work establishes a flexible and scalable framework for integrating multi-source remote sensing data for forest inventories of tree species, contributing to more adaptive, data-driven approaches for monitoring and managing forest ecosystems.