Support Vector Machine Algorithm for Mapping Land Cover Dynamics in Senegal, West Africa, Using Earth Observation Data

Abstract

This paper addresses the problem of mapping land cover types in Senegal and recognition of vegetation systems in the Saloum River Delta on the satellite images. Multi-seasonal landscape dynamics were analyzed using Landsat 8-9 OLI/TIRS images from 2015 to 2023. Two image classification methods were compared, and their performance was evaluated in the GRASS GIS software (version 8.4.0, creator: GRASS Development Team, original location: Champaign, Illinois, USA, currently multinational project) by means of unsupervised classification using the k-means clustering algorithm and supervised classification using the Support Vector Machine (SVM) algorithm. The land cover types were identified using machine learning (ML)-based analysis of the spectral reflectance of the multispectral images. The results based on the processed multispectral images indicated a decrease in savannas, an increase in croplands and agricultural lands, a decline in forests, and changes to coastal wetlands, including mangroves with high biodiversity. The practical aim is to describe a novel method of creating land cover maps using RS data for each class and to improve accuracy. We accomplish this by calculating the areas occupied by 10 land cover classes within the target area for six consecutive years. Our results indicate that, in comparing the performance of the algorithms, the SVM classification approach increased the accuracy, with 98% of pixels being stable, which shows qualitative improvements in image classification. This paper contributes to the natural resource management and environmental monitoring of Senegal, West Africa, through advanced cartographic methods applied to remote sensing of Earth observation data.

1. Introduction

1.1. Background

Land cover and land use change are widely known as key elements of landscapes. Maps showing classification of land cover types are key data sources for assessment of landscape dynamics and for evaluating environmental trends. Land cover types (or land use types) refer to the physical components of Earth cover that are physically present and visible on the Earth’s surface. Compared to land use types, which are defined as landscape patches used by people, land cover types typically represent natural physical characteristics of the Earth’s surface in terms of structure, patterns, and components. Characteristics and features of land cover types are visible on satellite images recorded from space. Therefore, remote sensing (RS) data, such as satellite images, are widely used for land cover change analysis using classification maps that show spatio-temporal changes in land cover/land use patterns. Hence, land cover maps are the background geoinformation products necessary for analysts and decision makers. The use of land cover maps enables the monitoring of environmental changes and evaluation of risks to sustainable development. Therefore, the use of land cover maps is applicable in diverse research and development sectors such as governments, civil engineering, and industrial planning [1].

In monitoring landscape dynamics, remote sensing (RS) data are widely used and applied in environmental mapping with the aim of detecting identical landscape patterns in a time series of images and analyzing changes using recognition of landscape patches. In this regard, there is a strong need for machine-based, automated geospatial data processing and image analysis that turns the technical values of pixels in the satellite images into information that provides knowledge and insights for environmental experts and planners. The identification and recognition of land cover types becomes possible using the analysis and synthesis of remote sensing data, such as spaceborne images. Environmental studies often use such products for landscape analysis, for example, images of Landsat missions.

Major types of satellite images for effective interpretation of the Earth’s landscapes include Sentinel 2 Data [2,3], NOAA-AVHRR [4,5,6], SPOT [7,8], and Landsat [9,10], and the combinations thereof [11]. For instance, the recent Global Land Cover 2000 map was implemented using the analysis of the Vegetation Sensor on board the SPOT-4 satellite image [12]. Another example of such products is presented by the time series of annual global maps of land cover types derived from the ESA Sentinel 2 imagery at 10 m resolution [13]. The Landsat scenes used for landscape mapping include data acquired from its various sensors, such as the Multispectral Scanner (MSS), the Thematic Mapper (TM), the Enhanced Thematic Mapper Plus (ETM+), and the recent product called the Operational Land Imager and Thermal Infrared Sensor (OLI/TIRS).

Certainly, although diverse images can be used in various combination as data sources for landscape interpretation and environmental mapping, the general approach of RS data processing relies on using the values of spectral reflectance obtained as digital numbers (DNs) of pixels on the image scene [14]. The effects of reflectance and absorption of various land cover types in different wavelengths enable the discrimination of major land cover types in the multispectral images: water, land, forests (with various types of vegetation), urban zones, and desert sands [15]. Such information can be extracted from variations in brightness, primary colors (red, green, blue), hue, saturation, and intensity of light. In experimental studies, these phenomena are utilized to assess land–water borders and distinguish them from other land cover types, leading to an improvement in the recognition of the Earth’s objects [16]. The information provided by the environmental descriptors can be obtained from the analysis of values of pixels recording spectral reflectance of diverse land cover types. In RS, this can be derived from the combination of different bands of the multispectral imagery, e.g., NIR/Red bands for computing the Normalized Difference Vegetation Index (NDVI) and other vegetation indices [17,18,19,20].

1.2. Problem Formulation

One of the challenging problems in identification of land cover types in the satellite images consists of automatic extraction of the reliable features visible in the image scene. This is true especially in terms of highly heterogeneous landscape patches in a complex mosaic of intermittent land cover types [21,22]. In this regard, image processing and analysis for landscape interpretation are fundamental issues in environmental mapping and thematic cartography. Identification of land cover types in the RS data requires diverse image processing techniques, such as classification, clustering, and segmentation. Classifying mixed terrestrial ecoregions typical for Senegal requires advanced methods of image analysis [23]. Senegal’s landscape is distinctive because it consists of a complex blend of diverse, semi-arid landscapes, such as forest–savanna mosaics, which are periodically filled with Sahelian acacia and West Sudanian savannas with diverse tree species, as well as coastal regions with dominant Guinean mangroves [24,25].

The current practice of geographic information system (GIS)-based supervised image processing involves manually marking different features in a landscape, such as regions of interest (ROIs), singular patches, and landscape corridors, because the extraction of features from processed images is difficult [26]. The human factor in landscape assessment involves some concerns about the repeatability and reliability, which suggest the need for the development of automatic methods for image interpretation and object recognition [27,28,29]. Manual interpretation of land cover types can lead to the misclassification of pixels in the images, which might result in misinterpretation of the selected landscape patches. As mentioned earlier, implementing quantitative data interpretation for prediction and analysis is challenging in African countries where fieldwork data are limited [30]. Therefore, RS-based vegetation mapping and agricultural landscape monitoring are vitally dependent on using satellite data processed with a high level of automation [31,32].

Effective image analysis relies heavily on the type of classifier and the algorithms used for data handling [33]. Another important factor is the quality of imagery for the evaluated region of interest (ROI) in the study area, which includes minimal cloudiness and haze, correct time of image capture and coverage. Finally, image analysis is based on the methodology and approaches used in RS, including data acquisition, preprocessing, description, recognition, classification, segmentation, interpretation, and accuracy analysis. The development of computer-based programming methods has led to the growth of novel image processing methods, which are the primary factors in effectiveness [34,35,36,37,38]. Development of the advanced tools of image processing and related cartographic tasks has resulted in many approaches to utilizing effective methods for automation [39,40,41,42,43,44]. One of these methods is the machine learning (ML)-based Support Vector Machine (SVM) algorithm used for image analysis and synthesis for interpretation of vegetation patterns [45,46,47,48].

In this respect, the research questions formulated in this study are as follows: (1) Can the Support Vector Machine (SVM) ML-based land cover detection method be used with GIS for satellite image analysis? (2) How can the SVM algorithm of ML be employed in geoinformation extracting tasks using scripts? (3) What is the workflow for RS data processing for an effective chain of data handling to perform the necessary tasks during SVM-based image classification? Answering these questions enables us to evaluate the effectiveness of SVM as one of the most powerful ML applications in cartographic tasks and RS data analysis.

1.3. Related Work

Many reports have recently been published illustrating land cover changes in Senegal [49,50,51], effects of climate change on landscape variability and vegetation communities [52,53,54], environmental dynamics, and sustainability [55]. Nevertheless, these approaches do not take into account the benefits of programming scripts for satellite image processing. At the same time, using ML methods increases the automation of mapping and the speed of data processing, ensuring the accuracy and objectivity of land cover classification. Indeed, image processing using ML methods such as SVM partitions the image into structured land patches using automatic analysis of spectral reflectance of the pixels. The obtained values of pixels are then assigned into various classes using principles of machine-based vision and pattern recognition. Through this approach, it is easy to visualize and estimate changes in land cover types using images from different years as part of a time series analysis. Additionally, it overcomes the problems with conventional software programs used for image processing, which may include drawbacks in data processing.

A drought period in Africa in the 1970s and 1980s resulted in climate changes that include such consequences as decreased rainfall and precipitation, raised temperatures, and increased evaporation [56,57]. These climate-related processes triggered land cover changes in natural vegetation patterns of Sahel ecoregions that affected cultivated areas and croplands, increased spots of bare soil, and led to changes in surface water [58,59]. To this point, one of the priorities of image processing methods for environmental landscape analysis is to support land cover monitoring through the interpretation of land cover types using time series analysis. Regular monitoring of the Earth’s landscapes enables us to detect environmental changes, land and soil degradation, and variations in vegetation cover [60]. Another essential application of image analysis is monitoring coastal zones to evaluate hydrological changes, to evaluate the extent of river flooding in the delta, to detect soil erosion [61,62], and to monitor the degradation of mangroves and coastal ecosystems [63,64].

1.4. Objective and Motivation

Inspired by the existing examples of the ML methods used to analyze multispectral satellite images in environmental applications, this study integrates advanced cartographic scripting methods with applied techniques of SVM image processing for monitoring landscapes in Senegal; see Figure 1.

The environmental analysis was performed using a series of satellite images evaluated for the period from 2015 to 2023. To this end, we explore the classification approach of SVM, which was employed for RS data processing by the scripting techniques of the Geographic Resources Analysis Support System (GRASS) GIS. The SVM classifier is applied as a widely implemented supervised learning approach. Other advanced deep learning (DL) methods that have been developed recently include such methods as autoencoders, a special type of neural network that operates with dimensional latent representation, and vision transformers, which differentially weigh the significance of each part of the image. Nevertheless, SVM is still a sufficient and effective approach for many remote sensing (RS) and GIS tasks. Given the availability of SVM in GRASS GIS, this method was employed for the image classification.

The main goal of this study is to map and evaluate changes in diverse land cover types in West Senegal, which include both inner regions and coastal mangrove forests in the delta of the Saloum River; see Figure 1. To achieve this goal, the practical objective is to classify land cover types in the coastal region of West Senegal using a series of Landsat 8-9 OLI/TIRS multispectral satellite images processed using ML methods, including SVM as an advanced method of automated image classification. Contrary to the existing GIS, the programming approach to satellite image processing has a high automation in data handling. Additionally, it is a very fast and robust approach for classifying and comparing satellite images. This is achieved by a combination of the embedded GRASS GIS modules is used separately for computing vegetation indices and plotting the classification maps.

According to recent studies, mangrove habitats are drastically disappearing. This is because these ecosystems are being impacted by anthropogenic activity, climate change, and local effects of salinity in the soil [65,66]. By evaluating gains and losses in landscape patches, RS data can be spatially analyzed to understand the dynamics of these distinct wetland ecosystems. Our comprehension of the changes in Senegalese landscapes is aided by satellite image analysis, which shows how these landscapes respond to the effects of West Africa’s changing climate, including rising temperatures, evaporation, and falling precipitation.

The goal of using the GRASS GIS scripts for RS data analysis is to quantitatively evaluate changes in land cover types in Senegal as a result of climate fluctuations during a recent eight-year period. The approach to achieve this goal is based on using automated unsupervised classification and supervised classification using the ML method of SVM. In this way, the maximal likelihood and SVM classification algorithms present an effective means by which the Landsat scenes can be processed automatically. The RS data obtained from the United States Geological Survey (USGS) repository were used for environmental analysis and monitoring to evaluate and visualize variations in landscape patches. This analysis is technically based on the ML-based estimation of spectral reflectance on the satellite images, which supports environmental monitoring of the West African landscapes responding to climate and anthropogenic effects.

The main purpose of this paper is to provide an advanced approach using ML to RS data analysis and processing. Such an approach can evaluate the nonlinear behavior of the landscapes in West Africa over time, which indicates the development of climate settings and the related response of vegetation patterns. To this end, we propose a script-based classification and image enhancement technique using GRASS GIS software to analyze land cover changes in the coastal region of West Senegal. This technique aims to address the main challenges associated with traditional image processing, such as structured noise in the classified scenes, poor quality of the images, the misinterpretation of pixels, and the low speed of data handling.

The ML-based image analysis is based on using the RS data, which serve as a cartographic basis to support optimized decisions made by environmental policy makers and planners. Satellite images processed by the machine-based methods allow for the identification of vulnerable landscapes and enable the monitoring of degraded lands in Senegal. In order to do this, a series of the multispectral satellite images from Landsat with recent OLI/TIRS sensors taken on different years was processed with various modules of GRASS GIS. This time series data analysis enabled us to perform an environmental analysis for monitoring landscape dynamics in Senegal.

Importantly, the study presented here is not intended to suggest that scripting methods replace the role of conventional GIS. In contrast, the programming codes supplement the traditional cartographic methods used in the RS software. However, the integration of a scripting workflow with image analysis and cartographic tasks provides an effective approach by which the analysis of land cover changes can be visualized in an automatic way. In light of this, the main contributions of this paper are as follows:

• A combination of the GRASS GIS modules ‘i.group’, ‘i.maxlik’, and ‘i.cluster’ are proposed for the unsupervised image classification. The land cover types are classified based on the difference in spectral reflectance of the pixels in each of the images.
• A set of GRASS GIS modules, including ‘r.learn.predict’, ‘r.learn.train’, and ‘r.random’, is used for supervised ML-based image classification using an SVM modeling algorithm.
• The land cover types in coastal Senegal are defined as groups of pixels according to the structural similarity between the landscape patches. Orientation, location, and frequency of mangrove fields in the Saloum River Delta are estimated from the analysis of landscape patches in the coastal area.
• To balance the robustness and accuracy, a coarse-to-fine strategy is proposed using the interpretation of landscape patches according to the FAO scheme of land cover types. The data for the whole country were downscaled to the ROI of the selected study area, with images covering the region of Cape Verde Peninsula and the surrounding territories.
• The proposed image processing and ML-based classification algorithms utilizing the SVM techniques of GRASS GIS outperform the existing traditional software with graphical user interface (GUI) approaches in terms of repeatability of scripts and automation through computer vision. This methodological workflow is useful for future similar studies of land cover type analysis in Senegal or surrounding areas of West Africa.

2. Materials and Methods

To increase the speed and accuracy of image processing, the GRASS GIS software was applied for advanced image processing due to its extended functionalities [67]. Generic Mapping Tools (GMT) Version 6 [68] was used as a cartographic tool for topography, and RS tasks are based on scripting techniques. The main advantages of scripts include their high speed of data processing and repeatability [69,70,71].

2.1. Data

We use six Landsat 8-9 OLI/TIRS images for evaluation of land cover changes in western Senegal from 2015 to 2023. Major metadata are summarized in

Table 1. Identifiers (IDs) of the six multispectral Landsat 8-9 OLI/TIRS images.

Date	Landsat Product Identifier L1	Landsat Product Identifier L2	Scene ID
25 February 2015	LC08_L1TP_205050_20150225_20200909_02_T1	LC08_L2SP_205050_20150225_20200909_02_T1	LC82050502015056LGN01
17 February 2018	LC08_L1TP_205050_20180217_20200902_02_T1	LC08_L2SP_205050_20180217_20200902_02_T1	LC82050502018048LGN00
7 February 2020	LC08_L1TP_205050_20200207_20200823_02_T1	LC08_L2SP_205050_20200207_20200823_02_T1	LC82050502020038LGN00
25 February 2021	LC08_L1TP_205050_20210225_20210304_02_T1	LC08_L2SP_205050_20210225_20210304_02_T1	LC82050502021056LGN00
20 February 2022	LC09_L1TP_205050_20220220_20230427_02_T1	LC09_L2SP_205050_20220220_20230427_02_T1	LC92050502022051LGN01
22 January 2023	LC09_L1TP_205050_20230122_20230313_02_T1	LC09_L2SP_205050_20230122_20230313_02_T1	LC92050502023022LGN01

.

The images were downloaded from the United States Geological Survey (USGS) EarthExplorer repository; see Figure 2. The data contain scenes taken during the winter period (January/February). The selection of the winter period is explained by the fact that images taken during winter enable us to better detect plant phenology in a semi-arid climate. In contrast, those collected during summer are not suitable for plant detection due to the high level of droughts in the desert areas, especially in the inner regions of the country. Additionally, the images were collected with minimized cloudiness and haze below 10%. Hence, the best quality and suitability of images was in the winter period, during the months of January and February. All Landsat images are multispectral and have a 30 m resolution in visible spectral bands. The images are projected in Universal Transverse Mercator (UTM) Zone 28 for western Senegal. The Landsat images used in this study are presented in Figure 3.

It is well-known that the high-resolution commercial satellite images (e.g., SPOT, Pleiades) are expensive due to the expenses related to the launch of the satellite and related operational costs. Therefore, such scenes are not always easily accessible as a data source. Hence, the use of the open source Landsat 8-9 OLI/TIRS is crucial, and the value of these products to environmental monitoring is high. This is especially true for replication of this study in future similar work. The aim of the multispectral Landsat images was to examine the impact of both anthropogenic and climatic factors on the process of land cover change in western Senegal. This was achieved by analyzing and processing the spectral reflectance characteristics of the pixels obtained from the RS data.

The images were selected for different years (2015 to 2023) for comparability. Using RS techniques, the environmental characteristics of the landscapes were evaluated in order to analyze how changes in climate are affecting the types of land cover and, in particular, the distribution of mangroves along the Saloum River Delta. The metadata for the images were evaluated for each scene: four images were selected from the Landsat-8 OLI-TIRS sensor, and two recent scenes were selected from the Landsat-9 OLI-TIRS sensor instruments.

2.2. Methodology

The methodology of this study is based on scripting software for data processing. Technically, we propose a novel mapping method using the SVM algorithm in GRASS GIS for improving the image classification task, and we compare it with the conventional method of unsupervised classification using k-means clustering. Major programming scripts are shown in Appendix B for technical reference. The full codes have been placed in a GitHub repository, including ML techniques of SVM). The advantage of the GRASS GIS approach to RS data processing consists of adjusted modules enabling us to analyze spectral properties of the satellite images. Additionally, the GMT was applied for cartographic mapping. This paper presents a series of scripting experiments that utilize both software applications. Methods of applied programming enabled us to evaluate recent dynamics of land cover types in Senegal.

The advanced script-based cartographic methods of monitoring and mapping land cover types are built on a growing demand for time- and cost-effective approaches in cartographic workflow and RS data analysis. Their use enables us to save time and resources while processing large geospatial datasets using satellite images obtained from open sources. Hence, three major script-based tool sets were used in this study for geospatial data analysis and environmental monitoring in West Africa:

• The GMT script was used for plotting the topography of the country; see Listing A1.
• An R script was used for plotting the methodological flowchart; see Listing A2.
• GRASS GIS scripts were used for image processing; see Listings A3 and A4.

The main idea was to apply a single workflow framework for both the image processing and the cartographic mapping tasks using scripts run from the console. This enabled us to reveal the advantages of the technical performance of automated geospatial data processing. More specifically, using a programming approach that includes various programming suites from the script-based software presents a novel workflow for geospatial data processing. This method demonstrates how adding scripts expedites the process of image analysis and cartographic plotting while also providing an overview of the study methods. As mentioned earlier [72], the automation of the cartographic procedure and the workflow’s repeatability in future research are the primary benefits of scripts. Moreover, scripting enables us to save significant time due to automation. We were able to identify changes in the analysis of landscapes of western Senegal for the assessment of environmental impacts from climate change by using the information obtained from the automated image analysis and interpretation. This was done by evaluating a series of images and analyzing RS data.

The workflow of the data processing and major methodological steps is shown in Figure 4, with the goal of performing image classification and interpretation in order to detect changes in land cover types. The data processing was technically implemented using GRASS GIS software [73] and run on a Mac machine with Apple M1 chip and the MacOS operating system. After preprocessing and data import, the images were grouped to select the necessary multispectral raster bands for image processing. This was performed by the ’i.group’ module of GRASS GIS [74]. Afterwards, the images were clustered using unsupervised classification. This was carried out utilizing the module ’i.cluster’, which automatically assigns pixels into groups according to their spectral reflectance. This method of image processing provided enough data to group the pixels into land cover types and to evaluate the values of pixels in the actual scenes for comparison. The average computation time for image classification and enhancement per Landsat scene is about 2.5 s.

The matching performance of landscape patches over this period is evaluated using GRASS GIS functionality as a sequence of modules utilized for processing each scene separately using a script. The images were read into the GRASS GIS environment using the ’r.import’ module; the list of the available bands was checked using the ’g.list rast’ command. The extent of the ROI (UTM Zone 28) and the resolution were set to the region of the current image, which was detected automatically along with geospatial data such as coordinates, location, and projection. The false color composites were plotted to distinguish the area of vegetation in the delta of the Saloum River (colored in red shadows), land (colored by beige and light brownish shadows), and water areas (colored black for seawater and navy blue for the Saloum River); see Figure 5.

The ML method of satellite image processing involves using the SVM algorithm for data partition and classification. To perform these steps with fixed ML parameters, we utilize the GRASS GIS modules ’r.random’ for generating a training dataset from pixels obtained from the land cover classification, the ’r.learn.train’ module for training a Support Vector Classification (SVC) model, and the ’r.learn.predict’ module, which predicts the pixel’s assignment to the target classes. We represent the image dataset as a time series of scenes from 2015 to 2023 represented as color composites from the multispectral bands of Landsat. We denote the number of estimators as 500 in the model and the number of seed points as 100 pixels, with a total of 1000 evaluated pixels (’npoints’ in the model’s syntax). The training model name is explicitly defined as ’SVC’, i.e., Support Vector Classification, as it is the algorithm embedded in the GRASS GIS software.

The standard requirements for a training dataset in the ML approach that includes the SVM algorithm include a supervised classifier to derive an accurate and precise description of the spectral properties of each land cover type. Therefore, the training data were derived from the previous raster dataset in TIFF format as samples that support vectors through data partition. The target pixels were selected as representative cells located on the edge of the class distribution in feature space. Using GRASS GIS syntax, generating the training pixels from the land cover classification was performed using the following code: ‘r.random input=L_2015_clusters seed=100 npoints=1000 raster=training_pixels’. Here, the number of pixels was selected as 1000 for a larger and more representative dataset, and the seed for training the data was set to 100.

Afterwards, the contrasting regions that have distinguished spectral reflectance values (water, savanna, forest, croplands, and urban spaces) were used for identification of similar classes in the evaluated periods. Then, the land cover classes were identified using the SVM algorithm with the modules ‘r.learn.train’, which trains an SVM model, and ‘r.learn.predict’, which performs prediction of pixels’ association with each of the 10 target land cover classes based on the spectral reflectance of pixels in the images.

3. Results and Discussion

3.1. Interpretation of Key Findings

In this section, the ability of the two algorithms—unsupervised classification by k-means and supervised classification by SVM—to detect land cover changes in the satellite images is evaluated. Specifically, we assessed the dynamics in the distinct habitats of Senegal ranging from arid landscapes with savanna to humid coastal conditions with mangroves. The data for the land cover types in Senegal were visualized in the QGIS software using vector layers in the Environmental Systems Research Institute (ESRI) format; see Figure 6.

According to the Food and Agriculture Organization (FAO)/United Nations Environment Programme (UNEP) classification scheme, the land cover types in Senegal include 10 major classes. These were adopted for the target extent of the ROI through generalization at the country level using the Land Cover Classification System (LCCS) of FAO/UNEP:

• Mosaic grassland or shrubland/cropland;
• Mosaic vegetation/cropland;
• Mangrove wetlands;
• Rainfed cropland;
• Consolidated bare areas;
• Mosaic grassland/forest or shrubland;
• Closed broadleaved forest or shrubland permanently flooded—brackish water;
• Salt hardpans;
• Post-flooding or irrigated or aquatic cropland;
• Water bodies.

The quantitative estimations of changes in land cover types over the evaluated period (2015 to 2023) are summarized in the tables in Appendix C. They report the computed changes in land cover classes for each processed image. The analysis of these values shows that the northern and southern coastal ecosystems of western Senegal have a relatively small degree of change over the estimated period. In contrast, in the arid north, areas associated with forests and mosaic types of vegetation demonstrate higher magnitudes of change, suggesting a higher level of instability compared to the estuaries in the south. The classification of the land cover types was adapted to the larger extent of the ROI with details for the coastal area, including mangrove forests. The territory of Senegal comprises eight major biological zones with 11 different forms of land use. Of these, major types include the savanna, cultivated lands, forests, and steppes as the dominating types of land cover in the semi-arid zones [75].

A special focus of coastal Senegalese landscapes is mangrove swampland, a crucial habitat in the west of the country. Mangroves provide a source of natural resources [76] and a habitat for wildlife species [77]. At the same time, diverse factors have strongly correlated with mangrove losses. These include climate-related processes such as changes in temperature and precipitation, flooding, and increased salinization of waters, as well as human-related factors such as increases in intensive urbanization [78] and active agriculture practices [79]. Overall, the analysis of land cover types over the coastal area shows the variability in mangroves (bright cyan in Figure 6). Coastal dynamics are explained by high interannual fluctuations in the delta of the Saloum River: the upper tidal zone is flooded during high tide, while during the drought period, it remains drained, with high evaporation resulting in soil salinization.

3.2. Clustering-Based Classification

The overview of Figure 7, on which are overlaid the 2015–2023 Landsat OLI/TIRS time series and the land cover types for each case, shows the diversity of habitats and the intensity of changes detected in the images classified by the k-means algorithm. The analysis of the classified images based on clustering shows a trend in land cover dynamics of the coastal region of Senegal over Dakar and its surroundings.

The results demonstrate that the dominating land cover type of Senegal remains savanna (73%), although the areas covered by savanna showed a slight decrease since 2015. In contrast, the landscapes covered by croplands and agricultural areas expanded to 22%. Furthermore, the decline in the area covered by forests and mangroves is notable. Nevertheless, in general, the dominating land cover types in Senegal remain savannas, woodlands, and forests, covering over 75% of the country, a result aligned with similar studies [24].

The maps of land cover change based on the classified imagery display different spatial patterns. While forest and savanna in the subplots have a significantly higher magnitude of values than the background, this is not the case for the bushland and grasslands, where the difference between the values inside and outside is minimal. The croplands demonstrate moderate changes that indicate the measures taken for sustainable management in Senegal. Overall, landscape heterogeneity can be observed in comparing the processed images. The southern sector of Senegalese vegetation is characterized by the Sudano-Guinean forest savanna and the Guinean forests [80].

The most pronounced variations in land cover types in western Senegal detected on the processed and classified satellite images reflect the relationships and links between changing vegetation and soil characteristics that are subject to climatic factors, moisture level, salinity, and organic matter, as noted earlier [81]. Moreover, the pattern of changes in the agricultural land cover types over time correlate with seasonal fluctuations in the crop–fallow cycle. Agricultural plantations and irrigated lands are distributed in sporadic settlements along the populated and semi-urban areas. Earlier studies also noted [82] that climate variability is the dominant factor that leads to increasing the actual evaporation and, as a result, to land salinization in the basin of the Senegal River.

3.3. Support Vector Machine (SVM)-Based Classification

The SVM-based classification of the dynamics of land cover types in Senegal based on the Landsat OLI/TIRS scenes (2015–2023) resulted in the generation of the thematic maps of land cover types shown in Figure 8, in which the most important classes are identified as follows: the rainfed cropland class is depicted in aquamarine; water areas are shown in blue; mosaic vegetation is colored bright red; and grassland, forest, and shrubland areas are shown in orange. The numerical results obtained for the different images using classification models of GRASS GIS are summarized in Appendix C. Generally, the classification of the Landsat scenes using the SVM algorithm applied to the reflective multispectral bands of the image (channels 1 to 7) provided excellent results for each kernel of the 10 categorized classes.

The aim of the SVM application is to improve the accuracy of recognition of land cover types and to increase the quality of mapping. The feasibility of land cover change analysis depends on the quality of cartographic techniques when processing RS data. Erroneously recognizing trends may arise from the use of land cover regions derived from unadjusted classifiers that exhibit imbalanced misclassification between distinct groups. Therefore, in our work, we used the effective approach of the SVM algorithm proposed by GRASS GIS, which presents a ML approach for image classification. The scripts used for SVM-based image processing are presented in Appendix B.

The detected landscape dynamics evaluated using a time series of the satellite images indicate the cumulative effects of anthropogenic activities and climate change in the semi-arid region of West Africa. This involves such processes as variability of precipitation, increase in temperature, land erosion, and wildfire. These factors control the distribution of croplands, natural vegetation systems, and coastal mangrove colonies in western Senegal. The main triggers for changing types of vegetation and landscape patterns in the Senegalese Sahel include climatic factors and anthropogenic activities, which have resulted in the decline and local extinction of woody species [83].

Among other land cover types, the relics of the Guinean forest are distributed in the northern part of Senegal as a narrow band of landscapes stretching parallel to the Atlantic coast. In the southern and eastern regions off Dakar and around the region of Thiès, there are many species that are abundant in the north of the forest and represented by the scattered trees in the Sudano-Sahelian savanna.

To the east of Dakar, the northern silvopastoral zone of Senegal presents a vast silvopastoral zone characterized around the Ferlo region, which is notable for its extreme aridity, dominated by land cover types of shrubby savanna and steppe vegetation [84]. The enclaves of hygrophilous vegetation surrounding the coastal wetlands are located to the north of Dakar, with mangroves as the dominant land cover type. Additionally, this region includes the occasional groves of the oil palm as spontaneous landscape features of the Guinean forests.

3.4. Accuracy Analysis

Figure 9 shows the results of the accuracy assessment. The correctness of the assignment of pixels to the target classes was evaluated based on the pixel confidence levels with rejection probability values. Using the algorithms embedded in GRASS GIS, six classified Landsat images were evaluated against the probability of the identified pixels being classified into the correct land cover class. In this way, possible misclassification that may arise due to the similar spectral reflectance values in land cover types was assessed. It can be seen that there is a good linear correlation between the contours of the land cover classes and the accuracy of the classification of pixels. The estimation of the values of spectral reflectance indicates a general assignment to a given land cover type. The results are similar for the classified Landsat images processed by means of scripts and correspond to the different land cover types, such as coastal areas, water, savanna, Guinean forests, and agricultural lands.

3.5. Implications and Discussion

Land cover types in Senegal present fluctuations in vegetation patterns as a result of long-term landscape evolution in West Africa [85,86,87,88]. The cumulative effects of climate, environmental, and anthropogenic factors led to a decrease in mangrove colonies, which decline in high-salinity waters [89]. This is also supported by previous studies that report the retreat of the mangrove in the advance of the tannes [90]. At the same time, the trend in rainfall recovery in West Senegal since the 1990s [65] has resulted in the rehabilitation of mangroves, which restored their colonies accordingly, as also visible on the images. This corresponds with the results of the existing reports on the location and distribution of Rhizophora mangle and Avicennia germinans mangrove species in western Senegal [91,92]. Earlier works also noted the increase in mangrove trees in the estuaries of Senegal [93]. In addition to the coastal ecosystems, the areas occupied by crop production follow the same trend in changes, possibly due to vegetation removal associated with plot preparation on agricultural land.

Human-induced activities such as agricultural, demographic, and socioeconomic changes have affected the environmental settings and habitats of Senegal. Although the human impacts on the region were negligible for decades due to the relatively low population density of Senegal, recently, a natural increase in the population has accelerated the exploitation of natural and mineral resources. As a result, this has triggered land degradation in the coastal areas. Consequently, select landscapes of Senegal are subject to the deterioration of valuable habitats, fragmentation of patches, and land cover changes. Major human-induced factors in Senegal are related to dynamic agribusiness in terms of agricultural commodities, which increases the areas of irrigated vegetation and croplands [94]. Although such activities are vital for supporting the local population, the environmental consequences on Senegalese landscapes include changed land cover types. Additionally, land degradation in the savanna landscapes correlates with trends in land cover changes and a significant decrease in woodlands [95].

To summarize, this paper reports the results obtained from experimental investigations into land cover types in Senegal analyzed using a ML approach and the SVM algorithm, compared and evaluated against the traditional unsupervised classification method of k-means clustering. The performance of SVM demonstrated reliable results in image processing, outperforming clustering in terms of accuracy and speed of data processing. To this end, a time series of Landsat scenes was processed using GRASS GIS software. The image processing was designed using a set of modules processed by scripts for automation. Interactive processing and scripts adjusted for each scene provided an effective way to process satellite images in workflow chain. Such automation enabled us to represent complex interactions within the landscapes in a short-term perspective by accurate ML-based mapping using a comparison of the SVM and k-means methods. In this way, this study has presented an advanced data-driven method of image analysis for automatically detecting changes in land cover types in a selected ROI of Senegal.

4. Conclusions

Mapping land cover types, discriminating mangroves, and monitoring savanna ecosystems are essential for proper land management in Senegal. In this study, the potential for employing the SVM algorithm classifier in conjunction with Landsat OLI/TIRS multispectral satellite images was evaluated to map areas of land cover in the environment of Senegal. In an effort to drive research in this field, this paper proposed a series of image classification experiments using the GRASS GIS scripting approach, which served as a means of ranking landscapes. Using processed satellite images, land cover development was evaluated in the condition of the unstable environmental setting of western Senegal, which has been affected by climate change and human activities. The practical aim was to evaluate the variations in land cover types in the western segment of Senegal around the Cape Verde Peninsula and Dakar and its surroundings during 2015–2023. The presented results confirm that RS data can be effectively used to evaluate the variations in landscapes in West Africa.

To evaluate the performance of GRASS GIS in image processing, the multispectral bands of six images were analyzed for matching landscape comparisons for different years, and the classification accuracy is reported as rejection probability for pixels in each Landsat scene. Hence, the advantage of the use of the RS data for environmental monitoring is that they provide an efficient way to map and visualize the remotely located regions that are otherwise difficult to reach, such as West Africa. The analysis of the extent and distribution of the selected landscapes in a series of satellite images and the decrease in minor patches enabled us to detect environmental landscape dynamics in a selected ROI. We also provided remarks on the factors affecting changes in land cover types based on the analysis of the processed images compared to previous case studies in the published literature. The major driving factors include the related environmental climate processes and human activities that resulted in land cover changes in Senegal. Such notes can be used in similar works with a focus on environmental changes, and land cover type changes in Senegal can continue to be investigated.