**3. Results**

### *3.1. Calibration and Validation of Land Use Change Models*

First, we predicted all land uses in 2015 for all sample points using the input data for 2005–2010 and assuming the Reference scenario. The overall accuracy of prediction was in the range 82–83% (Table 3). The kappa statistic was 0.76–0.77. It seems that the factor inclusion strategy in the calculation of transformation potential had no significant effect on prediction accuracy; the kappa statistics did not differ with statistical significance, with the highest value of the *Z* statistic being 0.148 (not presented in Table 3). The prediction accuracy statistics of the most encountered land use classes are summarized in Figure 4. The most accurately predicted land cover class is forest land—both the producer's and user's accuracies yielding nearly 99%. The development of built-up areas is also accurately predicted; the F-score is 96%. It is noteworthy that practically in all cases the producer's accuracy (~94.5%) is lower than the user's accuracy (~97.5%), suggesting that other land use classes are more often incorrectly predicted to be transformed into built-up land, rather than vice versa.

The accuracy of predicting the producing land was notably better than that of cultural grassland/pastures, natural grassland, or natural grassland with trees and brush. On average, producing land was predicted with 84–87% accuracies, and the producer's accuracy was higher than the user's accuracy. Cultural grasslands/pastures, natural grasslands, and natural grassland with trees and brush resulted in the lowest prediction accuracies (if considering the most abundant land uses). Only the prediction accuracy for cultural grasslands/pastures reached 50%, and the producer's and user's accuracies did not differ. More in-depth analysis of error matrices confirmed that the abovementioned land uses were mixed with each other during the prediction. Therefore, cultural grasslands/pastures, natural grassland, and natural grassland with trees and brush are combined into one class—grassland. Following this combination, the overall classification accuracy increased by 7–8%, but the increase in kappa was not statistically significant (Table 3). After the merge, grasslands were predicted with 73–80% accuracy, and the producer's accuracy was lower than the user's accuracy. Land with brush was predicted with ~60% accuracy, but the area of this type was relatively small.

The modeling exercise was repeated using the assumptions of Scenario 3: no grassland to producing land (2005–2010). Although the overall prediction accuracy improved by 1–2%, this improvement is not statistically significant. Different scenario conditions had an impact in predicting the grassland development when using detailed grassland subtypes. After combining the grassland subtypes, we achieved very similar producer's and user's accuracies, i.e., differing by no more than 1%.

### *3.2. Land Use Changes in the Future*

Predicted proportions of three major land use types—forest land, producing land, and grassland—are presented in Figure 5. The proportion of forest land is expected to increase regardless of the scenario. It should be noted that scenarios involving active efforts to increase the area of forest land result in larger forest land areas, although never exceeding 37%. Using the land use trends from 2010–2015 to model the transition potential resulted in larger forest land proportions. The area of producing land is expected to increase only if using 2005–2010 land use data to model the transition potential. However, if extrapolating the trends from 2010–2015, the areas of producing land decrease. Manual adjustment of Markov matrices, aimed to specify additional land use policy measures, resulted in even fewer areas of producing land, if compared with the Reference scenarios. If the land use changes during 2010–2015 continue into the future, the proportion of producing land in Lithuania will be reduced to below 30%. The area proportion of grassland is increased if considering the trends during 2010–2015 and, vice versa, decreased if using the 2005–2010 period to model the transition potential. The exception was the scenario with no grassland for producing land, where the grassland decrease stopped by adjusting the Markov matrix. If the land use change trends during 2010–2015 continue in the near future, the proportion of grassland will be projected to increase to 23–28%, depending on the scenario. The lowest grassland proportions were achieved in the scenario of grassland to forest (2005–2010), i.e., following the fast decreasing grassland areas from the half decade, since Lithuania joined the EU and introduced measures for grassland conversion into forest land. It should be noted that the projected trends of producing land development are inversely followed by the trends of grassland proportion.


**Table 3.** Prediction accuracy of all tested land use types.

\* all classes vs. grassland in the one-class Reference scenario, \*\* Reference scenario vs. scenario with no grassland for producing land (2005–2010) (in the numerator—all land use subtypes; in the denominator—grasslands merged into one class).

**Figure 4.** Predicting accuracy of some of the most encountered land uses, achieved using a strategy of driver variable selection based on optimization.

**Figure 5.** Projected development of selected land uses in Lithuania, depending on land use change scenarios: (**a**) forest land, (**b**) producing land, and (**c**) grassland.

None of the tested scenarios suggested carbon emissions from the LULUCF sector in Lithuania before 2030 (Figure 6). A larger absorption (up to 33%) was projected when considering land use changes that took place from 2010 to 2015 in modeling the transition potential. The largest overall absorption (above 1 ton of CO2 equivalent from 1 ha) was achieved in the scenario where producing land became forest (2010–2015), i.e., aiming to maximize producing land conversion into forest land.

**Figure 6.** Predicted carbon emission and absorption from the land use, land use change and forestry (LULUCF) sector in Lithuania, depending on scenario: (**a**) Reference (2005–2010), (**b**) Reference (2010–2015), (**c**) producing land to forest (2005–2010), (**d**) producing land to forest (2010–2015), (**e**) grassland to forest (2005–2010), (**f**) grassland to forest (2010–2015), (**g**) no grassland to producing land (2005–2010), and (**h**) no grassland to producing land (2010–2015). The value shown below each bar indicates the total carbon sequestration value. Numeric values can be found in Appendix A, Table A3.
