Prediction of land use and land cover change in two watersheds in the Senegal River basin (West Africa) using the Multilayer Perceptron and Markov chain model

ABSTRACT

Land use and Land cover change (LULCC) is a major global problem, and projecting change is critical for policy decision-making. Understanding LULCCs at the watershed level is essential for transboundary river basin management. The present study aims to analyse the past and future LULCCs in two significant watersheds of the Senegal River basin (SRB) in West Africa: Bafing and Faleme. This study used Landsat images from 1986, 2006 and 2020 and the Random Forest classification method to analyze past LULCCs in these two watersheds. The results revealed: In Bafing, vegetation, settlement, agricultural areas and water increased, while the bareground decreased significantly between 1986-2020. In Faleme, two periods have different trends. Between 1986-2006, vegetation, settlement, agricultural areas and water increased, while bareground decreased. Between 2006-2020, settlement increased, while vegetation, agricultural areas, water and bareground decreased. To predict LULCCs in 2050 under business-as-usual assumptions, the Multilayer Perceptron and Marcov Chain model (MLP-MC) was used. The MLP-MC shows better results on Bafing than on Faleme but without questioning its application on the two watersheds. Bafing has seen a trend towards ”more people, more trees”, while Faleme has seen a trend towards ”more people, more deforestation”. These results contribute to develop appropriate land management policies and strategies to achieve or to maintain sustainable development in the SRB.

KEYWORDS:

• Land use land cover change
• multi-temporal analysis
• MLP-MC
• Random forest
• Senegal river basin

Introduction

Land use and land cover change (LULC) is a challenge and a key focus of global change research. Land use (habitat, agricultural) and land cover (forests, wetlands, grasslands and water) have different meanings and are often used interchangeably (Tadese et al., 2021). LULC changes are the result of changes in the Earth’s land surface, such as the transformation of natural land cover (forests, grasslands and deserts) into human-dominated ecosystems (cities, agricultural and industrial areas) (Liping et al., 2018; Winkler et al., 2019). These changes significantly affect critical elements of our natural capital, such as vegetation, water resources and biodiversity (Chang et al., 2018; Solly et al., 2021). The dynamics of LULC change are not the same in all regions of the world due to different driving factors (Berihun et al., 2019). Although anthropogenic factors have been proven to be the main drivers of changes, factors such as climate, slope, appearance and altitude can also influence these changes (Anwar et al., 2022; Mekonnen & Manderso, 2023). In West Africa (WA), notable LULC changes have been observed (Andrieu, 2018; Barnieh et al., 2020; Cabral & Lagos, 2017; Diallo & Zhengyu, 2018; A. Traore et al., 2018; S. S. Traore et al., 2022). These authors shown that WA has undergone significant changes over the years, and the causes are generally attributed to rapid population growth and increased agricultural areas. Even if the information on LULC changes at the regional or national scale is available, it is essential to study at the local scale (Fikadu & Olika, 2023). Indeed, studies of watershed-based LULC changes are crucial tools for policymakers, planners and local communities to formulate appropriate policies and strategies in the future (Berihun et al., 2019; Thiam et al., 2022).

Advances in remote sensing and Geospatial Information Systems (GIS) have resulted in high-resolution products and LULC models for mapping, detecting and predicting LULC changes (Liping et al., 2018; S. K. Singh et al., 2015; Wang et al., 2021). Google Earth Engine (GEE) is an open-source, cloud-based geospatial processing platform that provides free access and open-source satellite datasets (Landsat, Modis, Sentinel) with a high spatial resolution for extended periods (L. Yang et al., 2022). Google Earth Engine (GEE) is the most popular big geo data processing platform, which provides a set of state-of-the-art classifiers for pixel-based classification used for LULC mapping. The main advantage of GEE is the close link between the data and the algorithms, both of which are accessible via an Application Programming Interface (API) (Gorelick et al., 2017; Shelestov et al., 2017). Due to its accessible and user-friendly design, it has grown in popularity recently (Dubertret et al., 2022; Jampani et al., 2020; Nasiri et al., 2022; Ougahi & Mahmood, 2022).

The success of any LULC mapping depends on several factors, including the choice of an appropriate classification algorithm (Lu & Weng, 2007). Advanced classification algorithms such as Regression Trees (CART), Random Forest (RF), kNearest Neighbor (k-NN), Support Vector Machine (SVM), Artificial Neural Network (ANN), Multinomial Logistic Regression (MLR), Maximum Likelihood Classification (MLC), and Bayesian classifiers have attracted considerable attention in image classification for LULC mapping (Ma et al., 2019; Macarringue et al., 2022). Authors such as A. D. Kulkarni and Lowe (2016) and Talukdar et al. (2020) have conducted comparative studies between several classification algorithms to identify the most suitable and accurate classification algorithm for LULC mapping. Their results indicated that Random Forest showed the best performance for the LULC classification compared to other classifiers. The main factors behind its widespread use in multitemporal LULC classification are (1) effective management of outliers and noisy datasets; (2) satisfactory results with multi-source and high-dimensional datasets; (3) superior accuracy to other widely used classifiers, like SVM or MLC in many applications; and (4) speeding up processing by concentrating on essential factors (Amini et al., 2022; Noi Phan et al., 2020).

Modelling LULC changes using remote sensing data and determining factors helps answer the question of how LULC change and how it may change in the future (Sankarrao et al., 2021). Several methods have been adopted to predict future LULC changes. These methods can be classified into non-hybrid and hybrid methods. The non-hybrid approaches, such as Markov Chain (MC), Artificial Neural Network (ANN), Cellular Automata (CA), have been widely used to identify the transitions in LULC classes and have been accurate in predicting LULC changes (Silva et al., 2020; V. G. Singh et al., 2022). The Non-hybrid methods have limitations in predicting LULC changes (Noszczyk, 2019). To overcome the limits of individual models, hybrid models have been introduced by combining several modelling approaches to address the complexity of real-world systems (Gaur et al., 2020). The hybrid models can capture LULC changes with greater accuracy (Gaur et al., 2020; Sankarrao et al., 2021). Clark Labs developed an integrated LULC model based on GIS and remote sensing, known as the Land Change Modeler (LCM), to explore future changes in LULC using a Multilayer Perceptron neural network (MLP) and Markov chain (MC) (V. Mishra et al., 2014). Given its robustness and popularity, the MLP-MC model seems the most suitable for modelling spatiotemporal dynamics and projecting future LULC change scenarios. Gaur et al. (2020) used hybrid and non-hybrid models to capture LULC scenarios for the Subarnarekha River and found that the MLP_MC model was the best-suited model. Examples of applications can be found in (Fathizad et al., 2015; Leta et al., 2021; V. N. Mishra et al., 2018; Sankarrao et al., 2021).

The Upper Senegal River Basin (SRB), located in WA, is a transboundary basin bordered by four countries: Guinea, Mali, Mauritania and Senegal. It is formed by several watersheds. Understanding LULC changes at a watershed level in transboundary watershed management is crucial (Thiam et al., 2022). Unfortunately, there are few studies on LULC changes in the SRB. Only one previous study was conducted by Faty (2017). Faty (2017) used Modis-Terra satellite data from 2007 and 2014 and maximum likelihood classification to analyze the LULC dynamics across the SRB. The main limitation of this study is that the author evaluated the LULC changes considering that all watersheds constitute a single socioeconomic, land and agricultural environment. The author didn’t consider the influence of the socioeconomic configuration specific to each watershed. Although LULC change is a common phenomenon, it is challenging to generalize trends within a particular region, such as the SRB (Berihun et al., 2019). In addition, no study has attempted to simulate the future trends of LULC changes considering each watershed in the SRB. Therefore, as part of this study, we aim to fill these knowledge gaps to understand the spatiotemporal variability of LULC changes in two watersheds in different socioeconomic and agricultural environments at the SRB: the Bafing and the Faleme. This study aims: (i) to map the LULC (1986, 2006 and 2020) in the two sub-watersheds with the Random Forest (RF) classification method; (ii) to simulate the future trends of LULC change of 2050 based on the hybrid model (MLP_MC) in LCM and (iii) to analyze the spatiotemporal variability of LULC in these two watersheds. This study is relevant for the riparian states of the Senegal River basin grouped into the organization for the development of the Senegal River (OMVS) to evolve policies specific to each watershed.

Materials and methods

Study area

The study area is in the SRB, between Senegal, Mali, Mauritania and Guinea Conakry (Figure 1). The two watersheds belong to different administrative regions:

• The Bafing is the main constituent of the Senegal River and originates near Mamou (Guinea Conakry) (nearly 1,000 m above sea level). It extends between latitudes 10°30“and 12°30” N and longitudes 12°30’ and 9°30‘W and covers northwest Guinea Conakry and southeastern Mali. It drains the entire eastern part of the Fouta Djalon and feeds the Manantali Hydropower dam (Sambou et al., 2023)

• The Faleme also has its source in the northern foothills of the Fouta Djalon in Guinea Conakry. It is a major tributary of the Senegal River, which it joins at the meeting point of the borders of Senegal and Mali at the level of the city of Ballou. The Faleme lies between latitudes 12°11′–14°27′N and longitudes 11° 12‘−12°15‘W in the humid tropical regions and the southern margins of the Sahel. The main biophysical characteristics of the watersheds studied in the SRB are presented in Table 1.

Figure 1. Location map of the bafing and faleme in the SRB in WA, between senegal, mali, Mauritania, and guinea conakry.

Table 1. The main biophysical characteristics of the watersheds. Source: (Sane et al., 2020).

Characteristics	Bafing	Faleme
Area (Km2)	39.01	28.05
Mean Annual Temperature (C)	28.3	29.3°C
Agro-Ecology Zone	Guinean Sudano	Guinean Sudano
Mean Annual Rainfall (mm)	1166	800
Major Soil Type	Gres Qtz., Dolerites, Granite, Sandstone	Shale, Gres Qtz., Dolerites, Granites
Dominant Crop	Cassava, Maize, Fonio, Sorghum and Millet.	Maize, Rice, Millet, and Sorghum.
Dominant Livestock	Bovine	Bovine
Activities	Fishing, Agriculture area, Livestock, Bauxite Mining	Fishing, Agriculture area, Livestock, Gold Mining

Data source

The acquisition of satellite images is the first step in LULC mapping. The choice of satellite images is based on four criteria: spatial coverage, spatial resolution, available years, and cloudiness (minimum) (Horning, 2004; Martignac,). Landsat satellite images were chosen, in this study. It is the oldest of the earth observation programs, thus having an archive of images over a long period (more than 30 years) (Woodcock et al., 2008). It offers a sufficient level of detail to identify the characteristics of the LULC. Google Earth Engine (GEE) provides surface reflectance images that are corrected atmospherically and improve the detection of changes (Wahap & Shafri, 2020). The images selected are presented in Table 2. As supplementary input data, we included the digital elevation model (DEM) (ALOS World 3D–30 m), distance from the road, distance from the river, and distance from the settlement.

Table 2. Characteristics of Landsat images selected for the LULC mapping.

Dataset	Satellite	Sensor	Spatial resolution	Date of acquisition	Band
Image 1	USGS Landsat 5 (Surface Reflectance Tier 1)	MSS/TM	30 m	1986	Multispectral
Image 2	USGS Landsat 5 (Surface Reflectance Tier 1)	MSS/TM (Surface Reflectance Tier 1)	30 m	2006	Multispectral
Image 3	USGS Landsat 8 (Surface Reflectance Tier 1)	OLI/TIRS (Surface Reflectance Tier 1)	30 m	2020	Multispectral

Land use and land cover mapping

This study used GEE (Gorelick et al., 2017; Shelestov et al., 2017) to build the LULC map through Landsat image processing, supervised classification, and classification accuracy. Figure 2 describes the general steps followed to make the maps. Details will be provided in the following sections.

Figure 2. The procedures used for the LULC classification map.

Preprocessing

Orthorectified and atmospheric corrected Landsat Surface Reflectance Tier 1 images available for Landsat OLI/TIRS and MSS/TM sensors were collected in GEE for the years of interest (1986,2006,2020). The cloud cover function in GEE was applied to select the collection of annual images with a cloud cover of < 15%. The median ee.Reducer function on GEE was then used to “reduce” the image collection to a single output image representing the median of the images (Dubertret et al., 2022; Noi Phan et al., 2020).

Construction of features

The spectral and topographic features were combined as input features for LULC classification algorithm. For spectral characteristics, blue, green, red, Near Infrared (NIR), and Shortwave Infrared (SWIR) spectral image bands were selected because they have the potential to discriminate similar spatiotemporal phenomena, thus improving the separability of LULC classes (Thiam et al., 2022). In addition, numerous studies have shown that the use of spectral features from indices such as normalized difference vegetation index (NDVI) (1), normalized water difference index (MNDWI) (2), normalized difference accumulation index (NDBI) (3) as input features for classification will effectively improve the accuracy of LULC (K. Kulkarni & Vijaya, 2021; Tsai et al., 2018). Indeed, the NDVI supplies information on the characteristics of the vegetation cover. The NDWI provides information on the characteristics of water bodies. The NDBI is used to get the artificial characteristics of the earth’s surface. In addition, topographical features such as altitudes and slopes increase the accuracy of land cover classification (Y. Yang et al., 2021). Hence, the elevation and slope data were extracted from the DEM as features for LULC classification (Table 3).

(1) $\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}}$(1)

(2) $\mathrm{M}\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}=\frac{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}$(2)

(3) $\mathrm{N}\mathrm{D}\mathrm{B}\mathrm{I}=\frac{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}-\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}+\mathrm{N}\mathrm{I}\mathrm{R}}$(3)

Table 3. Spectral and topographic features.

Type of features	Feature Name
Spectral Bands	Blue, green, red, NIR, SWIR
Indices	NDVI, NDBI, MNDWI
Terrain	Elevation, slope

Classification

In this study, the RF classification algorithm was used to produce LULC maps for years 1986, 2006, and 2020, respectively. Random Forest algorithm is a method using tree-type classifiers {h(x,Hk), k = 1,…,} where the “Hk” stands for independent identically distributed random vectors and “x” stands for an input pattern (Breiman, 2001). In training, the RF algorithm creates multiple CART-like trees, each trained on a bootstrapped sample of the original training data. It determines a split by examining a randomly selected subset of the input variables (Gislason et al., 2006). In the classification process, each tree provides a unit vote for the most popular class at input x, and the classification of each tree is referred to as a “vote” for that class (A. D. Kulkarni & Lowe, 2016). A complete mathematical presentation of RF is presented by Breiman (2001). The two important parameters the user must optimize to get more accurate results are the number of trees (Ntree) generated and the number of features randomly chosen to divide each node (Mtry). Based on the pretests from our data, the number of trees was set to 100, and Mtry was set to the default value (square root of the total number of features). Five LULC classes were used in the classification, namely (1) settlement, (2) water, (3) vegetation, (4) agricultural areas (5) bareground. The details are specified in Table 4. The same LULC classes were used for both watersheds to enable the analysis and comparison of the trends between the two watersheds. The choice of these five classes was based on information from key stakeholders in each watershed and followed the trend of previous regional studies, which employed comparable classes (Faty, 2017; Thiam et al., 2022). To perform RF, samples for each class were taken from Google Earth. These samples were used as regions of interest (ROI) to train the RF. Each ROI was given a certain LULC class designation. 70% of the sample was used for training, while 30% was used for internal model validation.

Table 4. Description of the five classes used in the study.

Class	Name	Description
1	Settlement	The human (urban and rural) settlement, housing, roads, transport, mining, and industry.
2	Water	Rivers, streams, ponds, lakes, reservoirs, estuaries, and wetlands,
3	Vegetation	Forest, savannah, riparian vegetation and mixed forest, and mixed forest land.
4	Agriculture area	Agriculture area (irrigated crops, rainfed crops), Pasture.
5	Bareground	Deserts, sand fields, exposed bareground rock, sand and temporary bare ground, transitional areas, mixed barren land.

Classification accuracy

Assessing the reliability of the classification results is essential in remote sensing applications. Many researchers recommend using a confusion matrix (Table 5), as well as the accuracy indicators derived from the confusion matrix, including overall accuracy (OA), user accuracy (UO), and producer accuracy (PO) to represent accuracy (Foody Giles, 2022; Szantoi et al., 2021). These indicators show the agreement between the LULC classification results and the actual LULC (Olofsson et al., 2014). In this study, a confusion matrix was generated in GEE. OA (4), UA (5), and PO (6) and kappa index (KA) (7) were then used to evaluate the reliability of the classification (Table A1). All these indices (4,5,6,7) take values between 0 (indicating disagreement) and 1 (indicating perfect agreement)). A value above 0.80 for the KA is considered satisfactory (Stehman, 2014).

(4) $O{A}_i={\sum }_{\mathrm{j}=1}^{\mathrm{q}}\mathrm{P}\mathrm{j}\mathrm{j}$(4)

(5) $U{O}_i=\frac{\mathrm{P}\mathrm{i}\mathrm{i}}{\mathrm{P}\mathrm{i}}\hspace{0.17em}$(5)

(6) $P{O}_i=\frac{\mathrm{P}\mathrm{j}\mathrm{j}}{\mathrm{P}\mathrm{j}}\hspace{0.17em}$(6)

(7) $K_i\hspace{0.17em}=\hspace{0.17em}\frac{\mathrm{N}{\sum }_{\mathrm{i}=1}^{\mathrm{q}}\mathrm{P}\mathrm{i}\mathrm{i}-\hspace{0.17em}{\sum }_{\mathrm{i}=1}^{\mathrm{q}}\left(\mathrm{P}\mathrm{i}+\ast \mathrm{P}+\mathrm{i}\right)}{{\mathrm{N}}^2\hspace{0.17em}-{\sum }_{\mathrm{i}=1}^{\mathrm{q}}\left(\mathrm{N}\mathrm{i}+\mathrm{N}+\mathrm{i}\right)}$(7)

Table 5. Typical confusion matrix for classification validation (This table has been adapted from Roland (2021)).

Classes	1	2	…	k	…	q	Total
1	P11	P12	…	P1k	…	P1q	P1+
2	P21	P22	…	P2k	…	P2q	P1+
…	…	…	…	…	…	…	…
K	Pk1	Pk2	…	Pkk	…	Pkq	Pk+
…	…	…	…	…	…	…	…
q	Pq1	Pq1	…	Pqk	…	Pqq	Pq+
Total	P +1	P +2	…	P+k	…	P+q	P

P = number pixels

Land use land cover change prediction

This study used the MLP-MC model embedded in the LCM to simulate future LULC maps. The prediction of future LULC was done in five steps: change analysis, identification of explanatory variables, creation of transition potential maps, change prediction, and validation. LULC maps of the years 1986 and 2006 were employed to analyze the trend of change, to calculate transition potential maps and to predict the LULC map of 2020. For model validation, the LULC map for 2020 was compared with the predicted LULC map of 2020. After demonstrating our model’s ability to predict the LULC map of 2020, the same simulation technique was used to predict the LULC maps of 2050 using the LULC maps of 1986 and 2020 based on the business as usual (BAU) scenario (J. Mas et al., 2014). The BAU scenario is a scenario in which future LULC distributions follow the trends observed in the past and are formulated based on the transition probabilities and driving factors as predicted by the model. We assume that the climate will not disrupt human activities. The general scheme of the study is presented in Figure 3.

Figure 3. The procedures used to simulate the future LULC map.

Changes analysis

The first step was the change analysis to define the transition classes. Changes are described as transitions from one class of LULC to another (Azari et al., 2022). The change analysis was performed by using the two earlier LULC maps with the module change analysis in LCM. The module change analysis estimates the gains and losses of each class between the two earlier LULC maps. Changes in terms of loss (Lij) and gain (Gij) are calculated using equations (8) and (9) (Thiam et al., 2021).

(8) $\mathrm{L}\mathrm{i}\mathrm{j}=\left(\mathrm{P}\mathrm{i}-\mathrm{P}\mathrm{i}\mathrm{i}\right)(\frac{\mathrm{P}\mathrm{j}}{\sum \mathrm{j}=1\mathrm{P}\mathrm{j}}),\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{i}\ne \mathrm{j}$(8)

(9) $\mathrm{G}\mathrm{i}\mathrm{j}=(\mathrm{P}\mathrm{j}-\mathrm{P}\mathrm{j}\mathrm{j})(\frac{\mathrm{P}\mathrm{i}}{\sum \mathrm{i}=1\mathrm{P}\mathrm{i}}),\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{i}\ne \mathrm{j}$(9)

where Lij is the proportion of loss from category i to j under random processes of loss, Pii is the proportion of category i that showed persistence between the two times, Gij is the proportion of gain from category i to j, Pj is the proportion of the landscape in category j at the end of time, Pjj is the observed persistent proportion of category j, and Pi is the entire area of category i at the starting point.

Identification of the explanatory variables

The second step was the identification of the explanatory variables that have driven past LULC changes. Based on the literature, we selected slope, elevation, distance to the river, distance to the settlement, and distance to the road as the main variables influencing the change in LULC over time (Chinwendu, 2019; Murgante et al., 2014). Slope and distance to the river were used to represent the accessibility of a neighbourhood. Distance to road and distance to settlement were selected to highlight the proximity of urbanization.

Transition potential modelling

The third step was the determination of transition potential maps with the multilayer perceptron (MLP) model in LCM. MLP is composed of a neural network based on a feed-forward algorithm with three layers: the input, hidden, and output layers (J. F. Mas & Flores, 2008). MLP adjust the weights of the input and output layers using the backpropagation process. It only incorporates conductive factors with strong predictive capacity into the computation procedure, resulting in different transition potential maps for each sub-model. These maps show the ability to change from one LULC class to another (V. Mishra et al., 2014). The MLP model was trained and tested using the explanatory variables and the change analysis obtained between the two earlier images as input. The MLP first created a random sample of cells that transitioned among LULC classes during the required time and started the automatic training process. The sample is divided into two equal parts, 50% of the sample for training and the remaining 50% for testing the performance. Transition potential maps were generated after the successful execution of MLP training for each class.

Change prediction

The fourth step was using Markov Chain (MC) in LCM to predict the LULC map for a specified future date. The historical rate change determined during the change analysis phase and transition potential maps are used as input in the MC model to predict the future LULC. The MC model is a stochastic process that shows the probability that one state will change into another. The MC model uses this information as the basis to predict future changes. By examining past changes, MC model creates a transition probability matrix of changes (V. N. Mishra et al., 2018). Based on a projection of the transition potentials into the future, the technique estimates precisely the amount of LULC expected to transition from the later date to the predicted date and provides a transition probability file.

Validation

The validation process aims to verify the accuracy of the predicted map compared to a reference map. The validation process in LCM involves cross-tabulation in a three-way comparison between the later LULC map, the predicted LULC map, and the actual map. Two approaches have been used for model validation. The first approach is to use the metric performance indicators relative operational characteristic (ROC) and the validation statistics of various Kappa indices (Kappa for no information (Kno), Kappa for location (denoted Klocation), KIA for Kstandard) between the classified map and the predicted 2020 map to assess the accuracy of the prediction (Pontius & Batchu, 2003). A LULC model is valid if the Kstandard is greater than 70%, according to Zadbagher et al. (2018). The second approach is to compare the predicted and actual area of each LULC class.

Results

Land use land cover maps accuracy

This section presents the results of LULC classification for the two watersheds. Figures 4, 5, 6 and 7 show the results of the independently classified images for 1986, 2006 and 2020. Table 6 presents the classification accuracies. The overall accuracy is equal to 96%, 95% and 95% for 1986, 2006 and 2020, respectively. All these indices (user accuracy (UA), and producer accuracy (PA)) take values between 0.75 and 0.98 (indicating perfect agreement). The Kappa index obtained for both maps is above the threshold of 0.8. A value above 0.80 is considered satisfactory (Olofsson et al., 2014; Szantoi et al., 2021). Therefore, the classification could be identified as accurate.

Figure 4. LULC maps of the Bafing for the years a) 1986, b) 2006, and c) 2020.

Table

	LULC Types_Bafing
Classe	Name	1986	2006	2020
1	Settlement	2.77	15.95	17.93
2	water	0.59	2.87	3.3
3	vegetation	35.66	42.73	43.85
4	cultivated_area	0.83	3.97	4.04
5	Bare ground	60.15	34.47	30.87
	Total	100	100	100

Figure 5. Percentage of area per LULC category for 1986, 2006, and 2020.

Figure 6. LULC maps of the Faleme for the years 1986, 2006, and 2020.

Figure 7. Percentage of area per LULC category for 1986, 2006, 2020.

Table 6. Classification accuracies: user’s accuracy (U.A), producer’s accuracy (P.A), overall accuracy (O.A) and the Kappa accuracy (K.A) for each of the selected images.

Landcover/land use class	1986		2006		2020
	PA	UA	PA	UA	PA	UA
Settlement	0.84	0.96	0.94	0.93	0.97	0.95
Water	0.99	0.99	0.98	0.99	0.93	0.96
Vegetation	0.99	0.99	0.98	0.96	0.96	0.95
Agricultural areas	0.75	0.87	0.90	0.97	0.85	0.93
Bareground	0.98	0.95	0.97	0.94	0.96	0.95
Overall accuracy (O.A)	0.96		0.95		0.95
Kappa accuracy (K.A)	0.95		0.94		0.94

Land use land cover change detection

Bafing

According to Figures 4 and 5, in 1986, the area covered by bareground was the most dominant LULC class, covering 60% of the watershed. Over the 34 years, this area has gradually decreased to almost half, from 60% to 30%. The vegetation area represents the second most dominant LULC class, covering 36% of the watershed. This class has continuously increased from 36% in 1986 to 44% in 2020. Settlement were the third most dominant LULC class and covered 2.8% in 1986. However, settlement significantly increased to 16% in 2006 and 18% in 2020. The agricultural areas were the fourth LULC class, covering 0.8% of 1986 in the watershed. The extent of agricultural areas increased from 0.8% in 1986 to 4% in 2020. The area covered by water was the lowest but increased continuously over the study period from 0.6% to 3.3%.

Faleme

Analysis of the LULC map of 1986 (Figures 6 and 7) shows a predominance of bareground, representing an area of 78% in the watershed. Overall, the area covered by bareground continuously decreased from 78% to 44% during the study period. Vegetation were the second most dominant LULC class, covering 20% in 1986, as in Bafing. However, the extent of vegetation area increased from 20% to 28% in 2006 and decreased from 28% to 23% in 2020. Settlement were the third most dominant LULC class, with an area of 1.7% in 1986. Settlement class increased steadily to 10% in 2006 and 28% in 2020, becoming the second most dominant LULC class. Between 1986 and 2020, the area covered by water increased from 0.46% in 1986 to 1.3% in 2006 but decreased from 1.3% (in 2006) to 1.1% in 2020. The agricultural areas represented the lowest coverage, covering less than 0.2% of the watershed in 1986 but increasing markedly from 0.2% to 4.7% in 2006. However, the agricultural areas class showed a downwards trend between 2006 and 2020, from 4.7% to 3.5% in 2020.

Table

	LULC Types_Faleme
Classe	Name	1986	2006	2020
1	Settlement	1.7	10.15	27.88
2	water	0.46	1.33	1.09
3	vegetation	19.62	27.77	23.22
4	cultivated_area	0.2	4.71	3.54
5	Bare ground	78.02	56.05	44.28
	Total	100	100	100

Land use land cover change prediction

Transition potential

The most significant surface state changes were analyzed between 1986, 2006 and 2020 to select the dominant changes in the modelling procedure. From 1986 to 2006 and from 2006 to 2020, conversions from bareground to vegetation, settlement, agricultural areas, and water were the most significant for Bafing. For Faleme, the overall distribution results for 1986, 2006, and 2020 showed that the most significant changes for the period (1986–2006) were transitions from bareground to settlement, agricultural areas, vegetation, and water. Over 2006 and 2020, conversions from vegetation and bareground to all other classes, including water, agricultural areas and settlement, were the most significant.

Model validation

Bafing

On the simulated LULC map of 2020 (Figure 8b), the areas occupied by settlement, vegetation, bareground, water and agricultural areas are 18.75%, 44.76%, 28.50%, 3.91%, and 4.08%, respectively, against 17.93%, 43.85%, 30.87%, 3.30%, and 4.04%, respectively, on the reference map of 2020 (Figure 8a).

Figure 8. Comparison of the Reference and Simulated Maps of Bafing for 2020.

Figure 10 shows no significant difference between the simulated and predicted areas. The visual comparison of the simulated 2020 map with the actual map is reasonably similar. In addition, the validation indicators provide values of ROC = 81.6%, κia (kstandard) = 78.34%, Klo = 79.86% and kno = 86.83% (reflecting the overall accuracy of the simulated map), which are considered satisfactory (Chinwendu, 2019; Olofsson et al., 2014; Roland, 2021; Tiné et al., 2019). These results indicate that the MLP- CA model reasonably simulated the LULC map of 2020 and can be used to project future LULCC in Bafing.

Faleme

On the simulated LULC map of 2020 (Figure 9b), the areas occupied by settlement, water, vegetation, agricultural areas and bareground are 11.10%, 1.73%, 30.10%, 5.88% and 51.20%, respectively, compared to 27.88%, 1.09%, 23.22%, 3.54% and 44.28%, respectively, on the reference map of 2020 (Figure 9a). There is a significant difference between the observed and predicted areas. This difference is due to the change in trend described above between the period 1986–2006 and the period 2007–2020. Although the validation indicators provide values of ROC = 75.92%, κia (kstandard) = 0.58%, Klo = 67.14% and kno = 67.14%, (reflecting the overall accuracy of the simulated map) deemed fair (Chinwendu, 2019; Olofsson et al., 2014; Roland, 2021; Tiné et al., 2019), the simulated maps for 2020 cannot be considered satisfactory. Thus, our model cannot be validated for the Faleme. Nevertheless, we have produced the simulated LULC map for 2050 based on the trend (2006–2020) that describes the current development in Faleme.

Figure 9. Comparison of reference and simulated LULC 2020 maps of Faleme.

Projected LULC maps

For Bafing, changes in LULC between 1986 and 2020 were first analyzed, leading to transition potential maps and a probability matrix illustrating the significant LULCC. The transition probability matrix for the classified maps is presented in Table 7. Analysis of the Table 7 results shows that between 1986 and 2020, the settlement, agricultural areas, and bareground classes were the most dynamic. Indeed, the settlement and agricultural areas classes indicate a 45% and 18% probability of not changing to another LULC class, respectively. At the same time, the area of agricultural areas and bareground has a probability of 23.6% and 219% of transforming into settlement. Similarly, the vegetation class also shows stability with a chance of 80.7%. Bareground have a high possibility of turning into vegetation at 22%.

Table 7. Transition probability matrix (%) of the LULC map for the period from 1986 to 2006 of Bafing.

Bafing makana	Settlement	Water	Vegetation	Agriculture area	Bareground
Settlement	0.4515	0.0158	0.2527	0.1149	0.165
Water	0.0491	0.06236	0.2236	0.0014	0.1024
Vegetation	0.0796	0.0358	0.8065	0.0139	0.13
Agriculture area	0.2364	0.0544	0.4893	0.1813	0.3122
Bareground	0.2089	0.0216	0.2157	0.0471	0.6836

Finally, the LULC projection scenarios for Bafing in 2050 were simulated based on the probability matrix obtained using the 1986 and 2020 maps. Figure 10 shows the predicted LULC maps for 2050. The results of this simulation indicate that vegetation will cover the largest area with 49% in 2050, followed by settlement with an area of 19%. Bareground will be the third most dominant LULC class and will cover 22% in 2050. Water and agricultural areas will each cover 4.8% of the area.

Figure 10. Predicted LULC maps 2050 of a) Bafing and b) Faleme.

For Faleme (Figure 10), based on the probability matrix obtained using the 2006 and 2020 maps, the simulation results show an increase in settlement areas, and agricultural areas will be observed by 2050. There is also a decrease in water, vegetation, and bareground.

Discussion

The global environment changes are acknowledged to be fundamentally and significantly influenced by LULC changes. This study aimed to evaluate LULCC over 34 years in two watersheds (Bafing and Faleme) of the SRB and to simulate future changes in LULC in 2050 with the status quo (BAU) assumption. This study used the RF classification algorithm and Landsat images from 1986, 2006 and 2020 for the LULC mapping. RF classification results are very satisfactory with good accuracy. We noticed that some pixels are poorly classified, especially in Faleme. This confirms the results of Zurqani et al. (2018), who suggest that the RF algorithm works better in areas where LULC types are dominated by vegetation. The analysis of the post-classification change detection has reported significant changes in LULC during the study period.

Analysis of changes in Bafing between 1986 and 2020 revealed the expansion of settlement and agricultural areas at the expense of bareground. These results corroborate those of Herrmann et al. (2020), who proved that the intensity of LULC change in settlement and agricultural areas was high in WA. Studies have showed that the increase in settlement and agricultural areas is caused by the increasing population in the WA (Assede et al., 2023). Indeed, the average growth rates are 2.5% and 2.7% for the countries covering our study areas, namely, Guinea and Mali. Tabutin and Schoumaker (2020) observed a high population rate in WA and the resulting socioeconomic impacts (increase in agricultural areas). These results also mirror those of Berihun et al. (2019), who also found that population increase was consistent and positively correlated with the expansion of agricultural areas between 1982 and 2006 in Ethiopia. The results on the increase in vegetation during 1986–2020 are in good agreement with the findings of previous studies conducted in Fouta Djallon Plateau of Bafing in Guinea. Indeed, These results are consistent with the results of UCAD (2019) and Descroix et al. (2020), who found a regreening in Fouta Djallon Plateau of Bafing in Guinea. The increase in vegetation that coincides with population growth suggests that population growth does not always lead to deforestation. Indeed, Descroix et al. (2020) pointed out that the densely populated areas of the Fouta Djallon Plateau of Bafing in Guinea are those where the vegetation cover is not threatened and where the ecological intensification of rural activities has long been established. Therefore, the claim “more people, more trees” proposed in Tanzania by Kabanza et al. (2013) seems valid for Bafing. In addition, it should be noted that several projects have been adopted in the Bafing in Mali to fight biodiversity losses after the construction of the Manantali Dam. Among these projects, we can mention the Bafing Faunal Reserve (Mali), the status of biosphere reserve (Mali), the Natural Resources Management Project (PGRN/World Bank) in the 1990s and 2000s, the Project for the Extension and Strengthening of Protected Area Systems, the Bafing transboundary area protected area project during the period 2010–2015 (Faty, 2017). Another significant result obtained is the increase in water between 1986 and 2020. The observed increase in water can be explained by the recovery of rainfall in this region in the 1990s, after the drought period of 1960–1970. Several authors (Bodian et al., 2020; Diop et al., 2016; Nouaceur et al., 2020) noted a recovery in rainfall in WA. Bodian et al. (2020) studied the recent evolution of hydroclimatic variables in the SRB from 1940 to 2013. These results show a recovery in annual rainfall in the SRB, which improves surface water availability. Recovery of annual flow was reported after the 1990s in Bafing (at Bafing Makana station) (Sane et al., 2017). It can also be attributed to the construction of the Manantali dam in Bafing. The Manantali Dam, built in 1988, has an area of approximately 477 km2 and a capacity of 11,791.8 million m³ (Bader, 2001). It aims to make surface water available and sustainable throughout the year and to satisfy energy production and flow regulation, especially in the context of climate change and variability (Bader et al., 2015).

In Faleme, the analysis of the post-classification change detection revealed that areas occupied by settlement increased from 1986 to 2020. The results also show that agricultural areas, vegetation and water increased significantly between 1986 and 2006. However, there was a decrease in agricultural areas, vegetation, and water between 2006 and 2020. These results can be explained by the artisanal extraction of gold in Faleme during the last decades. Overpopulation in Faleme is strongly linked to population growth and mining activities (industrial and artisanal). Faleme supports the gold resources of eastern Senegal and Mali Artisanal gold mining has increased significantly in recent decades, mainly due to the rising price of gold on the international market and the difficult socioeconomic situation (Bohbot, 2017). Artisanal gold mining is becoming a profession like agriculture and livestock in Faleme (Ministry of Economy and Finance,2018). This activity generates several sources of income in the different localities where it is practised, but it also causes several environmental impacts, including on soils, water resources and deforestation (Bohbot, 2017). The results of this study on Faleme are in good agreement with the conclusions of previous studies conducted in regions where gold panning is practiced. For example, results on the decrease in agricultural areas in Faleme are consistent with those of Atteyoub and Camara (2020) and Ndiaye (2020) who reported that the development of gold panning has led to the gradual abandonment of agricultural practice in this area. Similarly, Doucouré (2015) in his book “Des pierres dans les mortiers et non du maïs! Mutations in the gold-mining villages of southeastern Senegal”, describes the abandonment of agricultural areas in favor of gold panning. The deforestation and reduced water surface quality in Faleme is also consistent with the results of Traoré (2022) at kenieba in Mali and at Nzema in Ghana Kaku et al. (2021).

The predicted LULC map of 2050 was performed using the MLP-MC model embedded in LCM with the BAU scenario. LULCCs prediction involves two different aspects. The first is the amount of change, and the second is the spatial distribution of change. LCM provides the amount of change by comparing initial (1986) and second (2006) LULC maps with the MLP and then predicting future land cover (2020) using the MC transition probability matrix for the future. Model validation is performed by comparing the simulated LULC map (2020) with the classifier LULC map (2020)’. For Bafing, the model has been validated with satisfactory results. The simulation results indicate that the observed and predicted LULC map of 2020 were in good agreement. During the period 2020–2050, the prediction results (based on the past trend (1986–2020) revealed that vegetation would be the dominant LULC, but an increase in agricultural areas, water, and settlement will also be observed in 2050.

For Faleme, the model has not been validated with satisfactory results. The results showed that the observed and simulated LULC maps of 2020 were not in good agreement, and errors for all classes were noted. These results can be explained by the trend shift between the previously mentioned periods (1986–2006) and (2007–2020). The results of the validation of the MLP_MC model on the Faleme showed a limitation of deterministic prediction models such as MLP_MC in LCM. The principle behind MLP- MC projections is that the rates of change observed during the calibration phase will remain unchanged during the simulation period, which in many circumstances, is an incorrect assumption, according to J. Mas et al. (2014). Approaches based on historical tendencies may or may not be effective. The predicted LULC map 2050 was then obtained based on LULC changes from 2006 to 2020. The results indicated that the settlement area increased from 10.15% to 27.88% from 2006 to 2020, and the prediction results confirm that it will continue to expand to 32.58% during 2020–2050, at the cost of a reduction in vegetation (23.22% to 19.83%). Such a loss of vegetation is expected to reach a critical threshold in the coming years, showing the need to develop better spatial planning and adapted sustainable development strategies. These combinations of results suggest that population growth and anthropogenic activities appear to be the primary driver of LULC changes in these two watersheds.

These two watersheds show that while population growth can contribute to increased pressure on land resources and lead to unsustainable land-use practices, it is not the only determinant of these issues. Other factors, such as socioeconomic activities, agricultural practices, appropriate policies and regulations, education and awareness, also play an essential role. The results in Bafing showed that when population growth is accompanied by adopting sustainable land management practices, it can lead to better land and water conservation. The presence of a large dam in Bafing has led to the implementation of virtuous policies favourable to the environment. In Faleme, mining activities, many of which are uncontrolled, are leading to environmental degradation. These divergent development trajectories have different impacts on the water cycle and must be considered in water development policies for the Senegal River basin.

The main limitation of our study is that it did not consider stakeholder involvement in LULCC modelling processes. According to Hewitt et al. (2014), the information of stakeholders on LULCC drivers, reconstruction of timelines of major past events and their perspective on potential future trajectories of land-use change are essential to achieve holistic results in a participatory manner and complement model results. This technique bridges the gap between practitioners’ perspectives and those of technical or policy stakeholders (Thiam et al., 2022). In addition, the use of hybrid prediction models integrating several individual models could improve the prediction and allow the simulation of LULCC to be more realistic (Gaur et al., 2020; Girma et al., 2022; Sankarrao et al., 2021). Sankarrao et al. (2021) compare different LULCC modelling techniques to predict the future LULC by testing MLP-MC, Logistic Regression-Markov Model (LR-MC), Multilayer Perceptron Markov Chain Cellular Automata (MLP_MC_CA) and Logistic Regression Markov Model Cellular Automata (LR_MC_CA) models on Nagavali River Basin (NRB), in Southern India. The results revealed that after combining the MLP_MC model with the Cellular Automata, the model was improved in terms of the Kappa coefficient. In addition, the hybrid model MLP-MC-CA had a better agreement than the other models. Furthermore, the study did not incorporate climate change and variability drivers. Considering these limitations, future research can fill the gaps left by these shortcomings. This work was done to establish scenarios for the construction of hydropower dams and to assess the combined effects of LULC and climate change on the services these dams will provide in these watersheds (Sambou et al., 2023). This document allows the OMVS to develop adequate land and water resource management policies and strategies specific to each watershed, considering the sustainable development goals.

Conclusion

This study analyzed the historical LULCC from 1986 to 2020 with the RF classification and the projected LULC for 2050 by using the MLP-MC model in the Bafing and Faleme in the SRB. The main results revealed that spatial and temporal changes have occurred. During 1986–2020, a significant increase in vegetation, water, agricultural areas and settlement and a decrease in bareground were found in Bafing. The projections of LULC for 2050 show effect for the environment by the increase in vegetation, agricultural areas and settlement. The analysis in Faleme show an increase in settlement, vegetation, agricultural areas, and water, as well as a decrease in bareground between 1986 and 2006. Between 2006 and 2020, settlement increased. However, there was a decline in vegetation, agricultural areas, water, and bareground. By 2050 (based on 2006–2020 trends), land use changes will convert in a direction incompatible with a balanced environment. The analysis of surface state dynamics revealed that population growth and changing anthropogenic (socioeconomic) activities were the main drivers of LULC changes. The changes in LULC are both positive and negative. Based on LULCC, the SRB experienced a trend towards “more people, more trees” for the Bafing and “more people, more deforestation” for the Faleme. These two examples show that population growth, accompanied by adopting sustainable land management practices, can lead to better water and land conservation. These divergent development pathways have different impacts on the water cycle and must be considered in water development policies for the Senegal river basin.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author, M H Astou SAMBOU, upon reasonable request.

Additional information

Funding

This paper is a component of a PhD research financed by the German Federal Ministry of Education and Research (BMBF) and undertaken at the University of Abomey-Calavi, Republic of Benin, on behalf the West African Science Service Center on Climate Change and Adapted Land Use (WASCAL).