A new approach improving koala habitat prediction using hyperspectral airborne imagery

Koala populations are declining primarily due to habitat loss, making large-scale habitat quality prediction essential for conservation. A first approach to defining koala habitat quality involves identifying the number of different 'koala' trees species present, and their individual chemical composition. Previous studies have used remote sensing, from drones to satellites, to measure these factors. While the trade-off between scale and resolution is evolving, airborne hyperspectral data currently offers a promising avenue for developing predictive models. In the past, koala studies predicting nitrogen content using airborne data achieved relatively high accuracies, but they did not explore the potential of classifying koala tree species and incorporate this explanatory variable into nitrogen models. Here, we test these ideas while testing a new approach: using individual canopy pixels, rather than canopy averages, to train models. We argue that pixel-level training exposes models to noisy pixels (different shading, angle or background), forcing them to still find patterns and improving robustness to new data. Furthermore, this approach increases training data by leveraging within-canopy variability, reducing the need for large sampling effort. We found that species classification using individual pixels achieved 96 % accuracy, but over iterations converged to the same accuracy (74 %) as using average canopy reflectance. Furthermore, we demonstrate that pixel-level training and explicit incorporation of tree species substantially improves the nitrogen prediction model (R² = 0.61). Our findings support the use of this approach in future airborne data, allowing for improvements to models predicting koala habitat. Such models inform regional-scale habitat assessments and conservation decisions, helping prioritize high-value areas.