Maamora forest is located in North-eastern part of Morocco (Figure 1). Its 5 blocs subdivided in 33 groups and 460 parcels range over 68 km (West-East) and 38 km (North-South) [9] .

The climate in Maamora forest is Mediterranean, with an influence from the Atlantic Ocean. The bio-climate ranges from semi-arid with temperate winters in the eastern part of the forest to sub-humid with warm winters in its western part [5] . The average seasonal rainfall regime is W.A.S.S (Winter-Autumn- Spring-Summer). The topography of Maamora forest is characterized by a not so rugged relief, with slopes ranging between 0.6% and 0.8% in average, except the eastern part where slope is steep enough to provoke significant erosion [10] . The soils characterizing the forest are generally sand lying on clayey layer. The type of soil is important in water retention. In fact, with thick sand layers, water can become a limiting factor due to its unavailability for use by plants [11] .

The vegetation of Maamora is mainly comprised of cork oak, even though it has decreased in area lately, giving space for eucalyptus, pine and acacia plantations. Throughout its history, and with the products and services that it provides, Maamora forest is one of the most important forests in Morocco. It sig-

nificantly accounts for the needs in fuelwood, grazing for livestock, recreation, etc. for more than 10 communes [5] . Human activity is quite notable and mainly includes acorn collection, grazing and woodcutting for energy [4] .

In this section, data in the form of Landsat images was obtained at different dates, from the http://glovis.usgs.gov/ website. Available forest maps were also obtained from forest administration. Satellite images for the period between June and September, downloaded as individual layers, were stacked and sharpened to produce multispectral images. That period of the year (summer) was chosen in order to prevent the confusion that could be caused by shrubs during classification. All the available images for the summer period were stacked together to produce a composite image for each year. This was essential in order to reduce radiometric noise that would be notable if a single image were used per period.

The Figure 2, below, summarizes the methodology used for studying forest dynamics.

The classification of images for the three years considered (1987, 2000 and 2014) for studying the dynamics using the SVM (Support Vector Machine) algorithm made it possible to create a map showing the distribution of the main tree species of the forest. After classification and before obtaining the final map, sieving was done, considering a minimum mappable area [12] [13] of 2 ha which represented 20 pixels. A control for accuracy was performed using the Kappa index to validate the classification [14] . Ground truth was collected for the year 2014. However, for the years 1987 and 2000, we performed a visual interpreta-

tion, assisted by available maps, of at least 60 points per density-species class which we compared to the result of the classification in order to obtain the Kappa index. Afterwards, change detection was done by comparing the land cover maps. Finally, some drivers of change were identified.

The assessment of vulnerability was done using spatial multi-criteria analysis [15] . Eleven individual factors grouped in four groups of factors (biophysical, climatic, anthropic and silvicultural) were considered as result of forest dynamics study:

・ Biophysical factors: Natural terrain slope, slope of the clay layer, type of soil and depth of sand;

・ Climatic factors: Water stress and isohyets (proxy for continentality);

・ Anthropic factors: need in fuelwood, grazing pressure and population density;

・ Silvicultural factors: Age of stand and health state.

Data was collected from different sources (previous studies, Direction of national meteorology and forest administration). Water stress (WS) was determined from raw data and computed using the following formula:

W S = ( P + A W C ) – E T P (1)

where:

P is Precipitation (in mm),

AWC is the Available water capacity [16] ,

ETP is the Evapotranspiration [17] .

WS was computed for each month and then averaged for months with a water stress, i.e. when WS < 0. Each individual factor was spatialized, categorized (thresholds), averaged by group (unit of the forest) and combined at 2 levels (factor and final vulnerabilities) using the formula of Equation (2) in order to obtain the final vulnerability at the reference year, 2010 (year chosen for availability of all data needed).

V = ∑ 1 i W i * F i (2)

where:

V is vulnerability,

W_i are the weights,

F_i represents each individual factor or group of factors.

The method is summarized in Figure 3. The weights used to linearly combine the different factors were determined by Analytic Hierarchy Process (AHP) developed by Saaty [18] . The intensity of importance required for computing the weights are given by experts.

In order to forecast vulnerability at years 2045 and 2070, we projected factors considered changing over the period of study. Biophysical factors for instance were not projected given the fact that they wouldn’t significantly change over the

considered period. The climatic factor was derived by projecting temperatures and precipitations to the 2 different years considering 2 different scenarios, one we called optimistic (scenario RCP 4.5, RCP being Representative Concentration Pathway) and the other pessimistic (scenario RCP 8.5). Those scenarios were conceived based on possible trajectories of greenhouse gases [19] and an interface available on internet (http://climexp.knmi.nl/plot_atlas_form.py) was used to project climatic data. Anthropic factors were projected taking into account average growth of rural population in Morocco [20] . Stand age was recalculated supposing no regeneration would occur during that time.

The classification of the 3 images was followed by a control. Each time, the Kappa index (written on each map) was found to be higher than 80%, thus allowing us to validate the classification. The analysis of the obtained maps (Figure 4) showed that cork oak was currently mainly present in Blocs A, B and C, with high density only in bloc A. In the other blocs (D and E), it had almost disappeared, giving space to mainly eucalyptus and pine. Of all the forest species, acacia was the least represented.

To better interpret the maps of Figure 4, a synthesis of area was necessary in order to obtain a cover percentage per species, with all density classes put together. Figure 5 presents the trends for each species. Percentages are relative to the total area of the forest calculated excluding all enclaves, i.e. 131,020 ha.

The graph shows that:

・ The area covered by cork oak (49.2% or 64,461.38 ha in 1987) declined by

11.7% during the first period, between 1987 and 2000 (43.46% of total area or 56,937.4 ha in 2000), before increasing by 7.96% during the second period, between 2000 and 2014 (46.92% of total area or 61,471.65 ha in 2014);

・ The area covered by eucalyptus (34.13% or 44,719.49 ha in 1987) increased by 7.5% during the first period and then decreased by 11.31% during the second period (32.54% or 42,635.13 ha in 2014);

・ The area under pine plantations (4.4% or 5769.73 ha in 1987) increased significantly by 73.54% over all the studied period (7.64% or 10,012.84 ha in 2014);

・ The area under acacia (2.94% or 3849.94 ha in 1987) decreased by 33.67% during the first period before increasing by 41% during the second period (2.75% or 3600.9 ha in 2014).

That said, it is important to point out that the decrease in area of cork oak resulted in available space for the other species, including eucalyptus and pine. The increase in cover between 2000 and 2014 was most likely a result of successful cork oak regeneration by seed [21] . This was obtained after the decision of forest administration, in 2003, to restore the Maamora forest with its natural species which is cork oak.

These results are partly in tandem with those obtained in a study, done few years ago by the forest administration, as presented in Figure 6. According to that study, cork oak area decreased from 100,000 ha to 60,000 ha between 1951 and 1992 followed by an increase of up to 65,000 ha in 2006. In our study, we obtained an area of almost 57,000 ha in 2000. Hence, it is possible (if we place this value in the trend shown in Figure 6) that cork oak cover continued to decline until 2000, before increasing again.

However, the trend in cork oak area between 2000 and 2014, in our study does not match well with the study above mentioned. In fact, the evaluated cork oak area was 70,400 ha in 2011, whereas the classification in this work resulted in an area of about 61,500 ha in 2014. This difference could be partly explained by the omission that might be a result of very young plantations of cork oak not having sufficient crown growth. This is essential for reflecting light as trees, and helps in the identification of cork oak. Consequently, plantations for the last years might not be well accounted during the classification of the 2014 image.

The changes that occurred during the studied period are presented in the graph of Figure 7.

This chart illustrates the importance of the dynamics of forest ecosystems in Maamora forest. In fact, for both periods, only about 40% of the total area (or about 60% if we do not consider shifts to other density classes as changes) remained unchanged. Despite all the efforts for restoring cork oak ecosystems, a slight increase in the area shifting to lower density classes was still noticeable.

However, increased conversion and minor afforestation activities could, for the most part, be accounted for cork oak regeneration.

This dynamics was mainly a result of management actions, although other anthropic activities such as eucalyptus woodcutting for making charcoal, fuelwood, grazing, etc. could have contributed.

Below is a schema summarizing the most important drivers of change in the Maamora forest.

From the analysis that led to the elaboration of schema in Figure 8, 4 groups of important drivers of change in Maamora forest were identified, as follows:

・ Climatic conditions: aridity was more pronounced as we moved from West to East, with the influence of Atlantic Ocean being reduced. This factor would be more important with current trends in climate [8] .

・ Biophysical conditions: in Maamora, the most important biophysical factor could be considered to be the type of soil (sand lying on clay layer) as the slope of the clayey layer and the thickness of the sand layer could considerably influence water retention in the soil.

・ Health state: parasite attacks combined with aging of the forest weaken the vegetation and lead to its decline.

・ Anthropic factors: population increase coupled with increased livestock presence resulted in increased acorn picking, illegal logging and overgrazing.

These 4 groups of factors were used to assess the vulnerability of forest ecosystems to global changes.

The weights used to combine individual factors to obtain factor vulnerabilities are given in the table below:

The linear weighted combination of individual factors maps with the weights of the Table 1 produced the maps in Figure 9. The biophysical factor vulnerability map shows two groups (BG2 and BG7) as non-vulnerable, with no group of the forest being of high vulnerability considering only this factor. This implies that biophysical factors, even though not very favorable, are not the most limiting in Maamora. The climatic factor vulnerability map underlines the effect of continentality on vulnerability, thus highlighting that most continental ecosystems are the most vulnerable. As for the anthropic factor, unlike the biophysical factor, we do not have any non-vulnerable group, thus confirming the importance of human activity in the forest. The silvicultural factor vulnerability, on the other hand, is on average fairly low. This implies that even though Maamora forest might be old, especially in reference to cork oak, its age might not be the most important factor with regards to vulnerability of forest ecosystems to global changes.

After combining all the previous factor vulnerability maps using weights given in Table 2, the final vulnerability map at the reference year (Figure 10) was obtained.

The final vulnerability map did not highlight any non-vulnerable group, confirming the fact that all the ecosystems were, to a certain extent, vulnerable to global changes. It is important to note that only the DGI group was highly vulnerable in 2010, whereas all the other groups were lowly to moderately vulnerable.

After projection of changing factors (climatic, anthropic and silvicultural) final vulnerabilities obtained for each scenario and depending on the considered years (2045 and 2070) are presented in Figure 11.

Final vulnerability predicted for 2045 using the optimistic scenario (for pro-

jection of climatic factors) showed that 3 groups of bloc A and 1 group of bloc B would be of low vulnerability whereas the remaining groups of blocs A and B, as well as the groups of bloc C would be moderately vulnerable. The groups of blocs D and E would be of high vulnerability. When considering the pessimistic scenario, all the groups of blocs C, D and E would be highly vulnerable whereas those of blocs A and B moderately vulnerable.

In 2070, with the optimistic scenario, the obtained vulnerability map of forest ecosystems was similar to that of 2045 with the pessimistic scenario. The groups of blocs A and B would be moderately vulnerable, whereas those of blocs C, D and E highly vulnerable. Additionally, when considering the pessimistic scenario, some groups of bloc B would shift to high vulnerability.

・ Statistics of vulnerability per category at reference year 2010 and for the horizons 2045 and 2070 considering the 2 climate scenarios (RCP 4.5 and RCP 8.5).

The graph shown in Figure 12 presents a synthesis of percentage of total forest area per vulnerability class for the reference year (2010). It also illustrates the scenarios for horizons 2045 and 2070, with reference to the 2 climatic scenarios.

The graph shows that the vulnerability of forest ecosystems would increase with time as we go from 2010 to the horizons 2045 and 2070, with this being more or less significant depending on the considered scenario. In fact, forest ecosystems seem more vulnerable to global changes with the pessimistic scenario (RCP 8.5).

One important thing to note is the fact that no group presented an insignificant vulnerability, be it at reference year or at horizons 2045 and 2070. In 2010, 53% of the entire forest area had a low vulnerability whereas only 11% in 2045 would remain with low vulnerability with scenario RCP 4.5. The rest of the forest would shift to moderate and high vulnerabilities. At horizon 2045, with scenario RCP 8.5, and at horizon 2070, regardless of the considered scenario, all the groups of the forest would be moderately to highly vulnerable.

It should be noted that a certain level of subjectivity in the thresholding (low, moderate, high vulnerability) of the individual factors remains. This could be a source of significant variability in the resulting final vulnerability, from author to author. However, with the thresholds being defined rigorously based on our best knowledge of Maamora forest, the obtained results could be considered a reflection of the vulnerability of Maamora forest ecosystems to global changes.