Satellite remote sensing is increasingly used as a method of large-scale monitoring in ecology and conservation. Much attention has been given to the use of satellite remote sensing for detecting changes in land-use or cover change. However, existing studies typically only use a subset of ‘best’ cloud-free images and thus often discard vast amounts of potentially useful data. Furthermore, they frequently ignore uncertainty in the land-cover detection and classification process, which could limit their ability to accurately detect and classify ecosystem loss. Conventional statistical approaches in ecology, such as occupancy-detectability models, explicitly estimate these types of error implicitly in the modelling process. Despite their demonstrated performance gains, they have not yet been applied to remote sensing data and questions.
Here, we developed a statistical method to take detection probability into account, explicitly quantifying uncertainty associated with area estimates obtained from satellite images. We used all available passive satellite images within a region of interest from 1982 onwards, and a suite of relevant covariates that can affect ecosystem detectability at a given time point, including: sensor type, number of sensors, cloud cover, time of year, modelling algorithm.
Results/Conclusions
We show that by taking advantage of extensive multi-sensor satellite data using cloud-based platforms such as the Google Earth Engine, this new method better quantifies trends in ecosystem area, while estimating uncertainty within the process. This allows us to create time-series data for areas which are usually difficult to monitor using satellites due to cloud cover, and enables the use of images acquired by Landsat 7 after the scan line corrector failure.
By improving estimates of ecosystem area across a time-series, our approach can provide information on the drivers of ecosystem loss and the factors that influence area over time. Moreover, the methods developed will also be applicable to other metrics that can be summarised over the classified ecosystems (e.g. NDVI to estimate ecosystem productivity). This would allow monitoring that includes temporal variation, providing additional insights into e.g. phenology and the effects of seasonality on ecosystems. The outputs from our method are directly applicable for ecosystem risk assessments requiring understanding of ecosystem change through time, such as the Red List of Ecosystems. They will lead to more accurate ecosystem status categorisations, while also providing the means of investigating drivers and pathways of change.