What is the actual composition of specific land cover? An evaluation of the accuracy at a national scale – Remote sensing in comparison to topographic land cover

Abstract

Satellite imagery allows us to capture and collect land cover information for increasingly large areas. This allows us to represent current land cover on maps in a simple and standardized way; however, any land cover determined in this way is subject to some algorithmic uncertainty. This paper aims, for the first time, to indicate the magnitude of this uncertainty through the empirical probability distribution of a given land cover at a given location. By analyzing 3 data sources, i.e. the Corine Land Cover map, the POLSA land cover map and the classic map - the BDOT10k database of topographic objects. Empirical distributions of the occurrence of land cover class data in areas with a given land use on a topographic map were determined. The work was carried out on a large scale, i.e. on the maximum possible sample for Poland, i.e. on the area of the whole country. This makes it possible to introduce and quantify uncertainties. Spatial analyses were carried out using satellite-based methods to determine land cover or using a topographic map. This work and its results will be useful to all users who want to assess the occurrence of a phenomenon in a given area, taking into account the uncertainty of the land cover, and thus obtain more accurate and reliable results. It also provides, for the first time, a methodology for verifying such map correspondences, which can be replicated in work by other researchers, using the confusion matrix and as evaluation metrics the true positive rate (TPR) and weighted accuracy have been adopted. The paper proposes a link between land cover classes in all databases. It was shown that the TPR for BDOT10k was higher than 50% only with CLC Level 1 (72.0%) and POLSA Land Cover (61%), while the TPR for RS classes for each remote sensing data was always higher than 60% with BDOT10k. The class with the highest remote sensing classes was related to water, especially marine (92.0% for POLSA and 85.3% for CLC level 3), arable land (98% for POLSA, lowest for CLC level 3 (80%), and forests (coniferous POLSA – 89%, CLC level 1 and 2–85%), while low values were obtained for wetlands, peatbogs. The authors do not state which land cover approach is better, as each may have multiple uses, but the values presented in this work must raise awareness of uncertainties in land cover and critical implementation in decision-making processes for multiple areas of human activity. The study provides ready-to-use values of the probability of a given land cover class being present on a topographic map, given that remote sensing has classified it as such. These functions can also be used in reverse, to determine the probability of a given land cover class being present in remote sensing, given that a specific class has been identified on a topographic map. The results of the consistency assessment, with the composition structure, can be used by a wide range of users, including public administration, land managers, land architects, public services, academia and individuals.