1. Introduction
Substances of anthropogenic origin have a major influence on the climate system [
1]. Human economic activity contributes to thermal atmospheric pollution—increasing greenhouse gas (GHG) concentration enlarges the natural greenhouse effect and plays a decisive role in the rise of the average global temperature [
2,
3,
4]. GHGs are mainly generated by the combustion of fossil fuels in industrial and agricultural production processes, and, by a large proportion, from waste [
3,
5,
6,
7,
8,
9]. GHG absorption is usually associated with the physiological properties of green vegetation, as other solutions to sequester carbon have not yet been proven to be either technologically or economically effective [
10,
11]. Meanwhile, climate change is a global issue and needs to be addressed through global cooperation among countries to improve energy efficiency, develop and deploy clean technologies, and increase natural GHG absorption. In this context, the processes in and around land use, land use change, and forestry (LULUCF) are becoming crucially important. The LULUCF sector includes GHG emission and its removal from forests, arable or producing land, grasslands and pastures, wetlands, built-up areas, and other land plots. Emissions and removals of GHGs are accounted using internationally accepted approaches [
12,
13,
14]. However, in order to actively increase carbon absorption, it is necessary to know and manage the processes involved in the development of land surface layers and land use. Cognitive processes and management decisions will be hampered by a lack of access to scientifically based tools for modeling land use and hence GHG emissions.
Nowadays, many land use change modeling tools exist, differing in their methodological implementation [
15,
16]. They may cover universal or very specific application fields, with the focus on local case studies or continental exercises. There are several concepts of land cover and use modeling [
17]—economic models, system dynamics approaches, cellular automata, and agent-based models. Spatial economic or econometric models deliver generalized predictions of states of phenomenon by balancing various inter-related input factors determining their development. System dynamics or causality-driven models assume an empirical modeling of land cover or land use changes. This involves (i) an assessment of past changes first, (ii) a determination of relationships between land changes and factors driving such changes, (iii) an evaluation of change potential, and (iv) an allocation of land to the new land cover or land use types [
18]. Cellular automata usually operate in a raster domain, representing the landscape as an n-dimensional grid of cells. Each cell may acquire a finite number of states, which may change over time following some set of rules and depending on the state of neighboring cells. Models are iterated over time, delivering land cover or land use status within the cell at specific times [
19,
20]. Agent-based models are aimed at modeling the behavior of autonomous individuals (agents) who may perceive their environment and interact with individuals [
21]. Even though there are numerous potential solutions for land use change modeling, their applicability is heavily restricted by various legal, technological, and organizational aspects. The land use change modeling depends on the specific requirements of GHG emission accounting, the availability and specifics of input data, modeling tools, and experiences, especially when considering specific countrywide exercises.
There are many factors influencing GHG emissions and absorptions in the LULUCF sector, potentially resulting in uncertainties in both GHG accounting and projections [
22,
23,
24,
25,
26]. Simultaneously, availability, or often the lack of input data for land use change analysis, makes the task more challenging [
27]. Even though there are international standards to account for GHGs, there are always some specialties present in the operational approaches of each country. Lithuania, following its international climate change mitigation commitments, has developed an original LULUCF monitoring system, which is used for GHG reporting. This system predetermines the approaches of land use development projections. The core data source for GHG accounting from the LULUCF sector in the country is the National Forest Inventory (NFI), which is implemented by the State Forest Service [
28,
29]. Originally developed to provide statistical information on forest resources for strategic forestry planning at a country level, the Lithuanian NFI has recently been expanded to collect countrywide data on land uses and land use changes. The land uses are monitored in a systematic network of observation points through the whole country, while forest attributes are surveyed at points in the forest. There are operational solutions introduced in Lithuania to model the development of forest resources and forestry, ranging from forest stand-level simulators to systems manipulating aggregated countrywide data [
30,
31,
32]. The State Forest Service uses the European Forestry Dynamics Model (EFDM), developed as a harmonized forestry modeling tool for all European countries, based on NFI data. The EFDM has been used to calculate the forest reference level (FRL) for Lithuania following the European Union LULUCF regulation for 2021–2030 [
13]. The EFDM is a matrix-based model of a Markov chain type representing change by transition of areas (in this case, the NFI sample plots) between different fixed states of the forest [
33]. This matches well with the system dynamics or causality-driven models introduced above. The reference levels for land uses other than forest are based on historical data, thus, one may assume that no sophisticated modeling solution is needed. Nevertheless, successful land use management provides challenges for modern decision-support tools which are based on land use development scenarios. To our knowledge, the solution that has been widely used to make GHG projections in the LULUCF sector in Lithuania has been the land use, land use change and forestry emission accounting tool, LULUCFeat [
34]. LULUCFeat delivers GHG predictions based on aggregated LULUCF data and past trends, using information on driving factors and expert knowledge. Methodologically, this fits the economic models mentioned above. However, the solution is too focused on delivering certain GHG reports and underfitting expectations for a versatile land use change modeling system, based on all NFI data and compatible modeling principles.
Thus, the aim of the study introduced in this paper is to test the methodological principles for land use development scenario modeling for use in processes of GHG accounting and management. First, we ask what is the performance of the Markov chain analyses methodological approach in modeling land use development using standard GIS software? To conduct the modeling exercise, we use inputs available from already running in Lithuania inventory projects and freely available geographic databases. Then, we test the capacity of the LULUCF sector in Lithuania to accumulate carbon during the next decade, starting in 2020. For that, we project the development of major land use types in Lithuania until 2030 using several land use management scenarios and estimate potential contributions of different land uses on carbon emission/absorption. We hypothesize that the carbon accumulation in the LULUCF sector in Lithuania during the next decade should increase. Finally, we end with a discussion and proposals for both methodological enhancements of modeling solutions and land use management policies.
4. Discussion
There are two parts in the discussion that follows. First, we briefly address the choices, findings, and limitations related to the methodological approaches used to predict the land use changes. Second, we use our predictions to discuss potential land use development in Lithuania and the related land use policy implications.
The choice of methodological land use modeling approach was influenced both by the specificity of the general modeling environment and scientific considerations. First, the focus has been on the needs for current land use change modeling in Lithuania. We associate our study with the needs related first to the management of GHG emissions and absorption in the LULUCF sector. Thus, the input data on land use was based on information collected on a network of systematically distributed samples, inheriting differing estimation accuracies for specific land use types. Similar modeling studies usually focus on wall-to-wall land covers or land uses, even though they may be of rather coarse resolution, such as the CORINE database [
27]. On the other hand, pointwise land use estimates may make the availability of information driver variables easier, as we do not necessarily need to map wall-to-wall all the factors influencing land use development. Moreover, many driver variables used in the current study are extracted using distance-based, focal, or zonal GIS analysis. Our approach is to use only publicly available information on driver variables, usually downloadable from the spatial information portal of Lithuania (
www.geoportal.lt) or freely available from authorized institutions based on license agreements for research and education use. Unfortunately, we could not acquire data on land ownership, which use is commercialized using legal regulations. Finally, our methodological approach had to be compatible with that used by Lithuanian authorities to implement their international commitments, including the European Union land use, land use change, and forestry regulation for 2021–2030 [
13], that is, we choose a modeling engine that is compatible with the EFDM, which has been used to calculate the forest reference level for Lithuania and already used to facilitate forest policy building processes. Last but not least, the exercise was implemented using standard GIS software packages, including both data engineering and modeling, i.e., not requiring extra efforts for software development.
The prediction of land cover or land use development in general, and the use of models of the Markov chain type in particular, has always been a very challenging exercise. The most difficult task is to evaluate the transition potential from one land use type into all possible other types. Numerous methodological approaches are reported, such as the weight of evidence [
42], empirical probabilities [
43], logistic regression [
44], and neural network modeling [
38,
45]. Usually, the results achieved using different solutions are rather different, since studies address very different land use change tasks. Nevertheless, priorities are given to the use of the MLP algorithm, which is a type of neural network [
18]. This was also used in our study. Two other methods implemented in TerrSet LCM, a similarity weighted instance-based machine learning algorithm and logistic regression, were rejected in the initial stages of our study, mainly due to the ability to model only one transition at a time and because additional complexity in the modeling exercise did not increase prediction accuracy.
The best prediction accuracies were achieved for land use types that followed linear development trends in recent decades, i.e., forest and built-up lands. Forest land change trends were most stable not only during the modeled period but also throughout the entire accounting period. Afforestation/deforestation is a relatively slow process in Lithuania [
46], strictly regulated from a legal point of view, and therefore potentially the easiest to predict. Similarly, the development of built-up areas has also been steadily increasing since 1970. The areas of producing land were increasing in Lithuania since the country joined the EU, usually at the expense of grassland. Thus, the prediction of producing land and grassland changes is very important in supporting land use policies, especially for GHG management, because producing land is associated with carbon emissions, whereas grassland, in contrast, contributes to the carbon accumulation on average [
47]. Usually, the misclassifications of producing land as grassland and vice versa were the main types of prediction errors, e.g., ~16% of producing land points were wrongly predicted as grasslands and ~12% of grassland points were wrongly predicted as producing land. Identification of grassland management intensity or differentiating between, e.g., cultural and natural grassland, has always been a challenging task [
48]. Using nomenclature for grassland subtypes that is too detailed has resulted in lower prediction accuracies because the grassland types are mixed with each other. Land use type in the Lithuanian inventory system refers to the center point of the sample plot, so the presence of single trees or brush may also be neglected during the inventory, unless the land has not yet been converted into forest land. The increase of forest land is usually very strictly controlled during the inventory, which has always been first focused on evaluating forest resources and involves precise measurements of individual trees on 500 m
2 circular plots [
36]. Thus, we continued without specific grassland subtypes. No significant differences were found among the results obtained using different six-factor inclusion strategies for modeling transformation potentials. We explain this by the performance of the MLP, which is a type of machine learning algorithm. However, the number of input driver factors is limited in TerrSet LCM. Therefore, it is suggested that, in the future, the maximum amount of supporting information is collected and used in modeling the selected driver factors that are most strongly related to the land use transformations.
All scenarios tested suggested that the LULUCF sector in Lithuania will accumulate carbon, basically due to carbon accumulation in the forest land. Thus, a further increase of forest land area is extremely important to further contribute to GHG absorption. Nevertheless, none of the scenarios resulted in a forest land area percentage in the country exceeding 37% in 2030. According to official forestry statistics, forest land covered 33.7% of the country’s area in 2019 [
49]. Our prediction introduces some questions for official forest and land use policies in the country. The political objective is set to increase the forest land area in Lithuania by year 2030 to 23,000 km
2, i.e., 35% of the country’s area [
50]. Increasing the forest land area proportion in Lithuania is also among the key objectives of national forest policy, primarily associated with the management of GHG emission/absorption [
51]. Abandoned or unsuitable agriculture lands are usually identified as afforestation targets in regulations for afforestation and reforestation [
52]. In parallel, deforestation is strictly controlled and legally possible only upon the compensation of expenses for establishing new forest land [
53]. Therefore, our simulations confirmed that the political afforestation targets can be achieved by 2030. There are no scenarios that suggested forest land reduction, yielding steadily increasing GHG absorption potential. Even though the GHG accumulation in forest land is increased most by introducing active measures to facilitate producing land or grassland transformation into forest land, the first tested option (producing land to forest scenario) most improves the GHG balance from the LULUCF sector.
Assuming that there are limited possibilities to further increase the areas of forest land or reduce built-up areas and wetlands, the key factor to improve the GHG balance in the LULUCF sector will be the proportion of producing land and grassland. If land use management as it was in the period between 2005 and 2010 continues without additional measures to support specific land use transformation types (Reference 2005–2010 scenario), the GHG emissions from agricultural land could increase over the next decade from 2020 by ~9.5%. However, if continuing land use management policies as they were after 2010 (Reference 2010–2015 scenario), GHG emissions could decrease by 20–35% compared with the Reference 2005–2010 scenario, and by 2030 the emissions from agricultural land could be reduced by 11%. Therefore, we can assume that different suggested development trends are affected by changes in Lithuanian land management. Historically, several periods have shaped Lithuanian land management in the last three decades. The largest impact on the use intensity of agricultural lands could be associated with the agrarian reform in the country, which started in 1991. This reform resulted in fully changed formats of agriculture, land management, and land use relationships, and production capacities of agricultural subjects. The second group of impacts is associated with Lithuania joining the European Union in 2004 and the availability of EU and national budget resources to support agriculture and rural development. The factors influencing land use development are usually interdependent, and the outcomes of their inter-relationships during specific periods of socioeconomic development are shaped mainly by political and social factors, with natural conditions playing only a secondary role. Bearing in mind the time periods used to build the land use change models in our study, 2005–2010 could have been influenced by agriculture restructuring (which started in 1991–1992), and the next period is likely associated with the impacts of joining the EU and the state support for agriculture and rural development.
Additional measures to support specific land use transformation types were introduced in the models by manual adjustment of the Markov matrices. Three types of such measures are discussed. First, forcing the transformation of producing land into forest was associated primarily with the strategic forest policy objective to increase the proportion of forest land area, coupled with current land use management efforts to sustain grassland areas or at least to prevent their transformation back to producing land. The second type of measure (grassland to forest) was aimed to increase the forest land area on current grassland. It matches the first scenario, however, with no conditions regarding the efforts to prevent transformations from producing land into grassland. The third type of measure (no grassland to producing land) was associated with additional efforts to prevent grassland transformation into producing land only, i.e., leaving out the extra efforts to increase forest land area. Therefore, if a land use management policy generally follows that in effect from 2005 to 2010, additional measures to support specific land use transformation types will not result in reducing GHG emissions, either from agricultural land or the entire LULUCF sector in the decade starting at 2020. Conversely, GHG emissions from agricultural land are predicted to be reduced in the coming decade if the land use management policy used from 2010 to 2015 is followed. Introducing extra measures would support the reduction of GHG emissions from agricultural land. Especially important in this context is the reduction of producing land by its transformation into forest land (producing land to forest) or preventing the transformation of grassland into producing land (no grassland to producing land). The introduction of such measures may reduce GHG emissions in the next decade by ~16 and 28%, respectively.
Summarizing, in order to improve the GHG balance in the LULUCF sector in Lithuania over the next decade starting at 2020, the focus in Lithuania should be to increase forest and grassland areas. This objective is supported by national strategic political documents, especially those aimed at the effective use of EU support [
54,
55,
56]. The key contributor to the total CO
2 balance in the LULUCF sector will remain the total forest land area and the potential to increase it in the future. Thus, the EU contribution should be targeted to support the establishment of new forests, assuming that backward processes remain under strict legal restraint. The common agricultural policy (CAP) of the EU should further focus on green direct payment, especially maintaining permanent grassland, which not only supports carbon sequestration but also contributes to the protection of biodiversity (Regulation (EU) No. 1307/2013). In parallel, Lithuania should continue to maintain its permanent grassland [
55].
5. Conclusions
The prediction accuracy of land use types directly related to GHG accounting and emission/absorption management in the LULUCF sector in Lithuania was above 80% over a five-year period into the future. Land use types whose abundance changed relatively linearly during the last three decades—forest and built-up lands—were predicted with accuracies of 96% and above. The most challenging was the prediction of land use types on agricultural land, i.e., the separation between producing land and grassland. These results were obtained using a compatible methodological approach based on a Markov chain-type model as used by Lithuanian authorities to estimate forest reference levels for the country following the European Union land use, land use change, and forestry regulation for 2021–2030. It should be emphasized that driver variables affecting land use transformation over time were estimated from information freely available from GIS databases, as the modeling exercise was implemented using standard GIS software.
All scenarios tested suggested that the LULUCF sector in Lithuania would accumulate carbon during the next decade, starting in 2020. The main land use type contributing to the most carbon absorption will remain forest land. Even though the proportion of forest land area in Lithuania is predicted to increase, we did not manage to simulate forest land proportions exceeding 37% of the country’s area by either applying land use management approaches as they were applied since 2005 or by introducing additional measures to support forest land expansion. The key factors to improve the GHG balance from the LULUCF sector in the near future, assuming a stable development of forest land and strict deforestation control, are keeping the proportion of producing land and grassland and afforestation of abandoned and uncultivated agricultural lands.
To facilitate CO2 emission/absorption management in the LULUCF sector together with increasing socioeconomic and environmental benefits of Lithuanian rural landscapes, more sophisticated tools to support the monitoring, analysis, and modeling of land-related mitigation activities are needed. Lithuania has developed an original land use monitoring system that is used for GHG reporting, which, up to some level, predetermines land use development projections. However, even though the system is sufficient to fulfil the country’s international climate change mitigation commitments, it encompasses a number of limitations in both substantiating the methodology and the way it is operationally implemented. Further research is needed to improve the methodological framework for integrated land management, which can make use of the digital technologies for inventory and decision support to serve the needs of managers and policy makers with a specific focus on GHG management. More specifically, wall-to-wall mapped land use and land use changes would provide better inputs for land use development scenario modeling using the methodological approach tested in this study. The development of spatially explicit land use change scenario modeling and analysis tools could focus on the use of cellular automata and agent-based modeling approaches.