Chapter 3: Paper 2

2.2.1Problem Statement¶

High-resolution hyperspectral imagery enables precise discrimination of vegetation classes in complex forest canopies through detailed spectral signatures Modzelewska et al., 2020. However, hyperspectral data is seldom used for regional-scale mapping and monitoring in operational contexts due to the cost of sensors and processing complexity Santos et al., 2022. The imagery used for regional-scale vegetation mapping is typically supplied by aerial imagery vendors in proximity to the area of interest. Contemporary aerial imaging platforms are typically equipped with frame cameras sensitive to red, green, blue (RGB), and occasionally near infrared (NIR) wavelengths. Most operational imagery programmes target a spatial resolution of around 0.10 m, striking a balance between survey coverage and spatial detail. Satellite imagery has similar trade-offs, where the broad temporal, spatial, and spectral coverage come at the cost of spatial resolution. Some commercial satellite programs offer multispectral imagery with sub-metre spatial resolution, but the cost of these datasets can be prohibitive in many cases.

This disparity creates a critical gap between research-grade datasets optimised for algorithm development and the operational imagery available for conservation management applications. Without effective domain adaptation strategies, advances in deep learning-based species mapping remain confined to controlled research environments, limiting their application to real-world conservation practice where they are most needed.

2.2.2Proposed methods¶

This research will develop a progressive domain adaptation framework employing spectral knowledge distillation and multi-resolution training strategies. The approach will utilise paired hyperspectral-multispectral datasets to train teacher-student architectures, where hyperspectral-trained models guide the learning of spectrally and spatially constrained networks.

The framework will be evaluated using both synthetically degraded datasets and authentic operational imagery to assess real-world performance. Synthetically degraded datasets will aim to emulate common data characteristics from aerial imagery programs such as the Tasmanian Imagery Program, and from high resolution satellite programs such as WorldView-3. This approach will enable controlled ablation studies while validating the generalisability in genuine datasets.

2.2.3Key innovations¶

This work will establish a framework for leveraging the rich spatial-spectral information in hyperspectral datasets to make accurate dense predictions of vegetation type in lower-fidelity, but universally accessible imagery. The proposed dual-encoder architecture represents a novel approach to leveraging rich training data while maintaining compatibility with standard aerial platforms. By developing systematic knowledge transfer protocols, this research will support the widespread deployment of advanced deep learning methods using readily available imagery, expanding the practical impact of remote sensing technologies in conservation management.