Predicting vegetation dynamics in Deux Balé National Park, Burkina Faso, using land change modeler

In recent years, the natural vegetation of Burkina Faso has undergone unprecedented anthropisation. To date, most studies have focused only on diachronic analysis based on satellite imagery. Land use dynamics studies give an account of the past and present situation, but do not provide quantitative and qualitative information on the future of this vegetation. On the other hand, a prospective analysis of the dynamics would enable us to better assess the future of our forest ecosystems. Predictive data are essential for decision-making and implementing appropriate policy responses relating to sustainable forest resources management. The main objective of this study was to predict the vegetation dynamics of the Deux Balé National Park by modelling the year 2025. Thus, we selected the variables that best explained the anthropization process. A land change model was selected for the study. The model was calibrated using only variables with a V-Cramer coefficient greater than or equal to 0.1. This model is based on error budgeting and prediction. Visual and statistical comparisons of the simulated and actual 2016 maps allowed for better calibration of the model to simulate changes in the area of land-cover classes by 2025. The results revealed that, between 2016 and 2025, a significant regression of savanna vegetation will occur in favor of anthropized areas and gallery forests. Thus, from this period, gallery forests and anthropized areas covered up to 138 ha and 2914 ha, respectively. However, the savanna vegetation area decreased by 1186 ha, 961 ha, and 893 ha from 2016 to 2025, respectively. Therefore, urgent action must be taken to preserve the potential of the forest resources in parks.


Introduction
In most sub-Saharan countries, the main concern of environmental policies is to reconcile forest conservation and agricultural production in a context in where the demand for land is high and increasing (Bonnefoy et al 2001, Tankoano et al 2015).Since the colonial period (Kiema 2007), protected areas have been created with the main intention of preserving forests for their fundamental functions, including climate barriers, biodiversity conservation, carbon sequestration, and food provision.However, several authors have highlighted the severe degradation of forest cover and deforestation in Burkina Faso (DIFOR 2007, Bombiri 2008, FAO 2011, Tankoano et al 2015).It appears that this unprecedented demand for agricultural land causes severe problems in protected areas, which are biodiversity reservoirs in Burkina Faso (Ouoba 2006, Tankoano et al 2016).As a result, deforestation and the degradation of forests in these areas affect the viability of ecosystems and limit their ability to produce goods and services.This poses a threat to riverside populations and wildlife that depend on them (Tabopda and Fosting 2010, Andriamasimanana et al 2013, Kambiré et al 2015).Faced with this situation, it is therefore imperative to understand these phenomena to anticipate and limit their harmful impacts in time and space (Andriamasimanana et al 2013).Most studies have focused on forest regeneration dynamics in Burkina Faso and revealed a regressive trend in forest vegetation at both national and local scales (Zoungrana et al 2015, Tankoano et al 2016,Sanon et al 2019).Several authors have used satellite images to monitor savanna vegetation in Africa (Houessou et al 2013, Dimobe et al 2015, Ouédraogo et al 2010, Gansaonré 2018).In Burkina Faso, specifically in the Parc National des Deux Balé and its outskirts, the regression of vegetation cover was estimated at around 0.2% per year between 1985 and 2015 (Tankoano et al 2016).In the east of the country, Soulama et al (2015) showed that 30.8% of the vegetation cover had been degraded between 2001 and 2013.However, accurate data regarding deforestation and the degradation of protected areas are still lacking in Burkina Faso, especially in the Deux Balé National Park (PNDB).Nevertheless, these data were found to be crucial for planning forest management and rational utilization of plant resources (Mbow 2009).Changes in the landscape were clearly observed on the periphery of the PNDB, which were the result of agricultural expansion around the park (Tankoano et al 2016).Considering the high land pressure and adverse effects of climate change, it is necessary to identify and understand the drivers of deforestation (Kambiré et al 2015) to anticipate future landscape changes (Lambin et al 1999, Balzter 2000, Bonnefoy et al 2001, Mas et al 2011, Maestripieri and Paegelow 2013, Mahmoud et al 2016).Understanding changes in land use/cover could be the basis for sustainable preservation of protected areas.Therefore, modelling vegetation dynamics is required to design appropriate strategies to combat deforestation and promote sustainable land management (Oloukoi et al 2007, Oloukoi 2013, Abuelaish and Camacho-Olmedo 2016, N'Gamba et al 2016).Thus, an available model of vegetation dynamics in the PNDB would make it possible to predict the future status of various land use/land cover classes.This will aid the protection and management of protected areas.This study used the Land Change Modeler (LCM) implemented in Idrisi 17.0 Selva to analyze the dynamics and predict the future land cover map of the PNDB.This model has been widely used in several studies (Akadjé 2016, Mahmoud et al 2016, N 'Gamba et al 2016).According to N 'Gamba et al (2016), spatial modelling of the dynamics of the Yangambi Biosphere Reserve up to 2065 using the LCM and CA-Markov models revealed a significant reduction in forest area of almost 43% in favor of anthropogenic areas, which will increase by approximately 137.27%.A projection into the future of the state of forest ecosystems is very useful in the planning and sustainable management of protected areas.Thus, knowledge of the trends in forest ecosystem dynamics is necessary, and even fundamental, to properly forecast and regulate them.
The main objective of this study was to provide park managers with plausible information on the future status of the park's forest ecosystems to improve management decisions.Specifically, its aim is to predict the status of forest ecosystems in the PNDB by 2025 using the Land Change Modeler (LCM).

Study area
The PNDB was created in 1988 by Zatu AN VII/FP/PRES/MET, following the merging of two protected forests (Deux Balé and Dibon), and covers 80 600 ha (Kafando 2003).The park is located between the provinces of Balé and Tuy (11°25'−11°36' N, 2°45' and 3°12' W, figure 1).The PNDB extends on a peneplain with altitudes ranging between 240 m and 320 m (Coziadom 2009).The Park lies in the climate zone between the 750 mm isohyets in the north and the 1,000 mm isohyets in the south.This Sudanian-type climate is characterised by two seasons : a rainy season from May to October and a dry season from November to April.The average annual temperature is approximately 28 °C, with a range of approximately 7 °C (Coziadom 2009).

Material
The following material were used in this study • Raster format land use maps of the PNDB in 1986, 2010, and 2016(Tankoano et al 2016) • vector layers such as road networks, hydrographic networks, PNDB boundaries, and village and monitoring station locations (Geographical Institute of Burkina Faso4 ).These data, considered as variables of the LCM, allowed the creation of a series of distance maps for the road network, hydrographic network, villages, monitoring stations, and park boundaries.
• and Idrisi 17.0 Selva software used for the prospective modelling of vegetation dynamics.

Methods
The modelling process of land cover using the LCM was performed in three steps: calibration, simulation and validation.

Model calibration
The model calibration involved the following steps described below: The database (figure 2) included PNDB land use maps for 1986 and 2010, distance maps for the road network, hydrographic network, villages and monitoring stations, park boundaries (figure 3), and soil maps (figure 4).Before loading them into the LCM, the distance maps were resampled to the resolution of the land cover maps.The nearest neighbour resampling method was used.We chose this resampling method because it can be applied to both discrete and continuous values, whereas other types of resampling apply only to continuous data.Thus, the number of explanatory variables to be included in the simulation model of observed changes in land use is constrained by their availability, spatialization, influence on location, and changes in land use patterns.However, it should be noted that the number of factors presented and integrated was limited compared with the range of potentially explanatory variables.The legitimacy of these data in calibrating the model must be discussed (Maestripieri and Paegelow 2013).Indeed, all variables predate 2010, which is an asset for calibrating    2.2.1.2.Analysis of changes and transition potentials modelling Based on the 1986 and 2010 land use maps, the first estimate of changes (losses, gains, net changes, and stability) was realized by creating graphs and maps of changes and transitions.All transitions between the two land use maps (1986 and 2010) were automatically generated in a table after change analysis.In this table of transitions between land use classes, it is possible to select the desired transitions to build a sub-model of the model.In total, 9 transitions were selected on the basis of their probability of occurring between the two dates.These transitions include : (i) from Gallery forests to Anthropogenic areas, (ii) from wooded savannahs to Anthropogenic areas, (iii) from Dense shrub savannahs to Anthropogenic areas (iv) from Clear shrub savannahs to Anthropogenic areas (v) from Bare soil to Anthropogenic areas, (vi) from Anthropogenic areas to Wooded savannahs; (vii) from Anthropogenic areas to Dense shrub savannahs; (viii) from Anthropogenic areas to Clear shrub savannahs (ix) from Wooded savannahs to Clear shrub savannahs.After the analysis changes step, we proceeded to the transition potential modelling, where we identify the potential of land use classes to transition.At this stage, we created transition potential maps that are in essence maps of suitability for each transition.According to Megahed et al (2015), we can said that in this study, the transition potential maps represent the suitability of a pixel to turn into anthropogenic areas in each transition, based on factors explanatory that are used to model the historical change process.In LCM, a collection of transition potential maps is organized within a transition submodel.A transition sub-model can consist of a single land cover transition or a group of transitions that are thought to have the same underlying driver variables (Eastman 2015).

Selection of explanatory variables
After analysis of changes, we chose the variables that would explain the anthropisation of the park.Thus, each variable was linked to a map of changes between 1986 and 2010 to obtain an idea of its level of binding using the Cramer V coefficient, and then to evaluate its explanatory power.Several authors have used this approach (Paegelow et al 2004, Maestripieri and Paegelow 2013, Mahmoud et al 2016).Two types of variables can be used as inputs to the LCM : dynamic and static (Eastman 2015).Static variables (slope, soils, boundaries of the PNDB, etc) reflect the ability of each transition taken into consideration to remain stable over time.Dynamic variables (proximity to roads, existing land use types, infrastructure, etc) were recalculated at each step of the prediction process.Cramer's V coefficient is related to the correlation between variables and varies between zero and one.
The stronger the correlation, the closer the coefficient is to one, and vice versa.A variable is considered acceptable if its coefficient is equal to or greater than 0.4 (Eastman 2015).However, this result was rejected if the coefficient was less than 0.15.Nevertheless, it should be noted that a strong correlation does not consider the complexity of the relationships among variables (Eastman 2016, N 'Gamba et al 2016).According to Eastman (2015), the coefficient correlation acts as a guide to determine whether the driving force is worth being considered or not.Thus, we decided to include in the dataset the variables with a Cramer's V coefficient below 0.15, as there is a certain interdependence between the explanatory variables.In this study, we considered the road network, hydrographic network, riparian localities, park boundaries, anthropogenic areas, soil types and population density as variables.

Evidence likelihood or proof of likelihood transformation
The Proof of Likelihood transformation is a very effective method for integrating categorical variables into the analysis.Thus, categorical variables must either be converted into a set of Boolean (dummy) variables or be transformed using the likelihood-proof transformation option (Eastman 2015).The Evidence Likelihood transformation requires two inputs: (i) a Boolean map of areas that have gone through the transition being modeled ; and (ii) A categorical variable that has been binned into classes.According to Eastman (2016), the procedure looks at the relative frequency of pixels belonging to the different categories of that variable within areas of change.In effect, it asks the question of each category of the variable, 'How likely is it that you would have a value like this if you were an area that would experience change?' Modelling multiple transitions is available using the MLP option.Logistic regression requires them to be modeled separately (Eastman 2016).Moreover, the choice of MLP over logistic regression is explained by the fact that it is highly recommended by the editors of LCM (Eastman 2016).

Calibration of the MLP neural network
This neural network consists of hierarchically interconnected units (or nodes) with an input layer, one or more hidden layers that act as black boxes, and an output layer.The input value of the neurons is that of the explanatory variables of the model and evidence likelihood proof (Paegelow et al 2004).Each of these numerical values was multiplied by a certain number of weights to be added and transformed by a transfer function at the level of hidden layer neurons.Finally, the numerical values of the neurons in the hidden layer are multiplied by weights and their addition provides the values of the output neurons that model the explained variables.The weights were chosen during the learning phase on the test dataset, and the square error of this dataset was minimized.Finally, after setting up the MLP neural network, the process of creating potential transition maps was launched, which should then be used for simulation.It should be noted that MLP is a multilayer model that works better in transition modelling and can perform multiple transitions (up to nine transitions) using submodels (Eastman 2015).These are the modelling assumptions.Beyond these nine transitions, MLP will not work.It is important to achieve a success rate of 70% to expect sub-models that tend towards reality (Paegelow et al 2004, Mahmoud et al 2016).It should be noted that the 70% rate is a threshold proposed by these authors.This threshold is given to allow the users of the model to have a reference value to appreciate the quality of their prediction.They therefore suggest it for anyone using the LCM model to predict land use dynamics.

Simulation of the land use of the PNDB
Once the model has been calibrated, the choice of prediction date is specified to launch the land cover simulation process.Markov Chains are used in the simulations.A Markovian process is one in which the state of a system can be determined by knowing its previous state and the probability of transitioning from each state to each other state.For the first simulation test, the date of the reference 2016 map was chosen.This choice was based on the possibility of validating the model with known data before performing the medium-and long-term simulations.Thus, after validation of land use modelling in 2016, another simulated land use map for 2025 was created.The results of this simulation include a vulnerability map of the area and a projection map of future land use.

Validation of the Land Change Modeler (LCM)
An important step in the modelling of deforestation and degradation dynamics in the PNDB and the simulation of land use by 2025 is to validate the model using known data.The 2016 PNDB land use map, the most recent, was used as the basis for the first simulation test, calibrated on two previous dates (1986 and 2010) according to the second-order Markov chains (Eastman 2015).Two approaches were used to validate the model: visual and statistical comparison.First, for visual comparison, verification was performed by comparing the simulated and reference 2016 maps.Second, to reduce subjectivity and lack of accuracy of the purely visual comparative approach, a statistical comparison between the two maps was performed (Pontius et al 2004).In addition, considering recent criticisms of limitations inherent to the Kappa statistic (Pontius and Millones 2008) and based on a recent study that proposed a method to quantify and visualize the relationships between land cover changes and explanatory variables, another validation was accomplished (Chen and Pontius 2010).The focus is on 'budgeting' errors and correct predictions.These authors used four categories of pixels: (i) correct pixels owing to consistency between observation and prediction (null successes[N]), (ii) errors owing to differences between observation and prediction (false alarms[F]), (iii) correct pixels owing to observed and predicted change (hits[H]), and (iv) errors owing to observed but predicted change as constant (misses[M]).To assess the accuracy of the overall prediction of changes across the landscape, the authors proposed a method to measure errors (in % of the landscape) based on the quantity and allocation based on the results of the above budgeting (table 1).The predicted change quantity error (Q) measures the percentage of imperfections in the correspondence between observed and predicted change quantities.The allocation error measures the degree of approximation of the correspondence in the spatial allocation of changes considering the specification of the number of changes in the observed and predicted change maps.In order to be able to apply this method, the map of changes observed between 2010 and 2016 and that of the predicted changes must be cross-referenced.To achieve this, the CROSSTAB function of the Idrisi Selva 17.0 software was used to generate a contingency table.After validating the first simulation test with the reference year 2016, we produced a simulation map for 2025.

Relations between the changes observed between 1986-2010 and the explanatory variables
Assessing the degree of association between the main changes and explanatory variables is important for the rest of the modelling process.Only variables with a Cramer coefficient V greater than or equal to 0.10 were represented (table 2).In this model, it is important to consider only the variables that influence observed mutations.

Error curve of the Perceptron Multi-Layer Neural Network (MLP)
The error curve is also an important criterion for validating the process of creating potential transitions in LCM.The objective of the multi-layer artificial neural network (MLP) technique is to minimize the error between the calculated output and desired output during learning (Eastman 2015).The error curve should exhibit a decreasing trend without oscillations (figure 5).
Analysis of this curve showed a decreasing trend.This indicates that the MLP neural network performed well.The obtained rate was 80.73%, which means that the transition sub-models are closer to reality.

Visual comparison of the simulation with the reference image
The simulated and observed maps (figure 6) were compared.The results showed that the spatial distribution of land use classes on both dates was relatively well-simulated.Visual analysis of the 2016 simulated map showed  that all land use classes were well distributed in the park.However, it should be noted that there were new anthropized areas in the northern part of the park on the simulated map.Overall, the simulation results are similar to those of the actual and reference maps for 2016.

Statistical validation of the LCM
Admittedly, the visual comparison was satisfactory; however, statistical validation was necessary to avoid subjectivity in model validation.Model validation was performed by comparing observed and predicted changes between 2010 and 2016.We found that 78% of the observed consistency [N] between 2010 and 2016 was correctly predicted (table 3).The small variation in the 'water bodies and bare ground' compared to other classes partly explains this result.The error owing to an observed but predicted change in consistency was 10%[F], whereas the error owing to an observed but predicted constant change[M], which reflected the opposite process, reached 8%.The observed change predicted by the model was 4% The allocation error (A) was estimated at 16% and measures the degree of approximation of the correspondence in the spatial allocation of changes, given the specification of the quantity of changes in the observed and predicted change maps.The quantity error (Q) was evaluated at 2%, and measures the percentage of imperfection in the correspondence between the observed and predicted amount of change.These relatively low error values show that the LCM model provides a good prediction of the anthropisation phenomenon.

Land cover status in 2016: simulation and reality
The analysis in figure 7 shows that for all land use classes, there is a certain similarity between the two maps (simulated and reference) in terms of surface areas.The anthropogenic areas simulated in 2016 were similar to those observed in 2016.Most of the differences were found in the north of Park, where new patches of anthropogenic areas were created using the model.Finally, the water body class remained almost unchanged between simulation and reality.

Level of vulnerability of land use classes to anthropization
The LCM highlighted areas that were potentially vulnerable to anthropization that would not have emerged if an expert approach based on cellular automata (CA) or Saaty's weighting scale were preferred.According to  Eastman (2016), in the procedure for Multi-Criteria Evaluation using a weighted linear combination, it is necessary that the weights sum to one.In Saaty's technique, weights can be derived by taking the principal eigenvector of a square reciprocal matrix of pairwise comparisons between the criteria.The comparisons concern the relative importance of the two criteria involved in determining suitability for the stated objective.
Ratings are provided on a 9-point continuous scale.LCM produces two simulated maps (soft projection and hard projection).For hard predictions, a simulated map was developed for the prediction year, in which each pixel is allocated to a specific land use category

Dynamics of land use between 2016 and 2025
The predicted land use for 2025 is shown in figure 9.According to the model, the dynamics are progressive for some land cover classes (gallery forest classes and anthropized areas) and regressive for others (savanna and bare soil).Thus, between 2016 and 2025 (figure 10), gallery forests and anthropized areas increased by 138 and 2914 ha, respectively.During the same period, wooded, dense shrubs, and clear shrub savannahs decreased by 1186 ha, 961 ha, and 893 ha, respectively.The bare soil and water bodies remained stable between 2016 and 2025.

Choice of explanatory variables
Variables explaining the anthropisation of the PNDB were chosen to calibrate the model.These variables allowed us to obtain interesting results, considering the relatively strong similarity between the simulated and reality maps.However, these variables alone are insufficient to explain the phenomenon under study.The model is a simplification of reality because the number of explanatory variables in which the potential should be included is limited by the difficulty in representing some of them spatially.These factors include political, institutional, cultural, and economic factors.For example, the political dimension includes changes in forest management policies.In addition, the needs of local populations for non-timber forest products and the harvesting of fuelwood provide information that could help to better understand the phenomenon under study.However, these variables did not influence changes in parks.Unfortunately, these data are often partially or nonexistent, making it difficult to integrate them into the model.Several authors have reported this difficulty in integrating the explanatory variables (Maestripieri andPaegelow 2013, Mahmoud et al 2016).Among the variables retained, distance from villages or cropping hamlets seems to have played an important role in the main changes observed between 1986 and 2010.This confirms that agricultural activity plays a predominant role  in deforestation and forest degradation in Burkina Faso (Tankoano et al 2016).Some authors have also demonstrated the considerable influence of agriculture on deforestation (Marien 2009, Mbemba 2012, Kambiré et al 2015).

Validation of the LCM
Land use modelling using the LCM and Markov chains has made it possible to quantify probable changes and measure the risks of deforestation by 2025.The results from the LCM revealed a significant part of the observed and simulated constancy, that is, nearly 78%.Thus, it appears that this model simulates the evolution of land use classes.However, LCM presents some difficulties in simulating land use in park, with a total error of 18%.This total error value is still higher than other studies using the LCM model, where it was 10.9 and 12% (Olmedo et al 2013 andMaestripieri andPaegelow 2013).This situation highlights the complexity of the changes within the PNDB and its periphery.This difficulty should not be explained only by the limited number of variables (N'Gamba et al 2016), but also by the heterogeneity of the landscape of the park and the fluctuation of political decisions.A comparison of observed and simulated land use in 2016 provided acceptable results for spatial allocation.The LCM predicts results close to those observed.Error budgeting indicated that the LCM produced satisfactory results.

Modelling vegetation cover dynamics
The simulation results showed a trend towards the anthropization of parks.This trend is explained by the fact that land pressure is too high in the area and farming practices considerably impoverish the soil.The regression of natural formations to the benefits of anthropized areas would certainly lead to the loss of biodiversity in the park.It is clear, an alarming and continuous decline can occur if no significant measures are taken.These results confirm those obtained in similar areas by N'Da et al  (1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003).The authors concluded that the forests of Marahoué park will disappear in the near future if no protective action is taken.Thus, increased monitoring could be at the root of Park's low exposure to local anthropisation.The inclusion of policy decisions in the model would be of great benefit to better model the vegetation cover dynamics.Thus, Dempster-Shafer's prediction method is more widely used because it uses uncertainties (fluctuations in political decisions on natural resource management) in the simulation process (Mas et al 2011, Mohamed et al 2011).

Conclusion
The objective of this study was to determine evolutionary trends in the spatial and temporal dynamics of vegetation cover in the PNDB from 1986 to 2025.Remote sensing and GIS tools have been used for this purpose.
Prospective modelling was performed according to the principles of the LCM.Analysis of the results revealed an evolutionary trend towards anthropization to the detriment of natural plant formations, except for gallery forests.As the number of relevant variables was unlimited, only spatially explicit variables were considered.Despite the simulation being close to reality, the LCM failed to accurately predict the changes between 2010 and 2016.The comparison between the simulation and the reality of 2016 highlights a significant part of the observed and simulated constancy, approximately 78%.The 2025 scenario confirms an anthropization trend.If no measures are taken to strengthen monitoring, it constitutes a serious threat to park biodiversity.This scenario is part of the process of sounding an alarm so that urgent measures can be taken to conserve the Park's biodiversity.

Figure 3 .
Figure 3. Distance maps to the road network (a), hydrographic network (b), riparian villages (c) and the boundaries (d) of the PNDB.

Figure 5 .
Figure 5. Error curve of MLP neural network training.
[H].The total observed change [OC= M + H] is 12%, whereas the total predicted change [PC= F + H] is underestimated by 14%.The accuracy of the overall prediction of changes throughout Park was as follows: Quantity error (Q) =|F-M | = 2%, allocation error (A)= (F + M)-Q = 16%, and total error

Figure 6 .
Figure 6.Reference (a) and simulated (b) land use maps of the PNDB in 2016.
(Ayele et al 2019).Soft prediction is a projected map created to show the vulnerability in which each pixel is allocated a value ranging from 0 to 1.A smaller value indicates less vulnerability to change and a high value indicates a high susceptibility to change (Ayele et al 2019).The vulnerability map (figure8) shows all possible areas of change, from the least to the most vulnerable, between 2010 and 2016.The vulnerability index ranges from 0 to 1, with 0 representing stable areas and 1 representing the highest change index.The most vulnerable areas are mainly located in the northwestern and southwestern parts of the PNDB.These areas are close to those with a high human population and agricultural fronts.The vulnerability index is relatively low.It varied from 0 to 0.48.

Figure 7 .Figure 8 .
Figure 7.Comparison of the areas of the land use classes of the reference map with the simulated one in 2016.Legend (GF : Gallery forest ; WS : Wooded savannahs ; DSS : Dense shrub savannahs ; CSS : Clear shrub savannahs ; AA : Anthropogenic areas, BS : Bare soil ; WB: Water bodies.)

Figure 9 .
Figure 9. Land use map of the PNDB in 2025.

Figure 10 .
Figure 10.Evolution of land use areas from 2016 to 2025.Legend (GF: Gallery forest; WS: Wooded savannahs ; DSS: Dense shrub savannahs ; CSS: Clear shrub savannahs ; AA: Anthropogenic areas, BS : Bare soil ; WB: Water bodies.) (2008)  for the Côte d'Ivoire and by N'Gamba et al (2016)  for the DRC.Indeed, a prospective analysis of the spatio-temporal dynamics of the Yangambi biosphere reserve in the DRC revealed a regressive trend of natural plant formations in favor of anthropized areas by 2065 (N'Gamba  et al 2016 ).The same is true for Marahoué National Park (Côte d´Ivoire), where N'Da et al(2008)  noted that agricultural clearing led to the loss of 16,378 ha of forest in 17 years Eastman 2016).Neural networks are preferred because they are more efficient than multiple regression models, particularly for complex and nonlinear systems (Akadjé 2016, Mahmoud et al 2016).

Table 1 .
Errors and accuracies measurement.

Table 2 .
Relationship between the changes observed from 1986-2010 and the variables.

Table 3 .
Contingency matrix between observed and predicted changes.