Validation of Soft Maps Produced by a Land Use Cover Change Model

Abstract

AbstractIn Land Use Cover Change (LUCC) modelling, soft maps are often produced to express the propensity of an area to land use change. These maps are generally prepared in raster format, and have values of between 0 and 1, indicating the propensity of each pixel to change. In the literature, they are referred to as suitability, change potential or change probability maps. These maps are sometimes considered as the final product of a model (e.g. map of deforestation risk), but they can also serve as intermediate products that simulate the changes from which a hard-simulated land use/cover map can later be prepared using, for example, a cellular automaton. In both cases, it is essential to evaluate the soft map’s ability to identify the areas that are most susceptible to change. One way of assessing this ability is to compare the spatial coincidence between the real changes observed on the ground and the values estimated by the soft map. One would expect real change areas to coincide with high change potential values (near 1) and real no-change areas with low change potential values (near 0). This comparison can be made using various statistical approaches including Correlation Coefficient (Sect. 1), the Receiver Operating Characteristic (ROC) (Sect. 2) and the Difference in Potential (DiP) (Sect. 3). Other measures, such as total uncertainty, quantity uncertainty and allocation uncertainty (Sect. 4), are used exclusively in the analysis of soft maps. In this chapter, we describe the fundamental steps involved in these four statistical approaches to validating the soft maps produced by a model. The four sections are illustrated with specific cases: to validate soft maps produced by the model, to validate soft maps produced by the model against a reference map and to validate soft maps produced by various models against a reference map. We use the Ariège database to validate the different soft maps (change potential and suitability maps) produced by the model by comparing them with real land use maps of the Ariège Valley for two dates (CORINE 2012 and 2018). All these validation techniques are carried out using raster data. As commented earlier, the soft maps produced by the model are continuous, ranked variables. We designed exercises using this original format. In other chapters of this book, the soft maps produced by the model are validated after reclassification of the original maps.