Spatio-temporal Analysis Dynamics of the Landscape in the Classified Forest of Koulbi, Southwestern of Burkina Faso

Most of the protected areas in Burkina Faso are affected by deforestation and forest degradation. This study is done in order to analysis the spatial and temporal dynamics of the classified forest of Koulbi and to identify the drivers of deforestation in the forest. For that, three Landsat images from 1986, 1998 and 2014 were exploited. The supervised classification based on the artificial neural network algorithm, under the ENVI software, was used. A transition matrix was established to analyze land use/land cover changes. The standardized precipitation index of 1985 to 2014 was used to access climate trend. Survey data from riverside villages and field observations were used to assess the human impact. The results show that between 1986 and 2014, the woodlands and tree savannas areas were reduced by 7% and 0.49% per year, respectively. However, substantial yearly increases in the areas of shrub savanna (2.8%) and agricultural fields (1.4%) were observed. During this period, the rainfall had a sharp trend towards aridity, with 18 years of drought against 12 years of humidity. This show that deficit of rainfall is increase. The socio-economic survey revealed that the classified forest used to be highly occupied and 64% of the inhabitants were yam producers. Cattle breeders and gold diggers have also been observed. We conclude that human activities are the main cause of degradation of the vegetation cover of the Koulbi classified forest. And climatic factors to some extent. It will be necessary to take actions to better conservation.


Introduction
Classified forests refer to areas that are officially protected for conservation purposes [1]. In Burkina Faso, these forests cover an area of 3.9 million hectares, representing about 14% of the national territory [2]. Recent reports suggest that the classified forests of Burkina Faso are subject to increasing and uncontrolled occupation causing significant changes in the vegetation cover [3][4]. In the southwestern region, high immigration of agro pastoralists from the north [5,6] and emigration resulting from the crisis in Ivory Coast in 2002 [7] have increased pressure on the protected areas which are the mains areas of biodiversity conservation [8][9]10]. In addition, West Africa is prone to droughts, irregular rainfall patterns and temperature increases with potentially degrading effects on ecosystem dynamics [11][12].
The classified forest of Koulbi (CFK) gained its protected status in 1955. Main threats to ecosystems conservation in this area include increasing yam production [13], intensified gold panning activities and climate disturbances. In order to plan and sustain the conservation of CFK, it is necessary to gather accurate data of vegetation cover changes and to understand the main causes of deforestation. Such information can facilitate decision making aimed at promoting sustainable management of the forest.
Satellite remote sensing has been demonstrated as an effective and low cost technique for studying land cover dynamics at different spatial scales. Regardless of the difficulties caused by spectral confusion in savanna type environments [14][15]16], remote sensing of the vegetation cover can enable the accurate detection of land use/land cover (LULC) changes [17][18]19].
The purpose of this study was to (i) map and quantify the LULC changes in CFK from 1986 to 2014 using Landsat imagery and (ii) determine the effects of climate and anthropogenic factors on the observed changes in CFK.

Study Area
CFK is located in the province of Noumbiel, southwestern Burkina Faso, and covers an area of 40 000 ha ( Figure 1).
The CFK belongs to South-Sudanian climatic zone where the average annual rainfall is over 900 mm and the rainy season extends from June to September. Mean annual  rainfall and temperature in CFK are 1057 mm and 27.9°C, respectively. The hydrographic network is dense and depends on the Mouhoun River, which forms a natural boundary between CFK and the republic of Ghana. According to Geographic Institute of Burkina Faso [20], soils are mostly found to be ferruginous, brown, eutrophic and ferralitic. The vegetation cover is mainly by savanna, but isolated areas of woodland dominated by Isoberlina spp. [21][22] and Anogeissus leiocarpa (DC) Guill. and Perr. are also common landscape features [23]. In 2014, the population of the Noumbiel province was estimated to be 89 036 [2], mainly comprising farmers who cultivate yam as the most important cash crop. In addition, this area is favored by pastoralists and is experiencing increasing gold extraction activities.

Data
Landsat images from Path195 and row53 were chosen. Two of them were Thematic Mapper (TM) images form 11/18/1986 and 11/19/1998 and one was Operational Land Imager (OLI) image from 10/30/2014. They were obtained from the United States Geological Survey (www.earthexplorer.usgs.gov/). Data from 277 randomly distributed floristic inventory plots were collected and used as reference data to facilitate the supervised classification of the Landsat imagery. Rainfall data from Batié pluviometric station (Noumbiel province) over the period 1985-2014 were obtained from the National Direction of Meteorology of Burkina Faso. In addition, a socio-economic survey was carried out in eight villages within a 5 km distance from the CFK, including Donsièrè, Fadio Gangalma, Kpéré, Médicateon, Poni, Téheni-sud and Titinateon. In each village, thirty households were chosen randomly and semi-structured interviews were conducted, resulting in a total of 240 respondents. The interviews focused on i) perceptions of degradation of the vegetation cover, ii) causes of the degradation, iii) yam production and its impact on vegetation, and iv) land abandonment caused by the population's eviction that occurred in 2009.

Pre-processing
The Landsat images were subject to radiometric and atmospheric correction in order to compensate for sensor noise, as well as atmospheric water vapor and aerosol contamination, using ENVI 5.1 software and Fast Line-ofsight Atmospheric Analysis of Spectral Hypercubes (FLAASH) [24][25]26].

Artificial Neural Network Classification
Supervised classification of the Landsat imagery was conducted using Artificial Neural Networks (ANN), which has been shown to produce more accurate results in complex landscapes, such as savanna type environments, compared to conventional classification methods [27][28]. ANN is a multilayer perceptron and trained using the backpropagation method, a commonly used error-minimization technique. In this study, the structure of ANN was based on the input layers corresponding to the six Landsat bands, one hidden layer and output layers corresponding to the six LULC classes. Following [29] experiences, the back-propagation algorithm was implemented using five internal parameters. The learning rate (i) calculates the changes to be made to the weight values after each iteration. The momentum (ii), avoids the risk of error oscillation during learning process and measures the effect of a preceding iteration on the current iteration. The activation function (iii), often represented by a sigmoid or hyperbolic tangent, normalizes the values attributed to the neurons. The stop level (iv) varies between 0 and 1 and the iteration number (v) is set by the user. In this study, ANN parameters were calibrated following the experiences of [28] in complex LULC, with the iteration number = 1000, learning rate = 0.01, momentum = 0.9 and stop level = 0.1.
The Landsat classification focused on six LULC classes, including gallery forest, woodland, tree savanna, shrub savanna, field and fallow and water. From the reference data, 152 plots were used for training the 2014 classification, and 125 plots were used for independent validation. For the 1986 and 1998 classifications, training and validation sites were selected from areas where no change had occurred during the study period. These areas were identified using the 2014's LULC map and LULC databases from 1992 and 2002 of IGB.
In order to reduce "salt and pepper effect" and to improve cartographic representation, a 3x3 median filter was applied to all classification results. Classification accuracy was assessed using standard metrics, including confusion matrix, Kappa coefficient and error curve.
The annual average change rate for each LULC class was obtained using Equation 1 [31]: Where r represents the annual variation rate for class i; A 1 is the surface area of i th class and at time t 1 and A 2 is the surface area of class i at time t 2 .

Calculation of Standardized Precipitation Index
The amount and spatiotemporal distribution of rainfall control plant growth and need to be considered in an analysis of vegetation dynamics. We applied the Standardized Precipitation Index (SPI) to quantify rainfall shortages over defined time periods and related it to the observed vegetation dynamics; [32]. SPI was derived using Equation 2: Where Xi represents annual rainfall in year i, whereas Xm and Si represents mean and standard deviation of observed annual rain, respectively. The index values indicate details on the moisture or the environmental drought.

Validation of Classification
Accuracy assessment results showed that the overall classification accuracy was 83.0%, 86.3% and 87.2% for 1986, 1998 and 2014, respectively ( Table 1). The error curves had a decreasing shape (Figure 2) indicating that error was minimized between calculated error and that expected during classification. The results show that spectral confusion occurred between galleries forest and woodland. This has contributed to the low classification accuracy for gallery forest (Table 1).

Land Use/Land Cover Dynamics from 1986 to 2014
The LULC maps of CFK (Figure 3) show the composition and configuration of the landscape. In 1986, the area was dominated by tree savanna (46.0%) and woodland (22.4%; Figure 4). From 1986 to 1998, the woodland area experienced a strong decrease (22.4% to 5.7%), whereas large increases in tree savanna (46% to 58.4%), shrub savanna (21.3% to 27.3%) and agricultural fields (4.6 to 6.1%) were observed (Figure 4). In addition, between 1998 and 2014 the area of tree savanna decreased (58.4% to 40.2%) and shrub savanna increased (27.3% to 46.8%; Figure 4). During the period 1986-1998, tree savanna was the dominant class, whereas shrub savanna dominated in the second period (1998-2014).  The rate of LULC change (Table 2) showed that the woodland area decreased by 6600 ha from 1986 to 1998, representing an annual loss of 11.4%. A loss in woodland was also observed between 1998 and 2014, with an average decrease of 4.3% per year. In contrast, shrub savanna increased from 2.1% to 3.4% per year for the periods 1986-1998 and 1998-2014. The area of agricultural fields increased by 2.5% per year from 1986 to 1998 while it was stable from 1998 to 2014.
The tree savanna had an uneven variation during the two periods: it increased 2.0% per year from 1986 to 1998, but decreased by 2.3% per year between 1998 and 2014 (Table  2). The loss of woodland (1986-1998) favored the development of tree savanna, which then underwent degradation to shrub savanna in the second period. Table 3 provides details about the loss of woodland and tree savanna during the study period. The main trend is a change from woodland to savanna. For the same time, 19.87% of shrub savanna in 2014 was tree savanna in 1986. Finally, shrub savanna area grew up and became the main matrix of CFK.
In summary, during these 28 years, the vegetation cover has experienced major changes. This shows that despite protection efforts, the CFK has degraded. The causes of this degradation are elucidated in the following sections. Table 3. Land use/land cover transition matrix (area %) for observed time period (1986-1998-2014

Climatic Causes
SPI derived from a 30-year rainfall time series show that CFK experienced 4 years of high drought (HD), 14 years of fair drought (FD), 6 years of fair humidity (FH) and 6 years of high humidity (HH). Overall, drought years (18) were more frequent than humid years ( Figure 5), indicating that the area tends to be arid.

Local Perceptions
The degradation of the vegetation cover in CFK was noted by 95% of the respondents who attributed it to different factors, including demographic pressure (97%), bush fires (93%), transhumance (89%), yam cultivation (88%) and climate change (69%). More than 50% of survey respondents confirmed their stay in the classified forest before 2009 (Table 4). In Poni and Gangalma villages, respectively 93% and 83% of the respondents are former inhabitants of CFK.
The results suggest that the deforestation increased as a result of increasing yam production. Indeed, 62% of the respondents are found to be yam producers, among which 97% stated that after a year of production, the soil tends to be inappropriate for yam cultivation. These fields are instead used for the cashew (Anacardium occidentale L.) cultivation ( Figure 5c) and other activities for the coming season. The abandoned fields in the CFK consecutive to the population's eviction are estimated to 1676.7 ha (Table 4). Producers practice slash-and-burn shifting agriculture where the burned trees are used as support for yam (Figures 6a and 6b).

Validation of Classification
The overall accuracies of the classification and the kappa coefficients showed that the supervised classification resulted in a successful mapping of the LULC (1986, 1998 and 2014) in the CFK. The decreasing error curves confirmed the confusion matrix values, revealing that the classification was sufficiently accurate [28]. The confusion between gallery forest and woodland can be explained by the similar spectral properties of these two classes. Woodlands, especially those dominated by Anogeissus leiocarpa, were found in depressions and along water bodies [23] and tended to be mixed up with the gallery forest areas. In the Sudanian zone, mixed vegetation and bush fires often contribute to an increased spectral confusion [33]. This phenomenon was observed in this study where the limit between different savanna classes was very subtle. An increased number of training areas and the use of historical land cover data contributed to reducing the classification error [34].

Land Use/Land Cover Dynamics
The CFK has experienced major changes of the LULC between 1986 and 2014. A substantial increase in the area of the savanna classes area as a result of the decreasing woodland which is considered as islets [23]. A significant degradation of the primary vegetation cover has occurred during the last three decades with a large increase of shrub savannas and agricultural fields. The reduction of closed forests and increase of open tree cover classes has been reported in other areas of the Sudanian zone of Burkina Faso, including Sissili province, the Bontioli wildlife reserve and the classified forest of Tiogo 4]. The leading causes of shrub savannas development were that woodlands and trees savannas were either cleared for agricultural activities or subjected to selective wood cutting. The observed increase in agricultural fields between 1986 and 1998 was confirmed by the socio-economic survey which revealed a high land use in the forest before the people eviction that occurred in 2009. More than 50% of the respondents admitted to use protected land in the CFK. Results from this study show that an increasing area of land has been used for yam cultivation, which requires more space since yam fields are usually abandoned after a single production cycle. After the governmental decision to evict people from CFK, the formerly cultivated spaces were transformed into shrub savanna. Several authors 36] reported that pressures from land use and settlements were among the leading causes of the fragmentation of homogeneous ecosystems.
The SPI analysis suggests that CFK has experienced decreasing rainfall since the 1970s. These results are in accordance with evidence reported by [37][38]39] in other places in West Africa. For the period covered by this study, twenty-four (24) years of rain shortage were observed. The decreased rainfall along with sequential droughts may have disrupted the woody cover. Indeed, [40][41] reported that an isolated year of drought, even if extreme, often cause less negative effects to the woody vegetation than a two-year (or more) consecutive fair droughts. According to these authors, the persistence of even fair droughts would be a major risk in terms of tree cover degradation. The rain shortage reduces the forest and savanna ecosystems [42], increases the negative effects of bush fires [43] and affects the establishment or the recruitment of woody species as well as the natural regeneration [44].

Conclusion
This study analyzed the effects of climate and anthropogenic factors on landscape dynamics in the CFK. It enabled the quantification of the ongoing major changes and disentangled the effects of human activities and climate on the decreasing vegetation cover. The period 1986-1998 was characterized by a high decrease of the woodland area as a result of the increased trees savanna. From 1998 to 2014, the area of tree savanna also experienced a high decrease resulting from the increased of shrub savanna and that of the fields.
The study revealed that the forest was under high land use before 2009 and more than 50% of the survey respondents in the villages have confirmed to be former inhabitants in the CFK. Moreover, the high yam production was responsible for the degradation of the forest. Regarding the impact of climate, it was found that during a 30-year period, this region was dominated by high numbers of drought years (18 droughts). The observed increasing agricultural fields in 2014 and that of recurring gold panning activities in CFK despite the fact that population was forced to move by government suggested that focus should be on the monitoring of this forest by forest officials.