Cork oak forests cover 350,000 ha in Morocco. This is about 15% of world’s cork oak area. However, it contributes only for 4% to 6% in the world’s cork production [1] . Many factors are involved for explaining this situation (social pressure, intensive harvesting, climate) and regressive trends of these forests have been described [2] .

Maâmora as example is considered one of the largest cork oak forests in the world; it covers 133,000 ha. Unfortunately, cork oak covers only 70,383 ha. Since 1940, Maâmora has been managed according to four management plans. These plans revealed a cover regression trend which has been estimated at 30% of the cork oak area [2] . Therefore, since 1950, overgrazing and acorns picking has been considered a major obstacles to cork oak regeneration [3] , since then regeneration problems still prevail and overgrazing rate is currently estimated at 81.57% [2] .

To deal with this alarming situation of non-regeneration, forest managers have undertaken huge efforts for studying and regenerating cork oak. Results have been qualified as random [4] . Consequently, administration maintained and continued rehabilitation of cork oak. Thus, 20,000 ha have been planted under the scope of 2005-2014 program.

Cork oak regeneration studies in Maâmora pointed out that water balance (climate and soil factors) associated with the high anthropogenic pressure are the main factors which determine the success of planted areas. Moreover, previous experiments showed that initial density, soil type and waterlogging have a strong influence on the ability of the strains to reject. Further, Maâmora’s soils consist on a sand layer resting above a clay layer. Thus, sand depth and slope of underground clay floor seems to be crucial for cork oak seedlings success [4] [5] . Besides, rains follow a decreasing gradient from west to east by the distance to ocean and the location of the stands.

Despite that soil factors have been well documented in the Maâmora forest, no soils clay layer depth/slope or a soil suitability to cork oak regeneration maps are available for all the Maâmora. This could explain same cases of regeneration failure.

Habitat suitability models aim predicting lands which fit well for a species according to specified requirements, preferences, or predictors [6] . As suitability analyzes the interactions among three types of factors: location, development actions, and environmental elements [6] , they constitute a good tools for supporting decision making which improve the understanding of species habitat relationship in space and time and predict their likelihood of occurrence and abundance or performance based on habitat attributes affecting survival, growth and reproduction of considered species [7] [8] .

Habitat suitability models can be classified according to the approach and to the classification function used [9] . Nowadays, predictive techniques have been improved with a direct consequence on models quality and credibility [10] . Further, machine learning (ML) methods become promising tools for assessing suitability.

ML is nonparametric techniques for regression or classification which present some advantages compared to other statistical methods. They are able to: 1) deal with complex relationships between predictors; 2) process nonlinear relationships between predictors; 3) deal with large amounts of data; 4) process complex and noisy data and 5) don’t need any assumptions related to variable distribution [11] .

Among machine learning methods, many methods have been used for suitability assessment [12] such as classification and regression trees [13] , artificial neural networks [14] , heuristic methods such as genetic and simulated annealing algorithms [15] [16] , and Random Forests [17] . Several studies showed that Random Forest (RF) models, often reach top predictive performances compared to other methodologies [10] [18] [19] .

As summarized by Cutler et al. (2007), RF applications include: 1) regression; 2) survival analysis; 3) imputation of missing values; 4) clustering, multidimensional scaling and detecting general multivariate structure through the unsupervised learning; and 5) classification.

In this work, we used RF as a classification method and to explore relationship between observed cork oak plantation’s success rate and some environmental factors judged important. Our paper is organized as follows: 1) part 1: materials and methods which consists on a brief description of the study area, factors considered and modelling framework including a general description of the modeling procedures, the random Forest model and predicted maps; 2) part 2: results will illustrate the application of RF to Maâmora forest to obtain estimation of potential plantation success rate and discuss the results.

The out-of-bag estimates of the error rate (ERROOB) were used to select the optimum Random Forest parameters. For the calibration dataset (CD), the Random Forest was able to classify the samples with an out-of-bag estimate of error rate less than 10% (OOB error = 8.85%).

Figure 4 shows the ranking of predictors by their importance measured as the increased mean square error (%IncNodePurity) representing the deterioration of the predictive ability of the model when each predictor is replaced in turn by a random noise. Higher %IncNodePurity indicates greater variable importance. Only four of the predictors contributed noticeably to the estimation of suitability grade, namely precipitation (Prec), elevation (Alt), temperature (m & M) and Emberger’s quotient (Q2).

Moreover, Figure 5 shows partial plots which represent the marginal effect of a single factor included in the RF model on estimates of cork oak plantation suitability. In these partial plots of marginal effects, only the range of values can be compared between plots of different variables.

The RF model provided a good prediction of success rate grade for regeneration in the validation dataset. Per class OOB errors ranges from 7.3% for the fourth classes to 16% for the third class. Further, using a per-pixel approach for classification quality assessment (observed success rates from previous plantation vs. predicted ones) through a kappa index gives a value of 0.827 which mean an almost perfect agreement [28] .

The RF model application to the whole Maâmora is shown in Figure 6. The most suitable area (4th and 5th classes: with success rates greater than 60%) are located on the western part of the forest. While the eastern part represent the lowest predictions scores of suitability. This can be explained by continentally and the rainfall gradient.

The RF algorithm used does not require any assumptions related to normality of the variables and can deal with nonlinear relationships. These features may have a substantial utility within the context of forest management and could be used to model distribution shifts resulting from climate change. Moreover it constitutes a

promoting approach [7] [10] . Results obtained with the random forest method for predicting suitability maps are very encouraging; they can predict the success rate grades of regeneration. Further, compared to other approaches: LepoutreAbac [5] or suitability using AHP [22] , RF presents a highest predictive accuracy.

The RF variable importance measure indicated precipitation and altitude as important factors in the modelling (Figure 5). Even if precipitation is expected because of the Mediterranean climate of the Maâmora Forest, it was not expected that the altitude is also important as other factors.

When comparing the observed success rates from previous plantations with the predicted ones by RF model, the kappa statistic used indicate a good accuracy of the RF model.

As stated by Benito Garzón et al. (2006), biological validation and interpretation of results obtained by modelling process are very important. Thus, our results are encouraging because they agree with the main bibliographic conclusions [4] [5] [22] .

As compared to the larger previous cork oak distribution in the Maâmora Forest, what can be observed now is the result of historical and social dynamic on which area have been reduced mainly by intense anthropic activity (overgrazing, shifting species, etc.). Furthermore, Plantation success rates are not only affected by climatic, topographic, soil and social factors but with other factors. Moreover, seedling production and installation (seeds origin/quality, seedling production process, soil preparation and improvement, etc…), anthropogenic pressure, overgrazing and global warming issues, among others, have probably determined the current observed success rates. Models results should therefore be interpreted with full knowledge of these inevitable limitations. Assuming these limitations, the modelling framework may offer a good tool for ensuring sustainability of Maâmora forest, hence giving to the managers possibilities for planning regeneration and rehabilitation activities.

Succeeding Quercus suber plantation is a challenging operation for Maâmora forest managers. Indeed, suitability map remains a valuable tool which is very helpful to forest managers for planning forest regeneration and ensuring sustainability. The main purpose of this article is to produce this map.

The use of random forest algorithm made it possible to achieve the overall objective of this work. The produced map enables managers to choose the most suitable area to be planted and to guarantee the success of their regenerations efforts. Furthermore, it ensures suitability of the forest. The resulting map helps managers to make more rational decisions related to cork oak regeneration and to ensure the sustainability of Maâmora forest.

The research for this paper was financially supported by the wallonie bruxelle international, within the scope of the project “Comportement des plantations du chêne liège de la forêt de Maâmora, Maroc”.