Local uncertainty maps for land-use/land-cover classification without remote sensing and modeling work using a class-conditional conformal approach

• LULC maps are critical for multiple fields. • LULC uncertainty maps can fail to represent well uncertainty at local scales. • A class-conditional conformal approach is used to create local uncertainty maps. • These local maps can be created without remote sensing and modeling. • These local uncertainty maps can outperform global uncertainty maps. Land use/land cover (LULC) is one of the most impactful global change phenomenon. As a result, considerable effort has been devoted to creating large-scale LULC products from remote sensing data, enabling the scientific community to use these products for a wide range of downstream applications. Unfortunately, uncertainty associated with these products is seldom quantified because most approaches are too computationally intensive. Furthermore, uncertainty maps developed for large regions might fail to perform adequately at the spatial scale in which they will be used and might need to be customized to suit the specific applications of end-users. In this study, we describe the class-conditional conformal statistics method, an approach that quantifies uncertainty more uniformly for each class but that requires more calibration data than the conventional conformal method. Using the class-conditional method, we show that it is possible to create customized local uncertainty maps using local calibration data without requiring remote sensing and modeling work and that these local uncertainty maps outperform uncertainty maps calibrated based on global data. We use empirical data from Brazil (i.e., Dynamic World LULC product and Mapbiomas validation data) to demonstrate this methodology. The analysis of these data reveals substantial heterogeneity in observations of the same LULC class between Brazilian states, an indication that national-level data are not representative of the focal state, thus explaining why uncertainty maps calibrated using focal state-level data outperform maps calibrated using national-level data. Importantly, we develop straight-forward approaches to determine the spatial extent over which calibration data are still representative of the area of interest, ensuring that these data can be used to reliably quantify uncertainty. We illustrate the class-conformal methodology by creating uncertainty maps for a selected number of sites in Brazil. Finally, we show how these uncertainty maps can yield valuable insights for LULC map producers. Our methodology paves the way for users to generate customized local uncertainty maps that are likely to be better than uncertainty maps calibrated based on global data while at the same time being more relevant for the specific applications of these users. A tutorial is provided to show how this methodology can be implemented without requiring remote sensing and modeling expertise to generate uncertainty maps.