A continental-scale approach to ecosystem condition monitoring using satellite imagery and deep learning

Substantial declines in biodiversity over the past century demonstrate an immediate need to preserve ecosystems and further mitigate habitat loss. Monitoring changes in ecosystem condition at region thru continental to global scales can provide an important information of biodiversity declines and help facilitate targeted intervention. Efforts to use satellite imagery to map ecosystem condition change have experienced challenges with distinguishing observed changes from the natural variation of ecosystems. In this study we use an innovative deep learning architecture to pair time series satellite imagery with locations of known on-ground condition. Our model was developed using 209,041 on-ground records of native species present in the landscape, as a surrogate measure of ecosystem condition, coupled with Landsat time series data and topographic and climatological datasets. We predict ecosystem condition across the Australian continent for several years (2010, 2015, 2020, 2021, 2022) at 100 m. Arid regions in Australia's interior had predicted condition scores close to reference condition (1) for all years. Comparatively, highly modified landscapes in Australia's southeastern and southwestern regions had predicted condition scores closer to fully degraded (0). Mean predicted ecosystem condition across Australia was greater than 0.65 for all years, suggesting greater overall presence of native species rather than absence, however this was spatially variable. Our results demonstrate that using deep learning techniques and time series data can provide quantitative information on ecosystem condition, accounting for temporal variability of vegetation phenology and spatial variability across bioregions. Ongoing efforts to collect essential biodiversity variables from space must consider integrating with deep leaning approaches that have capacity for context driven spatial modelling. This will help ensure mapping products can support policy and inform intervention strategies. • Innovative deep learning techniques paired with satellite imagery can predict ecosystem condition. • We utilise 209,041 on-ground records of ecosystem condition for model development. • Proportion of native species was used as a surrogate measure of ecosystem condition. • Predicted ecosystem condition across Australia was greater than 0.65 for all years. • Our workflow provides a context-driven approach to predicting ecosystem condition.