The overarching aim of this project is to develop advanced approaches to automated vegetation mapping by combining remote sensing and deep learning technologies. This research will investigate the accuracy, applicability, accessibility, and scalability of these technologies using Tasmania’s vulnerable vegetation communities as a case study.
To achieve this aim, we will focus on the following objectives:
Investigate the role of hyperspectral information for dense prediction tasks of native vegetation target classes in complex forest canopies by comparing the performance of deep neural network architectures that capture the spatial-spectral relationships in the data with architectures that only consider the spatial relationships.
Improve the scalability and accessibility of DNNs for vegetation mapping by exploiting domain adaptation techniques such as knowledge distillation to learn from high fidelity (spectral-spatial) datasets and predict on lower fidelity datasets.
Evaluate if, what, and how, environmental contextual information and ecological domain expertise improves the performance of DNNs to discriminate flora species in the absence of large training sets.
Characterise the impact of incomplete species occurrence records on vulnerability modelling at the subpopulation level and determine to what extent these limitations can be addressed by automated tree inventories.