Analysing the uncertainty of the CORINE Land Cover time series (1990-2018) for Spain

ion that is difficult to achieve through semi-automated approaches (European Environment Agency, 2021). The Portuguese experience, based on the automatic detection of changes for further manual photointerpretation, may be a balanced solution to update CORINE ensuring its historical coherence and consistency. The problem of mixed categories When mapping landscapes at coarse scales, mixed categories are usually required to define areas made up of a mixture of different land uses or covers that cannot be attributed to a single one (Villa et al., 2008; Valcárcel Sanz and Castaño Fernández, 2012). However, these categories come with important limitations, usually because of their This manuscript has been submitted for publication in “International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' imprecise definitions, as they are created with a general purpose, to host different associations of uses and covers (Valcárcel Sanz and Castaño Fernández, 2012). Because of the absence of clear rules to define what a mixed category is, important variations may arise between mapping exercises from different users or following different methods. In addition, because of that flexibility, mapping LUC change is also very difficult when working with mixed categories. In almost all editions of CORINE, mixed categories are involved in most of the change mapped by CHA and CLC layers. In CHA layers and the first two CLC layers (90, 00) most of this dynamism is driven by the changes among natural vegetated areas and all changes from and to transitional woodland/shrub. Moors and heathland, a category whose definition is similarly imprecise, is also one of the main drivers of the mapped LUC change in all CORINE editions. In the two last editions of CORINE, CLC layers show a lot of changes where heterogeneous agricultural areas categories (complex cultivation patterns and land principally occupied by agriculture) are involved, which can be attributed to technical changes because of the different method of production of CORINE (see previous section). Maucha et al. (2011) analyzed the inconsistencies between CHA00 and CHA06 layers for all Europe. Mixed categories and, specifically, transitional woodland/shrub and complex cultivation patterns, were found to be behind most of the checked inconsistencies. Although this may be caused by the important dynamism of the areas mapped under these categories, it may there be a correlation between the use of mixed categories and the higher uncertainty of the mapping process. In fact, part of the inconsistencies detected by Maucha et al. (2011) were attributed to the subjectivity when mapping and differentiating these categories from other ones. The results of our study This manuscript has been submitted for publication in “International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' and, specifically, the categories involved in the technical changes detected in the 2012 and 2018 editions of CORINE, point out in this direction. The better characterization of the areas assigned to mixed categories could be a solution to this problem. In this regard, a survey carried out by the EEA among the CORINE participating countries revealed an interest of several countries to subdivide mixed categories such as complex cultivation patterns, land principally occupied by agriculture and transitional woodland/shrub (European Environment Agency, 2021). The subdivision of the last category was the one that attracted more interest among the countries that were consulted. The detailed information provided by national LUC datasets, like SIOSE, could be useful when carrying out this task. Nonetheless, the generation of the new CLC+ may already give an answer to this problem. It will provide LUC information at finer spatial and thematic scales that the current CORINE (Kleeschulte et al., 2017). On the other hand, when obtaining CORINE through semi-automated methods, like in the 2012 and 2018 editions, more attention should be paid to the uncertainty attached to the mapping of mixed categories. In this regard, these categories refer to abstract associations of elements that usually operate at specific scales. For CORINE, as pointed out by García-Álvarez and Camacho Olmedo (2017), mixed categories easy to delineate and understand at that scale, can refer to a different landscape conceptualization at finer scales like SIOSE. It is the so-called individualistic fallacy problem (Cao and SiuNgan Lam, 1997). Now, when CORINE is generalized from a very high detailed LUC database (SIOSE Alta Resolución), the problem can be bigger and the obtained result more uncertain. This manuscript has been submitted for publication in “International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI'


Introduction
Land Use and Land Cover (LUC) information is of uttermost importance for many studies in different fields, from social to natural sciences (Green, Schweik and Randolph, 2005;Bontemps et al., 2012). Researchers from these fields demand accurate LUC data at different levels of spatial and thematic detail that are able to characterize regional, national and continental areas, or even the global surface (Giri, 2016;Nedd et al., 2021;García-Álvarez et al., 2022). 4 Through this paper, we aim to fill the previous research gap by analyzing in detail the coherence and uncertainties of the CORINE Land Cover time series for Spain. We build on the previous work carried out by García- Álvarez and Camacho Olmedo (2017) and Martínez-Fernández et al. (2019), which have assessed in detail the uncertainty of CORINE for the period 2006-2012, when the method of production of CORINE for Spain changed. Since 2012, CORINE is obtained through generalization of SIOSE, a fine-scale national LUC database. However, the SIOSE method of production has also changed since the 2017 edition (Equipo Técnico Nacional SIOSE, 2020), which may have introduced changes in the last CORINE release as well. Thus, through the present paper, we aim to shed light on the uncertainties associated to the last update of CORINE for Spain at the same time that we analyze the full coherence and uncertainty of the Spanish CORINE time series.
To that end, we first provide a brief presentation of the CORINE Land Cover dataset, its history and characteristics. Second, we make a brief introduction to the selected study area (Asturias). In the third part of the paper, we explain the methods employed in our analysis. Later, the results are presented and discussed. Finally, we provide a brief conclusion.

Corine Land Cover
CORINE Land Cover is a European LUC dataset whose production dates back to 1985 (Büttner, 2014). Since then, 5 different editions of CORINE have been produced for the reference years 1990, 2000, 2006and 2018(European Environment Agency, 2021. New updates of CORINE are expected every 6 years (Büttner, 2014).
CORINE is produced at the national level under the coordination of the EEA, which defines the common characteristics of the dataset and ensures its coherence in border areas (European This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' 5 Environment Agency, 2021). Nowadays, the CORINE production takes part of the Copernicus Land Monitoring Service, together with other relevant LUC products at finer (HRL, Urban Atlas…) and coarser scales (CGLS-LC100) (García- Álvarez and Florina Nanu, 2022). A new product based on the CORINE experience and complementary to it is currently being developed. CORINE Land Cover + (CLC+), also called the "2nd generation CORINE Land Cover", will provide a single repository of European and selected national LUC datasets, which will be integrated in a common grid following the EAGLE data model and nomenclature (Probeck et al., 2021). The created grid will be based on a new product (CLC+ Backbone) obtained through Sentinel imagery segmentation and auxiliary data, which will also classify the landscape through a 18-categories classification scheme (European Environment Agency, 2021). The new CLC+ dataset will allow its use for multiple purposes, including the update of the traditional CORINE dataset (CLC legacy). CORINE maps have been traditionally obtained in vector format through photointerpretation of satellite imagery at 1:100.000 scale (Büttner, 2014). However, in the last decade, an increasing number of countries, including Spain, obtain CORINE through the generalization of national LUC datasets at finer scales (Hazeu et al., 2016). The Figure 1 shows the workflow for updating CORINE in Spain. Independent of the production method, CORINE mapping rules remain the same: a Minimum Mapping Unit (MMU) of 25ha, a Minimum Mapping Width (MMW) of 100m and a 3-level classification legend with a maximum disaggregation of 44 categories (European Environment Agency, 2021).
In addition to the production of a CORINE status layer (CLC), for each mapped period (90/00, 00/06, 06/12, 12/18) a specific layer of changes (CHA) is obtained. This layer maps all LUC changes over the considered period at the same scale and with a Minimum Mapping Unit of This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' 6 only 5ha, except for the first period (90/00). Different to the comparison of the status layers, the CORINE layers of changes only map the changes that really happened on the ground, without any technical change due to mapping errors or changes in the method of production.
The CLC layers are obtained since 2006 from the revision of the previous CLC layer of reference and the superposition of the CHA layer (Fig. 1). In this regard, every time a new CLC status layer is produced, the one for the previous year is updated, accounting for the detected error and inconsistencies. Because of the different mapping rules between CLC and CHA layers, the updated CLC layer with CHA changes is generalized to fit with the CLC MMU rule.

Study area
The Principality of Asturias is one of the 17 Spanish Autonomous Communities, the first level of This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' 7 administrative division in Spain. It is located in North Spain, as part of the Cantabrian Coast (Fig.   2). Asturias is a mountainous region made up of a succession of mountain ranges and deep valleys plus a plain coastal surface where most of the activities and infrastructures locate (Cortizo Álvarez, Fernández García and Maceda Rubio, 1990). At the center of the region, in a topographically favorable area, the Asturias Central Area hosts the most relevant urban centers of Asturias as well as most of the population and economic activity (Rodríguez Gutiérrez, Menéndez Fernández and Blanco Fernández, 2009). Most of the artificial users and covers of the region are placed here and in the coast. On the contrary, the rest of Asturias and, especially, its mountainous areas, are dominated by rural and natural landscapes made up of pastures and forests (Rodríguez Gutiérrez and Menéndez Fernández, 2005). Because of the difficult topographic conditions, the agricultural activity of Asturias is mostly driven by extensive livestock farming, with limited surfaces of arable land (Cortizo Álvarez, Fernández García and Maceda Rubio, 1990).

Figure 2. Location map of Asturias
This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' 8

Methods
We analyzed the changes showed by the CORINE Land Cover database in both the CORINE status layers (CLC) and the CORINE layers of changes (CHA). To this end, we made use of the available CLC and CHA layers for Spain: CLC90, CLC00, CHA00, CLC06, CHA06, CLC12, CHA12, CLC18 and CHA18. Figure 3 shows a flowchart of the approach followed to analyze these datasets.
This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' 9 Figure 3. Flowchart of the analysis of CORINE Land Cover data carried out for this study CLC vector layers were overlayed in pairs, matching the periods for which CHA layers are available: 90/00, 00/06, 06/12 and 12/18. After the layers were overlaid, we separated those polygons that changed between dates from those ones that did not undergo change. Then, we analyzed the polygons that changed globally and per transition for each of the three levels of the CORINE classification legend. In each case, we calculated the area and proportion of the changes for all Asturias and with respect to the total quantity of detected changes for each period. Changes showed by CHA layers were analyzed in the same way: globally and per transition for all Asturias and with respect the total quantity of CHA changes for each available period.
To better understand the changes showed by the dataset and their nature, we have grouped the transitions of each CLC and CHA layer in five groups, according to the type of uses or covers involved: mixed covers, moors and heathland, pure forest covers, agricultural covers and artificial surfaces (Table 1). These groups include the most representative covers of LUC change in Asturias and reflect well the different nature and characteristics of the covers involved. Finally, CLC and CHA layers for all available periods were independently overlaid (CLC t1234 and CHA t1234). This allowed to count the number of times a specific area underwent change in the considered period and the plausibility of the change timeseries for each case. In this regard, when the CLC and CHA layers showed a change from category X to category Y and the next change transitioned from a different category than Y, this was considered a non-plausible change.
In addition, we counted the number of times the same transition (e.g. from X to Y) happened in the considered period.

Results
Except for the first editions of CORINE (90, 00), CLC layers always detect more LUC changes than CHA layers (Table 2). This is especially true in the 2012 and 2018 editions of CORINE: whereas CLC layers detect change in 36,5% and 7,9% of the mapped area for 2012 (06/12) and  (Table 3). The areas that undergo three or the maximum of possible changes (4), although exist, are quantitatively of little relevance. In CLC layers, most of the areas that change in more than one period, undergo those changes in the last editions of CORINE (2012CORINE ( , 2018. 70% of the areas undergoing two transitions experiment these changes in the periods 06/12 and 12/18. On the contrary, for CHA layers, almost 50% of the areas undergoing two transitions experiment those changes in the periods 90/00 and 00/06.   (Table 6).
This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI'

14
In 2012 and 2018, the exchanges among agricultural areas are one of the biggest changes detected by CLC layers (Table 5). These exchanges are not meaningfully mapped in any CHA layers, neither in the in previous editions of CORINE (1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006). In CLC layers, changes in agricultural areas represent a proportion of the mapped changes which is 7 (2018) and 3 (2012) times bigger than the proportion that the same type of changes represent in CHA layers and 3-4 times bigger than the same transitions in previous editions of CORINE (Table 5). The most important exchanges of agricultural covers mapped by CLC layers in 2012 and 2018 are the transitions from and to heterogeneous agricultural areas (that is, mixed covers), which in many cases are not even mapped at any extent in CHA layers (Table 6). Both changes of agricultural and forest and semi-natural covers are mostly driven by the dynamism of mixed categories. In this regard, the most relevant mapped transitions in any period in either CLC and CHA layers are from or to a mixed category (Table 4). Transitional woodlandshrub is the forest and semi-natural mixed cover accounting for all these changes. Among agricultural covers, complex cultivation patterns and land principally occupied by agriculture are the two categories accounting for all this change (Table 6).
Among artificial areas, mines, dump and construction sites are the covers that usually account for most of the mapped changes (

Discussion
The results prove how CLC and CHA layers provide very different information for Land Use Cover Change (LUCC) analysis. However, this difference is especially noticeable since 2012, after the change of production of CORINE Spain, which has introduced important limitations in the use of the CORINE temporal series. In all cases, mixed categories account for most of the mapped change, which may introduce new uncertainties in our studies because of the flexible nature of these categories and the difficulty to define them with precision. In the next sections, we independently address each of these issues.

The difference between CLC and CHA layers
The EEA and the CORINE production teams advise to use the CHA layers of CORINE for LUC change analysis as CLC layers include technical changes caused by the correction of detected errors or variations in the method of production (European Environment Agency, 2021). Our results prove how LUC change analysis from CLC and CHA layers may end in very different results and conclusions.
The only exception to that general rule are the CLC layers for the two first years that CORINE is available (1990,2000). Changes from the comparison of CLC layers for these years are the same than the changes mapped by the CHA layer for that period. This is explained by the specificity of the first CORINE update. The first CHA layer was obtained in some countries from the intersection of CLC90 and CLC00, after the revision of the first layer (Büttner, 2014;European Environment Agency, 2021). Thence, there are not differences between CLC and CHA changes as both were obtained in the same way.
Nonetheless, this means that CHA90 does not fit the 5ha MMU of CHA layers. Although there are some differences in the information that they show, the pattern and sizes of the changes measured by both types of layers is very similar. Therefore, disagreements can be attributed to the generalization process.
Opposite to the previous differences, CLC and CHA layers in 2012 and 2018 show very different transitions and sizes of changes. In addition, a relevant part of the changes mapped by CHA layers in those years are not included or coded differently in CLC layers.
In this regard, in 2012 more than half of the changed mapped in the CHA layers is wrongly represented or not represented in the CLC layer. These disagreements cannot be explained by the generalization carried out when producing the CLC layers. It can only be explained by the variations in the method of production of the Spanish CORINE, which affected not only to the areas that underwent change, but to all the mapped landscape (section 6.2).
Despite of the official recommendations to use CHA layers for change analysis instead of CLC layers, it is still common the use of the last ones for LUC change analysis (Hewitt and Escobar, 2011;Rusu et al., 2020;Fernández Nogueira, 2021;Gemitzi et al., 2021). As the first editions of CORINE (up to 2006) did not show big differences between using one or the other layers for LUC change analysis, users were not faced with the same This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' uncertainties than they find now. This may explain that the use of CLC layers to assess LUC changes has remained a common practice along time. Thence, users require of more and better information regarding the correct use of CORINE database. Although the EEA and the Copernicus programme already provide relevant documentation, the national authorities in charge of CORINE production and distribution, such as the Instituto Geográfico Nacional (IGN) in Spain, supply very limited information on the database and do not warn about all these issues. In this regard, the Spanish IGN does not even provide with the last release of CORINE the most updated product manual (European Environment Agency, 2021), where all these issues are addressed in detail.
Many users still require a coherent time series of CLC layers. In this regard, in many cases, the CLC time series is used as an auxiliary layer for different mapping methodologies or as a required input for different spatial analyses (Burkhard et al., 2012;Goerlich and Cantarino, 2013;Kucsicsa et al., 2019). Other users demand to study LUC change for larger periods than the ones between CORINE editions. The EEA has recently developed a coherent time series of raster CLC layers to fit these user needs: the CLC accounting layers (European Environment Agency, 2021). They are obtained by backdating the last CLC layer (2018) with the information provided by the CHA layers.
The CLC accounting layers present a triple limitation. First, they are only produced in raster format at 100m, which may not fit the requirements of all the CORINE user community. Although more time and resource-consuming, the development of a similar product in vector format, the traditional one in which CORINE is distributed, could better satisfy the needs of users and favor the correct use of the dataset. Second, CLC accounting layers present MMU inconsistencies, with many patches below the 25ha threshold (European Environment Agency, 2021). This is of little relevance for many This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' users of CORINE, but could be relevant for those users interested in knowing the distribution of land uses and covers through time, which need coherent measurement rules across the different years. Finally, when there are incoherencies between CHA layers, these are translated to CLC accounting layers as well. The CORINE production team has already tested a method to correct this inconsistencies, although it has not been approved and applied yet (European Environment Agency, 2021). Nonetheless, previous experiences with the two first CHA layers show the path to follow, which does not demand a lot of effort nor time (Maucha, Büttner and Pataki, 2011). In this regard, as our study showed, the number of changes that present these inconsistences is very low, only affecting a very small proportion of the detected changes in CHA layers. It would only require a small investment to correct these errors in the original CHA layers, avoiding these sources of uncertainty for any user of the CORINE database.

The changes in the method of production of CORINE Spain and their effects
Since 2012, CORINE has been produced in Spain from the generalization of the national LUC database of reference: SIOSE (García-Álvarez and Camacho Olmedo, 2017).
Before, it was manually obtained through photointerpretation of satellite imagery. That

The problem of mixed categories
When mapping landscapes at coarse scales, mixed categories are usually required to define areas made up of a mixture of different land uses or covers that cannot be attributed to a single one (Villa et al., 2008;Valcárcel Sanz and Castaño Fernández, 2012).
However, these categories come with important limitations, usually because of their  Maucha et al. (2011) analyzed the inconsistencies between CHA00 and CHA06 layers for all Europe. Mixed categories and, specifically, transitional woodland/shrub and complex cultivation patterns, were found to be behind most of the checked inconsistencies. Although this may be caused by the important dynamism of the areas mapped under these categories, it may there be a correlation between the use of mixed categories and the higher uncertainty of the mapping process. In fact, part of the inconsistencies detected by Maucha et al. (2011) were attributed to the subjectivity when mapping and differentiating these categories from other ones. The results of our study This manuscript has been submitted for publication in "International Journal of Remote Sensing". Subsequent versions of this manuscript, after acceptance and review by the journal, may have slightly different content. If accepted, the final version of this manuscript will be available via the `Peer-reviewed Publication DOI' and, specifically, the categories involved in the technical changes detected in the 2012 and 2018 editions of CORINE, point out in this direction.
The better characterization of the areas assigned to mixed categories could be a solution to this problem. In this regard, a survey carried out by the EEA among the CORINE participating countries revealed an interest of several countries to subdivide mixed categories such as complex cultivation patterns, land principally occupied by agriculture and transitional woodland/shrub (European Environment Agency, 2021). The subdivision of the last category was the one that attracted more interest among the countries that were consulted. The detailed information provided by national LUC datasets, like SIOSE, could be useful when carrying out this task. Nonetheless, the generation of the new CLC+ may already give an answer to this problem. It will provide LUC information at finer spatial and thematic scales that the current CORINE (Kleeschulte et al., 2017).
On the other hand, when obtaining CORINE through semi-automated methods, like in the 2012 and 2018 editions, more attention should be paid to the uncertainty attached to the mapping of mixed categories. In this regard, these categories refer to abstract associations of elements that usually operate at specific scales. For CORINE, as pointed out by García- Álvarez and Camacho Olmedo (2017), mixed categories easy to delineate and understand at that scale, can refer to a different landscape conceptualization at finer scales like SIOSE. It is the so-called individualistic fallacy problem (Cao and Siu-Ngan Lam, 1997). Now, when CORINE is generalized from a very high detailed LUC database (SIOSE Alta Resolución), the problem can be bigger and the obtained result more uncertain. the CORINE layers of changes (CHA). However, these also present small inconsistencies. Although not very relevant, they could be easily addressed to reduce the uncertainty of CHA layers to the minimum. For those users that require status layers instead of layers of changes, the CORINE production team has also developed a specific product in raster format at 100m that provides a consistent time series of LUC data: the CLC accounting layers. Notwithstanding, this product is still being improved and, although less uncertain than the standard CLC layers, users should be aware about their limitations.
In Spain, the change of production of CORINE since 2012 has meant a lot of changes in the database that require further analysis. It is required a specific study about the impact of the change of methodology in the mapping of changes. In addition, information should be provided about the last update of CORINE in Spain to explain the technical changes that have been found.
We have identified a general lack of transparency and detailed information on the uncertainties and changes of method of production of CORINE in Spain. Although