Get reliable rangeland science

Environmental monitoring by land managers and land administrators using remote sensing has entered the era of big data, machine learning and open data. Remotely sensed data are commonly used in models as predictors for machine learning algorithms trained on field observations, to develop new products and applications as more data and greater computing power become available. Despite the wealth of remote sensing data, it is often difficult to obtain representative reference data for these applications, and complete coverage of the predictor space is rare. Users are left to assume that the model outputs will apply to regions where reference data are unavailable. This can result in situations where stated accuracy and error metrics show good model performance, however at a local and possibly regional scale s, the map products may not be representative of the true state and therefore not serve its intended purpose. Users of new remote sensing products and applications need to be aware of the uncertainty inherent in the products they use. In this paper, we present an approach to communicate uncertainty by adding a spatial component to performance metrics by applying the 'Area of Applicability' (AOA) of spatial prediction models, to the Joint Remote Sensing Research Program 's Fractional Cover 3 (FC3) product. The FC3 product is widely used in research and applied settings to monitor vegetation cover and bare ground and inform other models such as pasture biomass or land condition across Australia's rangelands. It is imperative that we continue to improve our understanding of fractional cover models, and their strengths and limitations, to provide appropriate advice and direction to users who are reliant on these data. It can also help to inform future investments in field data collection or other methods of training data collection.