Scenarios of future land use/land cover changes: impacts on cropland use in Šiauliai region (Lithuania)

Abstract Agricultural systems supply a wide range of ecosystem services. Projecting future agricultural land-use changes is key to anticipating the potential impacts of human activities. In this work, we assessed future agricultural land-use changes, using the Dinamica-EGO platform, under three scenarios: A0-(business-as-usual), A1-(sustainable agriculture), and A2-(agricultural intensification) for 2040 in the Šiauliai region (Lithuania). Spatial autocorrelation was evaluated by using a Moran’s I index and the spatial patterns with the Getis analysis and landscape metrics. The results showed that croplands will increase 29.6% in the A0, 14.95% in the A1, and 29.63% in the A3 scenario. According to the Getis results, cold spots are in the surrounding of Šiauliai city, and hot spots in the northeast of the Šiauliai region. It was verified a high cropland fragmentation in A1 and low fragmentation in A0 and A2 scenarios. These results are critical for land management to understand cropland impacts under different scenarios.


Introduction
Human well-being greatly depends on the ecosystems' capacity to supply multiple ecosystem services (ES) (MEA 2005). Agricultural ecosystems supply a wide range of ES, such as food, fodder, fibre, and biomass, for energy (Glendining et al. 2009). Therefore, changes in the agricultural systems resulting from agriculture intensification, land abandonment, climate change, and urbanization may cause drastic changes in ES supply (Costanza et al. 1997;Pereira 2020). In the last decades, human activities have changed the landscape in response to their needs (Kaplan et al. 2017). The improvement in social systems, growing population and economy, and the increasing mechanization of agriculture accompanied by the introduction of new crops and agrochemicals have led to unparalleled growth in agricultural production (Ritchie and Roser 2021). This agriculture intensification has dramatic impacts on climatic change, biodiversity loss, land degradation (e.g. soil erosion, compaction, acidification, air and water quality) and landscape homogenization (Firbank et al. 2008).
The impact of climate change and land use/land cover (LULC) changes have uncertain effects on agriculture. Recent studies demonstrated that human activities and climate change are likely to affect near-future land suitability for various crops such as horticulture production (Bryan et al. 2016). Due to fluctuations in agricultural systems conditions (e.g. population increase or changes in food habits), agricultural land is vital to meet food demand and ensure food security (Fukase and Martin 2020). Global and European policies have priority to protecting and restoring degraded ecosystems targets. An example at the global level is the United Nations 2030 Agenda for Sustainable Development Goals (SDGs), mainly goal 15 (life on land), which has the objective of "protect, restore and promote sustainable use of terrestrial ecosystems, sustainably manage forests, combat desertification, and halt and reverse land degradation and halt biodiversity loss." 1 In the same line, the United Nations Decade on ecosystem restoration 2021-2030, aims to "prevent, halting and reversing the degradation of the ecosystems worldwide." 2 This is essential to guarantee food security for populations (P erez-Escamilla 2017). The EU Biodiversity Strategy to 2030 and the European Green Deal 3 aims to preserve and restore ecosystems, promote sustainable agriculture, and recover the loss of biodiversity (Montanarella and Panagos 2021).
Mapping and assessing the impacts of LULC changes on cropland supply gained increasing attention over the past two decades (Winkler et al. 2021). However, agroecosystems are dynamic, and there is a need to project future LULC changes to identify potential future impacts on agricultural activity better. For instance, Mora et al. (2020) analysed the impacts of future LULC changes on food security at the global level, and Rega et al. (2019) measured the trade-offs between agricultural production and agri-environmental indicators (e.g. nitrogen surplus) using different future scenarios at European level. Assessing future agricultural land changes is critical due to the need to produce agricultural products, ensure food security, and preserve the ecosystems (Mora et al. 2020;Sundstr€ om et al. 2014). Spatial scenario forecasts can improve our understanding of agricultural areas' status, making it possible to recognize both potential threats (e.g. loss of biodiversity, land homogenization) and conservation opportunities (e.g. protecting natural areas). One of the approaches that have been used is the application of spatially explicit models that allow us to understand the interactions between driving forces and their relation to LULC changes. Choosing a model for a specific study depends on the research purpose. The most widely used is the Markovian cellular automata (e.g. Gomes et al. 2019), the artificial neural networks (e.g. Morgado et al. 2014), and the agent-based models (e.g. Murray-Rust et al. 2014). These models assess the interaction between human behaviours and environmental systems and project the LULC changes future spatiotemporal dynamics based on past LULC changes. Spatial explicit models coupled with different scenarios have been applied to study future agricultural land-use changes and measure its environmental and socioeconomic impacts (e.g. van Vliet et al. 2015). Preserving croplands is critical to guarantee the sustainable supply of multiple ES. Despite the number and multiplicity of studies projecting future LULC changes trends, little attention has been paid to projecting future agricultural LULC changes at a local/regional scale at a high resolution, combining different land-use strategies, socioeconomic and climatic scenarios measuring its impacts. This paper addresses this gap by projecting 'what-if' scenarios in a predominantly agricultural region. The objectives are: (i) to quantify future agricultural changes between 2021 and 2040 under three scenarios: business as usual (A0), sustainable agriculture (A1), and agricultural intensification (A2); (ii) to recognize the main driving forces responsible for the LULC changes; (iii) to assess the cropland dynamics and patterns; and (iv) to pointing out spatial planning strategies.

Study area
The study area is located in Siauliai (Lithuania) (Figure 1) and has an area of 8539.8 km 2 . Siauliai region has a total population of 261,452 inhabitants, with a population density of 31/km 2 (2020) (OSP. 2021). Between 2000 and 2020 the total population decreased by 45%, as a result of population aging and rural exodus (OSP 2021). Grassland, woodland and forest, wetlands, rivers and lakes occupy approximately 42% of the total area in 2020. In 2020, the Siauliai was the region of Lithuania with the largest area covered by farmlands (OSP 2021). This region's per capita GDP in 2020 was 13,200 euros, lower than the national average (17,445 euros) (OSP 2021; Lithuanian Agriculture Department, 2021).

Methodology
The methodology of this study included six main steps: (i) land use data and accuracy assessment: satellite image classification for 2000 and 2020 and accuracy assessment were performed; (ii) designing narratives of land use scenarios (A0, A1, and A2); (iii) driving forces of land use scenarios: identification of driving forces that may affect future land use and land cover changes for the studied scenarios; (iv) model calibration: this step was supported by transition matrices, multicollinearity tests, and weights of evidence modelling; (v) model validation: the 2020 land use map was simulated (using the 2000 land use map as a reference year and a selection of driving forces, and applying the CA approach). Using this approach, the model was validated by comparing the 2020 simulated land use map with the 2020 land use reference map; and (vi) future land use changes projections: the projections under the studied scenarios were performed using a CA approach. Figure 2 provides a methodology flowchart of this research.
2.2.1. Land use data and accuracy assessment Satellite images from 2000 to 2020 were classified into land use maps. These images (cloud-free) were obtained from Landsat 7 and 8 at a multispectral high spatial resolution (30 m) and were downloaded from earthexplorer.usgs.gov (USGS 2021) ( Table 1). All images were rectified to European Terrestrial Reference System 89 (ETRS89). These images were acquired in the vegetation growing season (May 2000 and May 2020) because it makes it easier to distinguish the different types of land use and land cover (Shi and Li 2012). Considering the aims of this research, and following the Biodiversity Information System for Europe, which determines the linkage between land use classes/types and ecosystem types, land use classes were divided into the following six categories: 1 -urban (e.g. settlements, industry, transportation); 2 -cropland (e.g. both irrigated and non-irrigated land), 3 -grassland (e.g. lawns, and sod fields); 4 -woodland and forest (e.g. natural or planted forests, evergreen forest land, mixed forest land); 5 -wetlands (e.g. non-forested wetland); and 6 -rivers and lakes (e.g. permanent open water, lakes, reservoirs, streams, and estuaries).
The land use classifications were produced employing a standard parametric pixelbased supervised maximum likelihood classifier (MLC). Despite the existence of new machine learning algorithms, like Random Forests and Supported Vector Machines  (Rahman et al. 2020) that have proven to be also accurate (Rana and Venkata Suryanarayana 2020), MLC is still one of the most widely used methods in satellite image LULC changes classification (MohanRajan et al. 2020). Here, pixels were assigned to LULC changes classes employing probability outlines by considering both the class signatures' means and covariances (Hogland et al. 2013). MLC's main drawbacks are its low classification accuracy for classes with analogous spectral responses and/or comprising many subclasses (Kavzoglu and Reis 2008). Thus, this is not a problem in our six classes' nomenclature.
To determine if the land use classification is reliable, an accuracy assessment was performed by comparing for each year the land use classification map with reference maps that are assumed to be appropriate for the validation process. To do this, for each landuse class, training samples were chosen through the representativeness of each land-use class. To evaluate land use classification accuracy, an independent sample was randomly selected for each land-use class (and for each studied period). This allowed to prevent pseudo-replication and improve the training data variety. Therefore, 517 points in the 2000 land use map and 513 in the 2020 land use map were randomly selected proportionally in each land-use class to evaluate the land use classification accuracy. The validation data was managed manually based on Google Earth images, Google Street View data, Sentinel-2 data and Landsat data of different periods. Crosstabulation between land use classification and reference class (ground truth) was applied to evaluate land use classification accuracy. User and producer accuracy were estimated based on this analysis. Postland use classification improvements were employed to decrease classification errors triggered by the similarities in spectral responses of some land-use classes such as urban and cropland. Lastly, a majority filter using a 3 Â 3 moving window was employed for each land use classification to recode isolated pixels classified differently than the majority class of the window (i.e. salt and pepper effect). The software used to perform these methodological procedures was ArcGIS 10.7.

Designing narratives of land use scenarios
Scenarios provide assessments that may promote active discussions about future territorial policy alternatives (Gomes 2020). They can help to explore and spatialize uncertainties. In this study, we used spatially explicit projections of future agricultural land-use changes for 2020 and 2040 and investigated crop supply potential. Based on hypothetical futures, we designed three different scenarios, namely an (i) business as usual scenario (A0); (ii) sustainable agriculture (more conservative) (A1); and (iii) agricultural intensification scenario (more liberal scenario) (A2). These selected scenarios are based on Global, European, and national/regional levels. Specifically, at the global level, we derived the narratives from the Intergovernmental Panel on Climate Change (IPCC) (in which climate data from these scenarios were used). At the European level, we derived the narratives from the scenarios created by the Joint Research Centre. They analyse four scenarios for European Union agricultural production in 2040 and measure its impacts on the environment and biodiversity (Rega et al. 2019). On a national/regional scale, we extracted the narratives from the Rural Development Programme for Lithuania (European Comission 2020). This program aims to create new agricultural jobs, improve the farms' economic viability, and integrate farms into the food chain (European Comission 2020). A description of the narratives and their main assumptions are presented below: A0 business as usual: This scenario aims to recognize, according to near past trends (between 2000 and 2020), the socioeconomic, demographic, and climatic dynamics and to mimic and project them into the future (2040). No additional spatial planning policies are verified in this scenario. LULC changes under this scenario reflect an extension of 2000s-2020s trends, which were dominated by urban growth (þ71.31%), and cropland areas decrease (-9.85%) (these values were estimated for the Siauliai region and were calculated based on the land use classification produced in this study for 2000 and 2020). A1 sustainable agriculture: This scenario provides incentives to practice sustainable agriculture. Strict land protection and sustainable management with high biodiversity conservation are considered in this scenario. Local solutions to economic, social and environmental sustainability are implemented, and there is local self-reliance. Spatial planning policy is restrictive, in which agricultural and land use policies promote landscape multifunctionality. This scenario shows (i) changes in food habits (e.g. dietary pattern); (ii) increased importance of the agricultural markets closest to the urban centres; and (iii) strengthening of local agricultural production. A2 agricultural intensification: This scenario assumes a substantial increase in the demand for agricultural products. This may lead to pressures to expand agricultural land, increasing landscape homogenization and biodiversity loss. In this scenario, the nature protection policy is weak. The intensity of landscape protection decreases due to agricultural intensification. There are high investments in agricultural mechanization. An increase in temperature is observed (IPCC 2021), leading to a development in the agricultural area in northern latitudes.
Several driving forces were used to simulate future LULC changes in each scenario to formulate these scenarios. Therefore, differences between scenarios are obtained by different driving forces and parameter settings that affect the output of each spatial scenario.

Driving forces of land use scenarios
A list of driving forces for each scenario was prepared based on the characteristics of each selected scenario according to the literature review (e.g. Plieninger et al. 2016;and Gomes et al. 2018) and expert knowledge ( Table 2). The driving forces might differ with geographical location and consequently with the specificities of each territory, such as socioeconomic, demographic, environmental, political, proximity, and climatic drivers. Table 2 shows the selected driving forces for this study area and each studied scenario. As can be seen from Table 2, the driving forces and suitability of each one of them were similar for scenario A0 and scenario A2. This option resulted from the likely trend toward greater agricultural intensification as promoted by the Rural Development Programme for Lithuania (European Comission 2020). In the case of the A1 scenario, driving forces such as distance to main roads, distance to croplands, distance to woodland and forest, and distance to grasslands were not included because the distance for these drivers was considered not to be a determining factor in a sustainable agriculture scenario. Moreover, the slope and roughness drivers were also not included in the A1 scenario since, according to the literature, they are not considered limiting factors in a sustainable agriculture scenario (Tarolli and Straffelini 2020).

Model calibration
The calibration applied in this study helped to adjust the model parameters and to improve the model's reliability. Therefore, different methodological approaches were used such as transition matrices, multicollinearity tests, and the weight of evidence method to determine the transition probabilities of each land use class. The different methodological steps are detailed below: i. Transition matrices: the past transition matrices were estimated by employing the probability of change to mirror the percentage of LULC changes between 2000 and 2020 (using the multi-step transition matrices). ii. Multicollinearity tests: to estimate the multicollinearity of the spatial driving forces, two measures supported by the crosstabulation matrix were used: the Cramer coefficient and a measure that belongs to the class of entropy known as joint information uncertainty. Values below 0.5, in both coefficients, correspond to a non-collinearity of the drivers (Bonham-Carter 1994). iii. Weight of evidence method: using this method the potential land-use transitions between the different land use classes were recognized. This is a vital process to recognize the importance of the driving forces that may cause each LULC changes. It is based on the Bayes theorem of conditional probability and the concept of posterior probability (Bonham-Carter 1994). It computes transition potential maps for each land use transition and allows to rank each driving force's importance in each transition (Gomes et al. 2021b).

Model validation
The reciprocal similarity comparison method was applied to compare the simulated land use map of 2020 (using 2000 land use as a base map and the selected driving forces shown in Table 2 for the A0 scenario) with the 2020 observed (reference) land use map. This method uses expanding analysis windows (1, 3, 5, 7, 9, and 11) to recognize spatial patterns within a specific cell vicinity (Mas et al. 2014). This method has been widely used in previous works, and has been applied to quantify the model's predictive reliability (e.g. Yang et al. 2019). The data obtained in the simulated model was compared to the observed data. The output ranges from 0 (total disagreement) to 100% (total agreement). The average similarity values of the 16 model iterations were visualized as a fuzzy similarity curve. When it is above 80% (Viera and Garrett 2005), the output is satisfactory, and future LULC changes projections for the different studied scenarios in 2040 can be projected.

Future land use projections
To project the future LULC changes for each scenario, we applied the Dinamica EGO (De Oliveira et al. 2020), a spatially explicit land-use model. This model is an exploratory tool to empirically simulate relationships between land use and driving forces and simulate future LULC changes under the A0, A1, and A2 scenarios.
The quantity the rate of LULC changes in each transition a potential map of change of land use was modelled through Markov Chain modelling. This random process is discrete in both state and time and uses a mixed approach between the observed past trends and neighbourhood effect. This model has been widely used to project future LULC changes (e.g. Sun et al. 2018) and has been employed to analyse the transition probability between an initial and a final state. Moreover, it comprises a series of transition-potential maps that characterizes spatial susceptibilities of changes from one land-use class to another. This approach was applied using the Dinamica EGO by setting the 2020 land use map (observed) as the base map of the future LULC changes projections. The simulation for each projection encompasses 19 interactions corresponding to annual steps between 2021 and 2040. Each scenario has its specific characteristics: (i) the A0 follows past trends to predict the future; the (ii) A1 scenario is more restricted regarding the agricultural intensification and landscape transformation; and the (iii) A2 is more liberal, allowing for greater intensification of agricultural activity. Thus, different parameterizations were used for each land-use transition and scenario. In Dinamica EGO, the sizes and percentages of new patches were set according to a lognormal probability function, where the parameterization was specified by the mean patch size, patch size variance, and isometry. This means that the parametrization can be modified to generate different future LULC changes and patterns. For instance, (i) an expansion in mean patch size results in a lessfragmented land; (ii) an expansion in the patch size variance results in a more heterogeneous landscape; and (iii) an isometry greater than 1 result in more equal patches. In practice, each of the parameterizations attributed to each land-use transition and scenario could be the result of a more conservative (limiting the increase in the size of agricultural parcels) or more liberal policies (allowing the parcelling of agricultural parcels) (Table S1-7).

Spatial and statistical analysis
The percentage of croplands per eldership was estimated for the reference years (2000 and 2020) and each scenario (A0, A1 and A2). A Kruskal-Wallis ANOVA was applied to assess the statistical differences between the reference years and the scenarios. Significant differences were considered at a p < 0.05 using Statistica 10.2. ArcGIS 10.7 was used to calculate the mean values at the eldership level. Data was mapped using the same number of classes (five) and the classification method (quantile). Spatial autocorrelation was carried out using the Moran's I index. The values range from À1 (dispersed pattern) to þ1 (clustered pattern). If the values are around 0, there is a random pattern (Moran 1950).
To identify cropland spatial patterns at the eldership level, a Getis Ord Ã hot/cold spot analysis was applied (Getis and Ord 2010). Lastly, a set of landscape metrics was calculated to recognize cropland dynamics and patterns. Specifically, landscape metrics helped to measure the structure and pattern of croplands in 2000 and 2020, and in each scenario. Farina (1998) states that to assess land fragmentation, it is required to recognize five different patch characteristics, namely, shape, edge, patch size, isolation/proximity, and connectivity. Following this suggestion for each patch characteristic, we selected the following metrics presented in the LecoS -Landscape Ecology Statistics plugin (QGIS version 3.16): mean patch shape ratio, edge density, mean patch area, and patch cohesion index (Table  3). To create spatial patterns, k-means clustering was performed.

Calibration: accuracy assessment of land-use classification maps
Land use classification accuracy in 2000 and 2020 were above 90%, which represents a good reliability (Kafy et al. 2022). Table 4 shows the outcomes of the user and producer assessment to test the spatial agreement estimated for each land-use class by comparing the reference and the classified land use maps.

Land use classification 2000 and 2020: past and present LULC changes
Land use classification results of the pre-processed satellite images for 2000 and 2020 in the Siauliai region are shown in Figure 3. Cropland was the largest LULC changes in the study area. The cropland area was 5336.26 km 2 and 4810.78 km 2 in 2000 and 2020, respectively (during this period, the cropland area decreased by À9.8%) (Figure 3). However, the northeast part of the Siauliai region still has a substantial area covered by croplands (Figure 3). Urban land use increased from 88.68 km 2 in 2000 to 151.93 km 2 in 2020. There was also an increase in the woodland and forest land from 1562.4 km 2 in 2000 to 2516.85 km 2 in 2020 (þ61%) ( Table S8).

Calibration and validation: accuracy assessment of future LULC changes
None of the Cramer coefficient values and joint information uncertainty was higher than 0.5 in both coefficients. Therefore, no driving forces were excluded in the future LULC changes projection. The reciprocal similarity comparison method results revealed a good similarity between the observed and the predicted 2020 land use map. The overall accuracy (with a window of 11 Â 11) was 0.94 ( Figure S1). The results for the calibrated model were satisfactorily validated. Table S9 shows the multi-step transitions from 2000 to 2020 estimated in the Dinamica EGO platform and used as an input in the LULC changes simulations. This matrix shows the rate of one land-use class i.e. urban areas, cropland, grassland, woodland and forest, wetlands, and rivers and lakes, changing to one of these land-use classes based on past land-use trends per united time step (1-year). The only exceptions in which there were no land-use transitions  were: (i) from urban areas to any landuse class; (ii) from grassland to rivers and lakes; (iii) from rivers and lakes to grassland. This matrix was used to count the number of cells to be changed (Table S9).

Weights of drivers of change (A0, A1, and A2)
The comparison between the A0 scenario and the current condition (2020) shows that croplands are likely to increase in 2040. For this transition in this scenario and from the conversion from grassland to croplands distance to main roads, distance to woodland and forest, and distance to rivers and water bodies were the most influential driving forces (Table S10). Moreover, the conversion from woodland and forests to croplands, slope, and distance to rivers and water bodies were the main driving forces identified (Table  S10). In the A1 scenario, distance to urban areas and distance to rivers and water bodies were the main driving forces responsible for the conversion from grassland and woodland and forest to cropland (Table S11). Lastly, in the A2 scenario, slope, distance to main roads, and land fragmentation were the main driving forces recognized as those responsible for converting woodland and forests and grassland to croplands (Table S12).

Cropland area assessment
The Kruskal-Wallis ANOVA identified significant differences among the reference years and scenarios in cropland %. The area was significantly higher in A0 and A2 scenarios than in 2000, 2020 and A1 ( Figure 6). Figure S2 shows the percentage of croplands per eldership in the Siauliai region for all scenarios. The elderships with the highest percentage of croplands in the three scenarios are located mainly north, northeast and east of the Siauliai region, especially in A0 and A2. Moran's I index showed a clustered pattern in 2000 and 2020, and all scenarios (Table 6). According to the Getis Ord Ã hot/cold spot analysis ( Figure S3), hot and cold spots have a similar spatial distribution in 2000 and Figure 4. A0 -business as usual; A1 -sustainable agriculture; and A2 -intensive agriculture scenario. Land use classes: 1 -urban; 2 -cropland, 3 -grassland; 4 -woodland and forest; 5 -wetlands; and 6 -rivers and lakes. 2020 in all scenarios. Cold spots were mainly located in Siauliai city and the surrounding elderships (e.g. Ku ziai, Kur s_ enai in the west part, and Kairiai and Tyruliai in the east part of Siauliai city) ( Figure S4). Hot spots have a similar pattern and were located in the northeast part of the Siauliai region, particularly in the elderships of Pa svitinys, Kepalia, and Joni skis. These elderships showed a confidence level greater than 99% ( Figure S3). Generally, there was a higher cropland fragmentation in the A1 than in scenarios A0 and A2 scenario (Figure 7).

Past LULC changes
The accuracy observed in the land use classified maps (above 0.9) suggested that according to the literature, they have good reliability (Jansen et al. 2008), as observed in other studies (e.g. Ahmad and Quegan 2012;Paris et al. 2019). This means that the land use classification maps are reliable with these values.
A decline of cropland areas by À9.8% was verified between 2000 and 2020. This decrease was generalized across the region, despite the European incentives that started to be applied in this period (Renwick et al. 2013). (e.g. Common Agricultural Policy incentives. Lithuania joined the European Union in 2004). This cropland decline is consistent with the observed in other studies conducted in eastern Europe, e.g. Estel et al. (2015), Lesiv et al. (2018), and Navarro and Pereira (2015). The cause of cropland decline was attributed to several factors, such as (i) the result of the breakdown of socialism leading to widespread abandonment of agricultural activity; (ii) an ageing agricultural population (Cvitanovi c et al. 2017;OSP 2021); (iii) land abandonment (Hartvigsen 2014); and (iv) population decrease (Bell et al. 2009;OSP 2021).

Future LULC changes
The validation process is an important step to increase the reliability of the outcomes achieved for the projected LULC changes. The achieved results following this procedure were higher than 90%. According to previous works, this indicates a high level of agreement (Chang et al. 2021). Therefore, the projected land use for 2020 is suitable to project future Spatio-temporal LULC changes.
Under the A0 scenario, it is expected that cropland areas will increase by 29.6%, and this transformation will be at the expense of grasslands and woodlands and forests. From grassland to cropland, this conversion was mainly verified in the south and southwest of the region, and from woodland and forest, conversion was verified throughout the whole region (Table S13). These land-use transitions are consistent with previous works in a business as usual scenarios, such as in a study developed by Gomes et al. 2021a in Lithuania and by Newton et al. (2021) in England.
Under the A1 scenario, it is projected that croplands will increase by 14.95%. This increase was verified at the expense of grasslands (526.72 km 2 ) and woodlands and forests (520.90 km 2 ) (Table S14). These new croplands in this scenario appear close to consumption areas due to the expected development of urban agriculture, essential to cities' sustainability (e.g. Ferreira et al. 2018). Distance to urban areas was the main driver of change responsible for the transition from grassland to cropland (Table S11). The conversion takes place in the southwestern part of the region and particularly around the city of Siauliai. This is relevant in a sustainable agriculture scenario since the proximity of crop production to urban centres is critical, as demonstrated by Feola et al. (2020), andNicholls et al. (2020). These authors showed that if croplands are located close to living areas, agriculture is more sustainable. These principles of sustainable agriculture or others, such as the preservation of biodiversity, the promotion of organic farming, or better soil management, have been promoted in the Lithuanian context by the Rural Development Programme (European Comission 2020; Mierauskas 2020).
Lastly, under the A2 scenario, it is expected that cropland areas will increase by 29.63%. This scenario is the one that demonstrates a high cropland area growth. However very close to what is expected in the A0 scenario. The Kruskal-Wallis ANOVA results confirmed this. Cropland % area was significantly higher in A0 and A2 scenarios than in 2000, 2020 and A1. This is directly explained by the driving forces selected for each of these scenarios and by the parametrization that resulted from the guidelines promoted by the Rural Development Programme for Lithuania, which involves increasing the size of farms to promote economic competitiveness (European Comission 2020).
The increase of cropland was mainly verified at the expense of woodlands and forests (832.93 km 2 ) and grasslands (673.65 km 2 ) ( Table S15). The conversion from grassland to cropland was mainly observed in the southwestern part of the Siauliai region. The conversion of woodland and forest to croplands was identified in all studied area, especially in the west and east, away from the main urban areas (Figure 4 and 5). The main driving force for the increase in cropland areas was the slope (Table S12), showing that the flat areas are more suitable for agricultural intensification. Previous studies have observed this (Widiatmaka et al. 2016;Chaplin-Kramer et al. 2015). Another driving force with some weight, particularly in the conversion from grasslands to croplands, was the distance to main roads (Table S12). This driving force is critical for an easier flow of products to the markets, as observed by Petit et al. (2011) and Persyn et al. (2022).

Future cropland dynamics
Morans' I result highlighted that the spatial pattern was clustered in all the scenarios, especially in the A1 scenario, showing that cropland areas were less dispersed than A0 and A2 scenarios. The differences were minimal, and this was confirmed by the Getis-Ord Gi Ã hot/cold spot analysis, where a high cold-spot was observed in the elderships surrounding Siauliai city, and this is very likely attributed to the urban development that occurred in this area and the consequent decrease of croplands as identified in other cities (e.g. Badreldin et al. 2019;van Vliet 2019). We identified a hotspot clustered in the northeast part of Lithuania in all the scenarios due to the substantial area covered by cropland, even in a sustainable agriculture scenario. A strong agriculture intensification in this area was also observed in previous works (e.g. Gomes et al. 2021b). In the future land use scenarios, landscape homogenization was particularly evident the A2 scenario croplands due to woodland and forest and grasslands loss. This landscape transformation was also identified by Zabel et al. (2019) and Garrett et al. (2018) as a consequence of agriculture intensification. Landscape homogenization has negative impacts on biodiversity loss  (Buhk et al. 2017), habitats loss (Nowakowski et al. 2018) and ES degradation (Zhao et al. 2015;Pereira 2020). Landscape heterogeneity was high in scenario A1, where croplands are mixed with woodland and forest and grasslands. A diversified landscape is essential to biodiversity (Morelli et al. 2013) and essential to ensure food production and security (Ntihinyurwa and de Vries 2021).

Limitations and uncertainties
As in previous modelling works (e.g. Gomes et al. 2021b), this work had limitations and uncertainties. Climate data had a different resolution from the other data, and this is an important limitation that may have imposed some uncertainties on the model outcome. Also, the usage of data from different years has the same effect. However, these problems were also found in previous modelling works (e.g. In acio et al. 2020). Land use classification maps for 2000 and 2020 were designed based on applying a maximum likelihood classification method, and due to the diversified and complex land-use patterns of the Siauliai region may have caused some errors or misclassification of the Landsat images as demonstrated by Kantakumar and Neelamsetti (2015) and Manandhar et al. (2009). We applied the post-processing function to overcome this limitation, as demonstrated in the results sections. Also, it can be challenging to recognize the causes and consequences of LULC changes simulations from complex spatial models due to the several interactions and feedbacks intrinsic within such spatial models. Therefore, the future LULC changes simulations might not have considered all potential driving forces. The use of 'what-if'scenarios (Acosta et al. 2018) is one of the biggest challenges to identifying possible future alternatives. Therefore, the 'what-if' scenarios should be understood as potential images of alternative futures by creating awareness of the potential impacts of alternative agricultural land use options. As with other agro-systems in eastern Europe, the future dynamics of the transformation of cultivated areas in the Siauliai region are overshadowed by uncertainty about agricultural land use policies (Prishchepov et al. 2012) and climate change (Vanschoenwinkel et al. 2016). This can lead to different landscape transformations of the landscape beyond those that are contemplated in these scenarios. The last limitation and uncertainty are linked to the future legislation of Lithuania that, for instance, may limit or facilitate cropland intensification in specific regions, which can affect the results of the land-use scenarios (Mik sa et al. 2020).

Implications for management
The work results showed an increase in cropland area under A0 and A2 scenarios, showing that the A1 scenario in the Siauliai region is affected by a substantial agriculture intensification. Although it is essential to ensure food security as highlighted in several works (e.g. Prosekov and Ivanova 2018;Viana et al. 2022), under A0 and A2 scenarios the increase of cropland supply ES will occur at the expense of biodiversity loss (e.g. Delzeit et al. 2017), regulating ES degradation (e.g. air, climate, flood and erosion regulation, carbon sequestration, water purification) (e.g. Liu et al. 2018;Pereira et al. 2018), another provisioning ES (e.g. medicinal plants, wild food) (e.g. Bengtsson et al. 2019), and cultural ES (e.g. landscape aesthetics) (e.g. Kalinauskas et al. 2021). Several works highlighted the trade-offs between cropland supply and other ES (e.g. Yang et al. 2020). It is critical to develop policies that support the A2 scenario and contribute to sustainable agriculture practices that avoid land consumption and the use of agrochemicals. For instance, at the global level, this will be beneficial to support SDGs, mainly goal 15 (life on land; reduce biodiversity loss and land degradation), but also another goal such as goal 3 (good health and well-being; by producing food with quality, free of agrochemicals), goal 6 (clean water and sanitation; by reducing drinking water contamination), goal 13 (climate action; by reducing the greenhouse gases released into the atmosphere) and goal 14 (life bellow water; by reducing diffuse pollution and fresh and marine water contamination, e.g. eutrophication) (Viana et al. 2022). At a regional level, a sustainable agriculture establishment is also aligned with EU Biodiversity Strategy to 2030 and the European Green Deal that wants to provide to EU citizens, among other things "fresh air, clean water, healthy soil and biodiversity" and "healthy and affordable food" 4 . It is also clearly highlighted that it is key "to ensure food security in the face of climate change and biodiversity loss 5 ". There are an important number of works showing that sustainable agriculture practices are not less productive than conventional ones as highlighted in a metaanalysis carried out by de Ponti et al. (2012). For instance, under drought conditions (that are expected to increase in a climate change context), organic matter has a high capacity to retain water, and the yields of organic farms are higher than conventional ones (Gomiero 2018). Therefore, there is a need to make a transition to sustainable agriculture police that protects soil quality as highlighted in the European green deal strategy "lead a global transition towards competitive sustainability from farm to fork" 4 . The European Green Deal also mention that it is key to "reduce the environmental and climate footprint of the EU food system" 4 and the continuous use of conventional agricultural practices is very detrimental to biodiversity loss, land degradation and climate change (Pereira et al. 2018;Viana et al. 2022). Finally, the European green deal also mentions a need to "strengthen the EU food system's resilience" 4 , which can only be a reality if we use practices that avoid land degradation and ensure the sustainable and continuous supply of food. For instance, in the EU, approximately 12 million hectares of agricultural areas are affected by severe erosion and loss per year of about 0.43% of their productivity (Panagos et al. 2017). This has a substantial economic cost: 1.25 million euros per year. Likely, sustainable agriculture practices could reduce these numbers.

Conclusion
Incorporating geospatial techniques employing remote sensing and spatial data to project future LULC changes can help identify the driving forces responsible for future land-use transitions. This study assessed the potential impacts of future agricultural LULC changes under different spatial scenarios projected for 2040 in the Siauliai region. Trends, patterns, and future impacts on agricultural LULC changes were assessed. Overall, we observed (i) a smoother and more compact growth of croplands in the A1 (þ15%) and (ii) a more accentuated growth of croplands (þ 30%) in both the A0 and the A2 scenarios. Getis Ord Ã hot/cold spot analysis showed that (iii) cold spots mainly were located in the Siauliai city and on its periphery, and (iii) hot spots were located in the northeast part of the Siauliai region. We also identified a high cropland fragmentation in the A1 scenario than in the A0 and A2 scenarios. This revealed how an agricultural region can face very different future land-use trajectories. Land-use strategies can directly or indirectly affect landscape transformation, i.e. they can restrict or promote land-use conversion from one land use to another. The results shown in this study for the different scenarios might be employed as a guide for decision-makers to improve land use planning strategies in a medium-term horizon to balance cropland supply and landscape conservation measures. Understanding future LULC changes can aid to indicate an ideal land use scenario. From the point of view of land use sustainability, scenario A1 could be chosen as the best solution by decision-makers, as it promotes a compact urban growth and cropland heterogeneity.