Machine Learning Methods of Remote Sensing Data Processing for Mapping Salt Pan Crust Dynamics in Sebkha de Ndrhamcha, Mauritania

The study area is located in the coastal region of West Mauritania, which includes the salt pan of the Sebkha de Ndrhamcha 1. Physical processes affecting the formation of sebkhas include the sedimentation in coastal, fluvial and lacustrine environments, as well as aeolian-dominated conditions (the strength and direction of winds). For instance, the Atlantic continental margin to the west of Sahara presents one of the largest dust sources on Earth and is strongly influenced by mass wasting processes (Henrich et al., 2010). Huge aeolian dune fields affect the sedimentary regime on the continental slope and surrounding sebkhas.

The prevalence of one or more of these settings favours formation of the salt pans and results in different level of their salinity (Tari et al., 2017). Thus, marine coastal sabkhas are more associated with the processes of evaporation and accumulation of marine deposits, while continental playas or inter-dune sabkhas are related to barchan formation and dune mobility (Handford, 1981; Ould Ahmedou et al., 2007). Different salinity of salt lake sediments enables to distinguish the north-west Africa hypersaline chotts with saline levels surpassing that of ocean water due to high concentrations of brines (Lemenkova, 2023a) and non-vegetated sebkhas with moderate salinity (Lemenkova, 2023b) from vegetated sebkhas with a relatively low salinity that ranges from 2 to 42 g NaCl per litre (Soulié-Märsche, 2008).

The geological setting of western Mauritania is presented by a series of Quaternary marine deposits formed in the gulfs of Senegalo–Mauritanian Tertiary–Cretaceous basin (Figure 2). During the uplift of the basin in Miocene, this area had undergone four successive transgressions and three regressions, which resulted in the accumulated continental formations covering marine deposits with sandy-glauconitic facies and diatomite layers (Barry, 2003). Several periods of eustatic movements of the world’s oceans formed the current Quaternary shape of the coastal zone in West Mauritania, with lakes and lagoons becoming sebkhas and bays becoming lagoons during the alternating periods (Einsele et al., 1974b; Elouard and Faure, 1972).

Figure 1.

Topographic map of Mauritania showing the extent of the study area of the Sebkha de Ndrhamcha salt pan (rotated yellow square). Data: General Bathymetric Chart of the Oceans (GEBCO)/Shuttle Radar Topography Mission (SRTM). Software: Generic Mapping Tools (GMT).

Map source: author.

Figure 2.

Surficial geology and lithologic units of Mauritania. Data: USGS, GEBCO. Software: QGIS. Map source: author.

Successive positive epeirogenic episodes resulted in reduced size of the marine gulfs off Nouakchott since the early Pleistocene to Holocene. The Tafaritian sandstones of the Aaiun-Tarfaya Basin belong to the continental environments, such as the lacustrine and sebkha deposits formed under shallow or aerial conditions of Saharan–Sahel environments. Subsequently, the former large marine Tafaritian embayment with uplifted sediments was replaced by the sparse lakes and sebkhas (Giresse et al., 2000, 1988). Currently, the mineral content of coastal sabkhas consists of a combination of terrigenous clastics, carbonate–sulphate (or gypsum) minerals, soluble salts (e.g. halites) and similar types of sediments (Kocurek et al.,1991). Such specific mineral content enables discrimination of these sediments on the satellite images using spectral reflectance of the saline minerals which differ by colours and brightness from the surrounding areas and sands. The lacustrine deposits discovered in other regions of Sahara gave rise to the notion of ‘Saharian Seas’ and current lakes as remnants of the endoreic depressions reflecting past geological processes (Lemenkova, 2023c; Plaziat, 1991).

The origin of the Mauritanian sabkhas is strongly related to the effects of climatic and geological factors. The diversity of geological setting of Mauritania geological history is characterised by the great variety of rocks and geological structures. Among them, the Reguibat Shield (500 by 2,000 km, stretching SW to NE) forms the northern part of the West African Craton and belt of Proterozoic rocks (Archean, Paleoproterozoic to Neoproterozoic ages) (Figure 3). The plate tectonic movements formed the West African Shield, which borders the Taoudeni basin (Hamoud et al., 2021). During the Holocene, in the area located on West African Shield and Ougarta Uplift (Figure 3) prevailed wet climate with repetitive marine transgressions and climatic oscillations (succession of humid/arid phases in the region of modern Nouakchott) (Einsele et al.,1974a). Such processes are now reflected in the Holocene stratigraphy and geomorphological landforms (Wissmann, 1982). This is also revealed in the palynological sequences according to stratigraphical and pollen analysis (Medus, 1987). Besides, the fluctuations in sea water inundation resulted in continuous calcification of the evaporite deposits and marine sulphates that increased the salinity of the area (Manning-Berg et al., 2024).

Figure 3.

Geological provinces around Mauritania. Data: USGS, GEBCO. Software: QGIS. Map source: author.

The climate factors are expressed in the temperature, evaporation and drought in the region, while physical processes determine the morphology of sebkhas and the sedimentary response to those processes. Moreover, the effects of climate-oceanological processes on the coastal habitats can be illustrated through the seasonal upwelling, which affects the coastal productivity and distribution of nutrients for aquatic habitats (Vázquez et al., 2023). Such cyclic circulation of water layers has a great ecological and socio-economic importance for the coastal region of Mauritania.

The distribution of minerals, geomorphic landforms and types of soil strongly affect the overall mosaic of habitats and land cover types; they also strongly affect the agriculture resources and crop management (Quebedeaux and Parks, 1984). For instance, in coastal areas, tidal flats and shallow waters favour the distribution of aquatic and regularly flooded vegetation (Wijnsma et al., 1999), while sandy desert regions are mostly covered by sparse herbaceous vegetation. The general map of land cover types and their distribution over Mauritania is presented in Figure 4.

Figure 4.

Land cover types in Mauritania. Data: FAO, OpenStreetMaps. Software: QGIS. Map source: author.

Similar to other regions of Sahara with salt lakes (e.g. (Ali and Abd Ellah, 2023)), the formation of salt pans in Mauritania is related to salt tectonics and accumulation of evaporite deposits. Besides strong climatic factors, the impact of geological setting on the formation of salt pans is explained by rift geometries that are defined by horsts and tilted fault blocks and generate base-salt flat topography, which affects salt accumulation, flow direction, diapirism and deformation of overburden layers (M. Pichel et al., 2019). The distribution of salt minerals is concentrated in desert areas with highly arid climate where evaporation exceeds precipitation. Therefore, the formation of salt lakes, salt pans, chotts and sebkhas in Africa is mostly located in northern regions close to Sahara (van Asten et al., 2003; Álvaro, 2012). The crust of salt deposits and salty compacted sands creates natural resources for the country where salt is being produced for market and industrial purposes.

Current environmental problems in the region include risks of flooding in Nouakchott, which are due to a combination of both natural factors (wave actions and wind strength) (Semega,2008) and human activities (infrastructure built along the coasts and fragility of the barrier beach) (Ould Sidi Cheikh et al., 2007; Senhoury et al., 2016). For instance, the semi-arid climate of Mauritania creates additional constraints for agricultural activities, which results in construction of large-scale irrigation schemes aimed at increasing yield and irrigation intensity (Borgia et al., 2012). Nevertheless, unsustainable crop management is a threat to ecological sustainability of coastal areas and contributes to the salinisation of soils. Lack of effective drainage system for rain and marine waters and a sewage sanitation system increases the problem of salt accumulation along the coasts of Mauritania and creates additional factors for soil salinisation around Nouakchott.

Flat topographic landforms of the Mauritanian coasts along the Atlantic Ocean, such as the extent of the Aftout Es Sahli (a geomorphological depression running parallel to the coastline) and Sebkha N’Dramcha (a salt pan continuing the Senegalese–Mauritanian sedimentary basin to the north off Nouakchott), create additional risks for the occasional inundations and floods (Vermeer, 2010). Sea water intrusion contributes to the increase of salinity in the coastal areas and supports formation of sebkhas and accumulation of salt in the submerged landscape patches. Another notable characteristic of the Mauritanian coasts is the upwelling tidal system, which is formed by the effects of the coastal slope topography, ocean currents and winds. As a result, the coastal region of Mauritania is characterised by almost permanent upwelling with maximal intensity in boreal winter and spring (Versteegh et al., 2022).

The cumulative effects of human activities and climate processes, briefly discussed above, lead to the decline of groundwater resources in Mauritania, which is especially notable for its semi-arid areas (Mohamed et al., 2017). All these factors lead to the crucial problems of water deficits and their associated shortages in Mauritania, such as the need for water desalination and the problem of water supply (Bayod Rújula and Dia, 2010), and contribute to the formation of the salt pan Sebkha de Ndrhamcha.

The data used in this study include six satellite multispectral Landsat images by National Aeronautics and Space Administration (NASA), whose main properties are summarised in Table 1.

The input raster scenes were based on the six Landsat images. The imagery data of the Landsat 8-9 Operational Land Imager and Thermal Infrared Sensor (OLI/TIRS) satellites were used in this study because they offer systematic coverage of the study with minimal cloudiness. Besides, technical characteristics of the radiometric and geometric features of the scenes are suitable for environmental monitoring in short-term period (2014–2023) using pixel-based methods of image processing. Such tools are applicable in GRASS GIS which includes a wide range of automated algorithms of image classification in its libraries.

Image analysis and classification of the Landsat 8-9 OLI/TIRS satellite images using pixel-based clustering method was done by k-means and an unsupervised method, which is included in the maximum likelihood discriminant analysis classifier by GRASS GIS. Pixel-based image classification techniques, such as k-means, are becoming more and more acknowledged as a useful cartography method for RS data processing to take advantage of the vast amounts of spaceborne data. Such methods enable creating representative maps for annual or seasonal time periods to use either as input for model building or as visualisation of the environmental situation in the given period. The raw images are presented in Figure 5, which shows the scenes for April month of years 2014, 2017, 2018, 2020, 2022 and 2023.

Figure 5.

Landsat 8-9 satellite images covering salt pan Sebkha de Ndrhamcha in western Mauritania in natural colours showing floodplain for 6 years (always April): (a) 2014, (b) 2017, (c) 2018, (d) 2020, (e) 2022, (f) 2023. Brief explanations of the colours: black colour indicates water areas of the Atlantic ocean, light beige colour represents sandy areas, brown colour indicates desert and bare lands, dark brown colour shows wet areas, grey colour indicates artificial surfaces (roads and urban areas) and cyan colour indicates salt pans on the original images. The colours are in Red Green Blue (RGB). Data source: US Geological Survey (USGS), downloaded from the EarthExplorer repository.

The map in Figure 1 shows the topographic setting and the location of the study area. The map is made using the Generic Mapping Tools Version 6.4.0 scripting toolset (Wessel et al., 2019). The programming scripts used for cartographic techniques are described in earlier works (e.g. Lemenkova (2022a,b,c)).

The images were processed using the scripts provided in Appendix. First, the data were imported into the GRASS GIS environment and preprocessed using the script provided in Listing 1. Then, the images were classified using unsupervised clustering methods with k-means algorithms. The GRASS GIS script used for unsupervised classification is provided in Listing 2. The results of this classification were used as training polygons for ML methods, providing the ‘seed’ information for supervised learning. The next step included the application of RF Classifier for image classification. The programming code for this is provided in Listing 3. The following step included the use of the DTC for image classification, with the script demonstrated in Listing 4. The algorithms of GBC for image classification were performed using the programming code shown in Listing 5.

Finally, the algorithm of SVM Classifier was applied for supervised image classification using the script in Listing 6. The performance of the algorithms demonstrated different time consumption and machine resources. Therefore, to note the difference in speed and the time required for execution, we summarise the performance of the methods in Table 2. The quickest algorithm was the DTC, while the longest period of data processing was demonstrated by the SVM Classifier. The data were processed using the MacOS 14.2.1 (23C71), Apple M1 Chip Processor.

Environmental and hydrological planning in arid regions of Africa is essential and is largely based on using geoinformation such as satellite images. Due to their severe climate and vast spatial regions, salt pans are extremely dynamic natural features of Mauritania that are challenging to investigate using in situ approaches. This study employs spaceborne multitemporal (2014–2023) multispectral Landsat data combined with thematic on topography and geology of Mauritania to document the spatial distribution of surface crust types over time in the western part of the country: region of Sebkha de Ndrhamcha. For this purpose, six satellite images were collected from the US Geological Survey (USGS) and analysed for spectral differences covering the period 2014–2023, which reflects different seasons and surface conditions of the salt pan. The data regarding the mineralogical and geological setting were analysed using the thematic maps (Figures 2 and 3) as well as distribution of land cover types according to Food and Agriculture Organization (FAO) (Figure 4).

An approach based on the GRASS GIS tools was developed to perform the spatiotemporal analysis of the salt pan crust dynamics in a dense time series consisting of six Landsat 8-9 OLI/TIRS cloud-free scenes between 2014 and 2023, covering major wet–dry cycles in Mauritania. With regard to this, the current paper contributed to mapping of the hydrologically vulnerable areas of Mauritanian coasts through detecting the fluctuating area of sebkhas and crust salt pan. Changes in the land cover types detected and visualised by GRASS GIS are generated from the Landsat 8-9 dataset and shown in the corresponding figures (Figures 6–11 according to the applied methods of unsupervised and supervised classification. RS data were used since they provide reliable source of information to clarify environmental dynamics of salt pans. In addition, they support further investigations on crucial limits on the sedimentological, mineralogical and hydrological history of the salt pans.

Figure 6.

Classified Landsat 8-9 OLI/TIRS images using clustering: (a) 2014, (b) 2017, (c) 2018, (d) 2020, (e) 2022, (f) 2023. Software: GRASS GIS. Source of maps: author.

Figure 7.

Maps of rejection threshold probability for accuracy analysis of image classification by chi-squared test: (a) 2014, (b) 2017, (c) 2018, (d) 2020, (e) 2022, (f) 2023. Software: GRASS GIS. Source of maps: author.

Figure 8.

Random Forest Classifier of Machine Learning (ML) classification methods applied for processing of satellite images: (a) 2014, (b) 2017, (c) 2018, (d) 2020, (e) 2022, (f) 2023.

Software: GRASS GIS. Source of maps: author.

Figure 9.

Decision Tree Classifier of Machine Learning (ML) classification methods applied for processing of satellite images: (a) 2014, (b) 2017, (c) 2018, (d) 2020, (e) 2022, (f) 2023.

Software: GRASS GIS. Source of maps: author.

Figure 10.

Gradient Boosting Classifier of Machine Learning (ML) classification methods applied for processing of satellite images: (a) 2014, (b) 2017, (c) 2018, (d) 2020, (e) 2022, (f) 2023. Software: GRASS GIS. Source of maps: author.

Figure 11.

Support Vector Machine (SVM) Classifier of Machine Learning (ML) classification methods applied for processing of satellite images: (a) 2014, (b) 2017, (c) 2018, (d) 2020, (e) 2022, (f) 2023. Software: GRASS GIS. Source of maps: author.

The land categories for years 2014–2013 and their statistics are presented in Table 3. Here, land cover classes are designated as the following categories: 1) Water bodies, 2) Shelf and coastal plains, 3) Sebkha, 4) Urban areas, 5) Sahelian grassland, 6) Salty sands, 7) Compact soil, 8) Stony desert and yellow dunes, 9) Sandy desert and white dunes and 10) Bare soil and rocks.

The results of the satellite image processing are presented in the maps shown in Figures 6–11 in the following subsections. The results of the ML-based classification for Sebkha de Ndrhamcha region are shown in the corresponding figures for each method applied (Clustering, RF Classifier, DTC, GBC and SVM), and numerical computation of the area occupied by each land category is summarised in Table 3. The accuracy assessment of the classification model is evaluated using statistical algorithms for computation of the overall accuracy, User’s and Producer’s accuracy and Kappa coefficient. The results are reported in Table 4.

Here, the accuracy is evaluated as the total number of correctly classified pixels on the Landsat scenes divided by the total number of pixels on this image using User’s accuracy and Producer’s accuracy followed by Cohen’s Kappa computation. Hence, the User’s accuracy is defined as the proportion of pixels classified to a specific land cover type among all the evaluated pixels that really belong to that land cover type. The evaluation was performed against the ground truth data referred to the FAO classification and includes the computed false negatives. Table 4 summarises and compares computed parameters of accuracy and Cohen’s Kappa coefficient.

The interpretation of the Cohen’s Kappa statistics is as follows. The highest possible value of 1.0 indicates the perfect precision and accurate classification of variables, while the lowest theoretically possible value of near zero indicates the lowest precision. For both User’s accuracy and Producer’s accuracy, higher values normally mean better results of the classification of the satellite images. Based on the computed values of Cohen’s Kappa coefficient and the accuracy method for the evaluation of classified data, the results suggest that the SVM and RF algorithms are rated as the most reliable approaches due to the high values.

The Producer’s accuracy indicates the proportion of pixels that are correctly classified and belong to a target land cover type. The Cohen’s Kappa coefficient measures the difference between the actual reference between the data based on the overall accuracy and the actual results of the land cover classification. In other words, it shows the actual changes between the referenced data and a randomly selected pixel classified using a given method. As other correlation statistics, the Kappa values can range from –1 to +1. The interpretation of values is as follows. Values of the levels above 60%–80% have a moderate accuracy of pixel’s classification, while those above 80% are acceptable and generally demonstrate accurate identification of target class.

Consequently, the values approaching above 90% and nearly 100% indicate almost perfect classification results. Thus, the SVM and RF techniques can be evaluated as robust approaches of GRASS GIS for RS data classification based on accuracy, followed by the DTC and GBC methods. The overall classification accuracy is above 80% with a corresponding overall Kappa coefficient above 0.85 (Table 4). Such results demonstrate robust conceptual agreement of the processed data with ground truth. This indicates that GRASS GIS has a reliable algorithm of RS data processing that can be used in similar works.

This paper showed the application of GRASS GIS modules of ML for satellite image processing. Practical use of the presented maps consists in the support to make use of geoinformation, adjust it for land monitoring and discover the climate factors which cause landscape dynamics in the coastal regions of Africa. Using the algorithms of image processing from statistical and programming libraries, the changed landscape patterns were detected and the cartographic visualisation was adjusted from a computer vision perspective using several ML-based algorithms, which are RF Classifier, DTC, GBC and SVM Classifier. The traditional unsupervised methods of classification were compared with the ML-based methods. While clustering demonstrated difficulties in correct detection of land parcels using spectral reflectance, the ML-based supervised techniques of trained learning, as a complement, detected, identified and recorded such information using computer vision algorithms in an automated manner. The outcomes can also be employed to enhance or control the quality of current land cover maps in Mauritania.

To cope with the problem of soil salinisation in the drought-prone region of arid Africa, such as coastal Mauritania, EO data are needed as effective data source for comparison and visualisation of changes and dynamics of salt crust. However, the use of the traditional methods of RS data processing is time-consuming and might be prone to misclassification. In contrast, ML methods applied to EO data processing ensure automated and effective way to visualise, classify and numerically estimate changes in land cover types using algorithms of statistical data processing and computer vision. With regard to this, this manuscript contributes to the development of such methods in West Africa.

This study presents the ML-based cartographic analysis of changes in land cover changes in the coastal region of southwestern Mauritania, including the salt pan of Sebkha de Ndrhamcha, to visualise dynamics in salt crust and surrounding landscapes. The data and methods were selected to achieve this goal according to the cartographic techniques of satellite image processing. Hence, the data used for evaluation included satellite Landsat images obtained from USGS as the RS dataset from 2014 to 2023. The methods used for evaluation and cartographic work included unsupervised and supervised classification used for comparison between traditional and ML GIS techniques of GRASS GIS. The results indicated that there was a change in land cover types. The changes between the class occupied by compact soil and bare soil and rocks correlate well with those in sandy desert and dunes. This is especially notable for the years 2020 and 2022, which shows regional climate impacts with the effects of winds and aeolian processes that increase the areas of dunes and sandy areas and decrease bare soils, accordingly. Likewise, the decrease of shelf and coastal plains is visible for the same period, which might be related to the active upwelling processes in this time.

In general, the coastal region of southwestern Mauritania, along Sebkha de Ndrhamcha, has experienced considerable changes in land cover classes over the past decade (2014–2023), throughout the area of categories, including their fragmentation and extent. However, changes have been significant in the central part of the salt pan, which also supports the main fluctuations of saline crust and salinisation of soils in the surroundings. Thus, the results of this study revealed that over the past 10-year period, salinisation of the sandy areas around salt pan increased, whereas the barren lands decreased in the study area and were replaced by the salty sands and dunes of desert. The details of land cover changes are shown in corresponding figures made with various techniques of ML algorithms using GRASS GIS. The overall results of the study highlighted that GRASS GIS techniques of ML are highly effective for satellite image processing as advanced and straightforward methods of supervised image classification.

The satellite-based evaluation of landscape in western Mauritania is focused on the Sahara–Sahel transition zone of arid and semi-arid climate. Scarce water resources and fluctuation of salt lakes and sebhka motivate the need for mapping of water availability and distribution of salt pans. Changes in land cover classes including fluctuations in salinisation were revealed using ML methods of satellite image processing using GRASS GIS. Practical application of this work consists in land management since visualisation of the extents of wetlands and water bodies enables to plan potential drainage networks based on the hydrological connections, as well as evaluate spatial location of aquatic habitats Campos et al. (2012). Hence, new created maps demonstrated changes in distribution of sebkhas and can assist land planners and environmentalists specialised on Mauritania.

The limited scope of this study was focused on the preselected tools with the available libraries of GRASS GIS, which present an interface to the scikit-learn library of Python. Therefore, future similar research can be proposed with the focus on further development of spatiotemporal analysis using GRASS GIS. For example, the use of GRASS GIS with the available ML tools may be extended towards its more sophisticated tools for spatiotemporal analysis (t.* commands). The quality control and enhancement of current land cover products in Mauritania could be achieved through land cover maps produced using such ML techniques. Furthermore, we also recommend the use of other satellite products such as Sentinel-2. The application of RS data with high resolution would be useful for broad-scale land cover mapping in West Africa over selected time periods. The Sentinel-2 data can be achieved based on Copernicus open-access repository. Finally, future work is suggested to produce and evaluate the effectiveness of a satellite time series of broad scale covering Mauritania for annual land cover mapping. The comparison of such data with existing reference land cover products would be useful for operative monitoring of arid areas of West Africa.

To conclude, the demonstrated case of environmental mapping by satellite data processing by using ML algorithms supports the analysis of salinisation in salty sand areas and sebkha of West Mauritania related to climate change. Using GRASS GIS as a powerful cartographic tool promotes the use of scripts as advanced techniques for processing geoinformation. Such technologies ensure future research efforts, which can continue further the presented study through focus on identifying salinisation of soil for other time periods. GRASS GIS-based methods of processing EO data and formulating optimal ML approaches among the existing algorithms enable to develop effective methods for satellite image processing. In turn, efficient cartographic instruments supported by programming methods for image processing provide technical support for the environmental climate visualisation and analysis of rare areas such as Sebkha de Ndrhamcha, West Mauritania.