Machine learning algorithms for optimization of image classification in spatially constrained regions: A case of Eritrea, East Africa

Vol. 4 • (2025) • No. 2 Engineering Today ORIGINAL SCIENTIFIC PAPER OPEN ACCESS © Faculty of Mechanical and Civil Engineering in Kraljevo. All rights reserved 13 Machine learning algorithms for optimization of image classification in spatially constrained regions: A case of Eritrea, East Africa Polina Lemenkova1* 1 University of Bologna, Department of Biological, Geological and Environmental Sciences, Bologna, Italy ARTICLE INFO * Correspondence: polina.lemen[email protected] DOI: 10.5937/engtoday2500008L UDC: 621(497.11) ISSN: 2812-9474 Article history: Received 15 July 2025; Revised 5 September 2025; Accepted 12 September 2025 ABSTRACT This paper presents the application of Machine Learning (ML) algorithms to solve the problem of optimization of classification tasks in Remote Sensing (RS) data processing. RS data is effective in spatial environmental monitoring since it enables detection of areas affected by natural hazards: droughts, desertification, coastal floods and deforestation. Vulnerable regions can be identified using analysis of spaceborne images for strategic land planning and decision making. The effectiveness of several ML models was tested using Geographic Resources Analysis Support System (GRASS) GIS software for satellite image analysis. Employing ML enabled to perform image classification tasks based on similarity of spectral reflectance of pixels. The following algorithms were tested and compared: Gaussian Naive Bayes (GNB), Decision Tree Classifier (DTC), and Linear Discriminant Analysis (LDA). The ML models were adopted to classify a time series of the Landsat 8-9 OLI/TIRS images and evaluate changes in land cover types in coastal and desert areas of Eritrea. This region encompasses the protected Semenawi Bahri National Park, notable for a diverse range of unique wildlife near the Massawa Channel, Red Sea. The results demonstrated changes in land cover types over the period of 2014-2024 which proved the climate-related effects on landscape dynamics. This paper demonstrated the efficiency of the ML methods in Geographic Information Systems (GIS) tailored to solve specific spatially constrained problems of land cover type identifying using scripting in GRASS GIS. KEYWORDS Machine learning, Image analysis, Africa, GRASS GIS, Remote sensing 1. INTRODUCTION Non-linear modelling and large volumes of spatial data are often the issue in complex task of detecting interactions between variables which require a high level of automation [1-4]. Modelling environmental interactions in African region is surrounded by uncertainties due to the landscape diversity and non-differentiable parameters that affect ecosystem formation: geology, topography, soil, climate and vegetation [5-6]. Remote sensing (RS) is a set of techniques used for Earth observation which can solve the problem of complex mapping through the analysis, and interpretation of objects visible on the Earth’s surface. More specifically, the energy of electromagnetic radiation, emitted or reflected by objects, is measured remotely without physical contact and then analyzed from the data captured on the satellite images. With this regards, environmental problems related to RS data processing require the use of the advanced methods, as their integrated use ensures high-quality geospatial analysis. Rapid development of machine learning (ML) methods in Geographic Information Systems (GIS) made a precious contribution to the advancement of spatial analysis and environmental sciences using RS data. In contrast to the traditional tools of image analysis, ML operates with provides a problem-solved solutions of image classification Engineering Today Vol. 4 • (2025) • No. 2 14 P. Lemenkova through spatial optimization [7-8]. Projects dedicated to Earth observation using high spatial and medium temporal resolution satellites have also been developed which increased the volume of data available for processing and analysis. In this regards, ML-based GIS analysis of satellite images supports processing of large amounts of data which requires building complex models, and performing complex spatial calculations [9-11]. In such cases, exact methodological solutions are resource-intensive and time-consuming. To optimize spatial analysis, heuristic approaches and automated ML methods allow finding prioritized solutions to satellite image analysis. Albeit not always perfect, the sufficient decisions in image analysis are acceptable for practical needs, speeding up the process and facilitating data analysis [12]. Specifically for environmental mapping, advanced methods of data analysis support decisions in complex cases where multiple factors and parameters (topography, climate, vegetation, geology) affect modelling [1314]. To assess the potential improvement of several factors (climate-topographic, geologic and land cover types, the RS data provides solutions to monitoring landscape dynamics through image classification. Such automated methods are performed using clustering search of pixels while categorizing land cover types and ensuring more efficient analysis of landscape dynamics. In this way, image analysis is used to detect deforestation and land cover change through integrated multi-factor analysis of geologic, topographic and vegetation variables. Moreover, RS-based mapping is time-efficient and effective in analysis and geospatial solutions process. Integration of ML and RS data methods support optimization of GIS analysis through optimized landscape mapping. In view of the effectiveness, this research presents automated methods of image processing and analysis for environmental monitoring. 2. STUDY AREA The study area encompasses coastal and desert areas of Eritrea around the Massawa Channel and includes the Semenawi Bahri National Park stretching along the west coasts of the Red Sea, Figure 1. Figure 1: Study are on the topographic map of Eritrea. Software: GMT. Map source: author. Landscape dynamics in Eritrea has a complex pattern which differ regionally demonstrating a variety of land cover types, Figure 2. Diverse habitats of the country include tropical and subtropical grasslands, grasslands in savannah, shrubland and xeric vegetation in semi-desert and desert regions, moist broadleaf forests along the mountainous slopes. Recently, climate warming and anthropogenic impacts (nomadism and land overuse) caused savannah encroachment in semi-natural grasslands in mountains and plateaus. Such processes lead to habitat fragmentation and land cover changes. Important environmental problems of Eritrea include desertification and changes in land cover Vol. 4 • (2025) • No. 2 Engineering Today FMCE Kraljevo 15 patterns. Landscapes of Eritrea are very contrasting, since they are located at the intersection of three African main climate zones: temperate, arid and tropical savannah. Major vegetation types include highland xeric bushes, Acacia woodlands and riverine forests. Coastal regions of Eritrea are one of the least known ecosystems in East Africa. Their extended location along the Red Sea makes Eritrea vulnerable to climate change. Dominated by mangrove forests and seagrasses, coastal zones are notable for high biodiversity, rare and precious bird species of the Red Sea. The Massawa Channel separating the archipelago of Dahlak Marine National Park from mainland is notable for rich wildlife. Here, high biodiversity includes unique fish species, migratory birds and rare mammals [15]. At the same time, this region is affected by climate warming: repetitive droughts, seasonal rainfalls with irregular patterns and high evaporation during dry periods. Land cover change is triggered by climate-related processes: extreme temperature in inner deserts, irregular and low precipitation, high evapotranspiration. Significant changes in air temperature in Eritrea are recorded with daily min/max values constantly increasing for ca. 0.20°C per decade [16]. Such processes cause water deficit which lead to the desertification and aridification of landscapes, and land cover types [17-18]. Figure 2: Land cover types in Eritrea. Data source: FAO. Map source: author. Extreme weather induces water deficit and affects vegetation coverage causing land degradation, overgrazing, deforestation, soil erosion and desertification. The environmental problems of Eritrea are exaggerated by the specific topography: the Danakil Depression is the third lowest located place on the African continent and the hottest place on Earth with regard to the mean temperatures. Typical landscapes are presented by hot and dry deserts with barchan dunes [19]. In such conditions, rare species depend on the availability of water and have to adapt to climate seasonality. Decadal variations in temperature and salinity of water in southern Red Sea affect regional environmental setting [20]. Rising temperatures contribute to degradation of mangroves and coral reefs where coastal plains provide habitats for migrating birds and endemic species [21]. Besides, riparian vegetation in the coastal regions of Eritrea are washed away over decades due to river widening [22]. Seasonal climate gradient is more pronounced in the south of shelf ecosystems [23] where unique landscapes support local inhabitants through essential services of livelihood. In coastal areas of the Red Sea, riverine resources are being actively used by population for economic reasons as source for timber, wood, fuel, charcoal, food and livestock. Such intense anthropogenic activities triggered land cover changes across various regions of Eritrea. 3. OBJECTIVES AND GOALS The research objective is to understand the potential influence of climate effects on landscapes and fluctuations of water level in eastern Eritrea using advanced methods of image analysis. A qualitative-quantitative approach of the environmental analysis using RS data allows understanding not only the patterns and trends of biodiversity and land cover change, but also compute the rates of changes to understand the reasons why they are changed. To this end, Engineering Today Vol. 4 • (2025) • No. 2 16 P. Lemenkova a series of Landsat imagery on 2014, 2018, 2022 and 2024 was processed and analyzed using heuristic cluster search. To find optimal method of image analysis, we aim at testing and comparing several ML methods of image analysis using advanced cartographic software Geographic Resources Analysis Support System (GRASS) GIS. The core issue is to employ the objectivity and automation of the computer vision algorithms of GRASS GIS and to use the advantages of ML techniques. The ML-based automated approaches of image analysis aim to extract information on land cover types from satellite images, limiting human intervention into the classification process. To achieve this, ML-based classification methods associate each pixel in the image automatically with a land cover class based on the value of spectral reflectance. Algorithms of ML are based on the principle of ML which simulates computer vision approaches of visual data analytics that enable machines to discriminate, interpret, and classify data into clusters. ML outperforms the traditional methods of classification through increased precision and accuracy of mapping. Technically, the ML approach of GRASS GIS for mapping and image classification relies on the algorithms of Pythons' Scikit-Learn library. Detailed statistical results on the classified images are included in the GitHub repository with reports for each year. 4. MATERIALS AND METHODS 4.1. Data In this study, we used the most well-known RS data sources are satellite images of the Landsat program launched by the National Aeronautics and Space Administration (NASA) in 1972. Initially dedicated to assessing grain harvests, this program now allows the study of the entire continental surface to analyze environmental dynamics of the Earth. A total of eight satellites were launched between 1972 and 2013, three of which are used in this study: Landsat-8 and 9 Operational Land Imager (OLI) Thermal Infrared Sensor (TIRS) provides images at a resolution of 30 meters for eight spectral bands with a revisit time of sixteen days. Figure 3: Original images of the Landsat 8 OLI/TIRS in RGB colours. Annotations of major geographical features are added on the first image. Data source: USGS. Compilation source: made by the author. Vol. 4 • (2025) • No. 2 Engineering Today FMCE Kraljevo 17 The data were obtained from the USGS EarthExplorer repository (https://earthexplorer.usgs.gov/) as time series analysis of the satellite images on years 2014, 2018, 2022 and 2024, Figure 3. There are numerous examples of the use of Landsat for the environment monitoring of the African coastal regions. [24-26]. Figure 3 shows the set of a satellite image time series and its associated characteristics. Spatial resolution, coupled with the imaged area, defines the image size, while the number of bands indicates the multispectral character of the images. Temporal resolution allows the same scene to be viewed multiple times which enabled time series analysis: 2014, 2018, 2022 and 2024. 4.2. Workflow Image processing and classification has been performed using console-based techniques of GRASS GIS software [27]. This software was selected due to its effectiveness and functionality: it processes diverse data format, supports complex functionality of image processing and imports the results into the bitmap data format. The embedded algorithms of GRASS GIS modules for image processing enable to perform accurate training of the models using scripting approaches [28]. The AI-driven GRASS GIS ML framework of image processing described below was employed to map seasonal variations in the Eritrea using satellite multispectral Landsat 8-9 OLI/TIRS data with repetitive time span. The performance of the both AI-driven and MaxLike methods of image classification was visualised, compared and evaluated with 30-m resolution Landsat OLI-TIRS images with as colour composites. The land cover data were obtained from FAO dataset: https://data.apps.fao.org/catalog/iso/c1f0e345-8ef1-45a2a0b3-13e183bf13ff using the GlobeCover dataset. As the original dataset depicts the land cover in globe scale, it was downscaled to the study area. The topographic data and a shaded relief background in the classified satellite images were collected from the General Bathymetric Chart of the Oceans (GEBCO): https://www.gebco.net The topographic map was plotted in Generic Mapping Tools Version (GMT) using existing methods of cartographic scripts [29]. The methodology of RS data processing employs Python's libraries of Scikit-Learn [30] with algorithms of supervised classification such as LDA, DTC, GND. These algorithms were compared among each other and with those obtained from the traditional approach of Maximal Likelihood Discriminant Analysis that uses unsupervised classification approach. The best results were demonstrated by the LDA due to detailed classification in mapping. The workflow scheme is presented in Figure 4. Figure 4: Schematic workflow of image processing and data analysis in GRASS GIS. Source: author. 4.3. Algorithms of image processing The scripts are available in the GitHub: https://github.com/paulinelemenkova/GRASS-GIS-scripts-ML-SVM-Eritrea The algorithms of the automatic image classification included the sequence of the technical steps of data processing which included the following cartographic workflow: First, the multispectral channels of Landsat were grouped to avoid the panchromatic bands using the following code: • g.region raster=L_2024_01 -p i.group group=L_2024 subgroup=res_30m • input=L_2024_01,L_2024_02,<...>,L_2024_07 The data were imported into the working folder using model 'r.import': • r.import input=/Users/<path>/<name>.TIF output=L_2024_04 extent=region resolution=region In this snippet of code, the functions 'extent' and 'resolution' define spatial borders and spectral resolution of the input data. In the same way, all the Landsat images and topographic shaded relief from GEBCO grid were imported. Engineering Today Vol. 4 • (2025) • No. 2 18 P. Lemenkova The content of the files was checked using 'g.list rast' module. The next step included creating colour composites. Colours composites were generated using module 'r.composite' because band combinations of the multispectral images detail features on the land surface and provide more information compared to the monochrome bands taken separately. Different combinations of colour composites demonstrate that the texture of land cover types is highly diversified due to the natural processes of water evaporation from the basin and drying desert landscapes. Creating colour composites was done as follows: • r.composite blue=L_2024_07 green=L_2024_05 red=L_2024_04 output=L_2024_754 The isolines were added on the maps using 'r.contour' module with 200 m interval using following code: • r.contour shaded_relief out=isolines step=200 --overwrite Spectral signatures for land cover types were generated using module 'i.cluster' of clustering algorithm, which performs the unsupervised automatic classification with a variable number of clusters. Specifically, it detects and recognises spectral reflectance values of pixels of the images. Practically, the code used for this step is as follows: • i.cluster group=L_2024 subgroup=res_30m • signaturefile=cluster_L_2024 classes=10 reportfile=rep_clust_L_2024.txt The results of data handling are saved for each image. The signature file was used as input for 'i.maxlik' to generate an unsupervised image classification using information in signature file. The maximum-likelihood classifier uses the cluster means and covariance matrices obtained from the signature file generated by the 'i.cluster'. Here, the algorithms define to which category based on the definition of spectral class belongs each cell in the satellite image. This partition is done based on the evaluation of the probability of belonging of each cell to the classes (here, the example for the image on 2024, repeated for each year). The classification model of GRASS GIS is defined by the module 'i.maxlik' which performs image partition using a maximum-likelihood discriminant analysis classifier using the following code: • i.maxlik group=L_2024 subgroup=res_30m signaturefile=cluster_L_2024 • output=L_2024_clusters reject=L_2024_cluster_reject 4.4. Algorithms of machine learning (ML) The ML algorithms included the Gaussian Naive Bayes (GND), Decision Tree Classifier (DTC), and Linear Discriminant Analysis (LDA). All the ML models were trained and randomly selected points from the test image. The ML approach was realised in GRASS GIS relevant modules and the following scheme. First, the training pixels were generated from an older land cover classification using the following code by 'r.random' module: • r.random input=L_2014_clusters seed=100 npoints=1000 raster=training_pixels This module generates randomly placed raster pixels which is useful for creating training data for ML approaches. First, we used the GNB classifier which is based on Bayes' theorem, the Gaussian Naive Bayes algorithm – a ML classification method that assumes that features have a normal distribution. It works especially effectively with datasets that contain continuous characteristics, such as landscapes of the Earth. The next approach of classification used the DTC which uses a decision tree to make predictions using hierarchical data assessment. The DTC is a method of ML that incorporates non-parametric supervised learning and is utilized for classification and regression. Hence, the principle follows a dendrite-like model of decisions taken by the algorithm and possible assignment of pixels to a given class of land cover types. The objective of DTC is to learn basic decision rules derived from the data features in order to build a model that forecasts the value of a target variable. A tree can be thought of as an approximation of a piecewise constant. The DTC functions by iterative splitting the pixels on the image into subclasses based on the most significant values of spectral reflectance of pixels that form classes representing the nodes of the tree. The technical approach employs n_estimators parameter in every ML model, included in the syntax of the code. The number of estimators parameter is derived from Scikit-Learn library of Python and affects only tree-based estimators, such as Random Forest. In this context, it specifies the number of separate decision trees trained on subsets of observations. A supervised ML technique of LDA is employed for dimensionality reduction and classification. Its specific feature is that it looks for a linear feature combination in a dataset that best divides two or more classes. Therefore, the major approach of the LDA algorithm is that it solves the multi-class classification problem in highly diversified patterns of landscape with repetitive mosaics. It separates multiple classes which have diverse several features by reducing the data dimensionality. The programming code of the LDA algorithm in the GRASS GIS is as follows. First, the data are trained using the function ‘training_map’ in the 'r.learn.train' module as follows: • r.learn.train group=L_2024 training_map=training_pixels The use of supervised ML classification methods to obtain land cover maps requires training samples to train the classification algorithm. These training samples are described by a vector of variables extracted from satellite data Vol. 4 • (2025) • No. 2 Engineering Today FMCE Kraljevo 19 and a label provided by the reference data. The label is an indication of the land cover class of the training samples, which is used in decision-making to assign the class of new observations. The quality of the labels is therefore directly related to the accuracy of the decision rule learned by the classification algorithm. In this study, we used the available FAO map for training polygons to train the ML models. Afterwards, the ML model was selected for discriminating categories. • model_name=LinearDiscriminantAnalysis save_model=lda_model.gz Afterwards, the fitted Scikit-Learn estimator is applied to rasters using 'r.learn.predict': • r.learn.predict group=L_2024 load_model=lda_model.gz output=lda_classification_2024 Following that, the raster files were visualised using the combi nation of modules: “d.rast” for visualization of grids as explained above, ‘d.vect ’ was used for plotting the isolines as follows: • d.vect isolines color='100:93:134' width=0 Different colour palettes were applied for visualising classification results of various ML models (byr for LDA, roygbiv for GNB and plasma for DTC). This was done for distinguishable visualization of the results and comparison of maps. Using the selected palette for all the maps was supported by the legends where the same set of classes is present and described in classification according to the designated classes. The cartographic grid was added using ‘d.grid’ module: • d.grid -g size=00:30:00 color=white width=0.1 fontsize=16 text_color=white Two legends were added using the ‘d.legend’ module with defined parameters of the cartographic code: • d.legend raster=L_2024_clusters title="Clusters 2024" • title_fontsize=19 font="Helvetica" fontsize=17 bgcolor=white The same principle of mapping was used for all the visualised maps. The given procedure was completed for all the maps and visualised for GNB algorithm of ML. The key approach of the GNB as a ML classification technique consists in an assumption that supposes that each land cover class follows a normal distribution. It presumes that each parameter of vegetation reflectance has an independent capacity of predicting the output variable. This is advantageous for homogeneous classes. However, for landscapes of Eritrea, the input data are heterogeneous and selected classes (xeric vegetation in semi-desert areas and mangroves along the coasts of the Red Sea, mosaic grasslands and shrubland) do not contain such normal distribution and have regional distribution patterns. Since RS have limitations and inherent errors that can be introduced during image processing, numerical and visual classification, rectification to real-world coordinates, the accuracy assessment was performed as post-processing and recording of data accuracy on the classified images. Hence, advanced methods of RS data processing were used to evaluate accuracy of mapping through script-based image analysis. The overall performance of the tested ML classifiers was assessed using metrics of accuracy, F1 score and Cohen's kappa coefficient. The Cohen's kappa is a quantitative metric used to analyze the correctness of pixels' assignment to various land cover classes and the dependability of rating coefficients [31]. The Cohen's Kappa method was selected because of its robustness for data evaluation reported earlier in research papers [32-34]. 5. RESULTS AND DISCUSSION The presented series of maps show changes in land cover types and desertification indicating climate effects of drought, raise in temperature and aridification in Eritrea, Figures 5-7. In the context of land use mapping, ML classification algorithms exploit the spectral variability of objects on the surface visible on the satellite data [35]. Indeed, this variability induced by different factors that affect vegetation and soil properties. In turn, these depend on several factors such as climate processes, human activities, and differences in soil properties that complicate the appearance of land use classes on the images. Furthermore, this variability becomes more contrasting for large areas, such as African regions. The classification of the satellite images revealed significant changes in land cover in Eritrea during recent decades. These included including a notable desertification and reduction in forest and vegetation cover. The reasons of such phenomena refer to resource exploitation, population pressure, and climate change. In particular, the comparison of the classified images from 2014 to 2024 indicated the decline in forest cover and the increase of lands occupied by deserts. Deforestation and land degradation in Eritrea may have negative effects on the ecosystem services, resulting in food and water insecurity. The comparison of the results with available data on earlier periods show the consists of trends of land degradation in Eritrea. Thus, according to Global forest water survey, as of 2000, Eritrea had 5 ha of tree cover, equivalent to less than 0.1% of its land area and < 0.1% of the global total. Besides, less than 0.1% of land cover in Eritrea was classified as tree cover with over 30% of canopy density. Moreover, the top regions representing 93% of Engineering Today Vol. 4 • (2025) • No. 2 20 P. Lemenkova all tree cover belongs to the Semenawi Keyih Bahri which had the most tree cover at 40 ha compared to an average of 7 ha, as of 2010 [36]. In addition, changes in land cover types and habitat loss impact wildlife habitats and species distribution which can have implications for biodiversity. Moreover, environmental degradation increases vulnerability of the ecosystems to climate change, and extended and repetitive droughtse. Such environmental problems indicate the persistent issue of land degradation in East Africa and the need for measures on sustainable land management to support the environmental sustainability of this region. Figure 5: Classified images Landsat 8 OLI/TIRS using GNB algorithm. Source: author. Spatial analysis was performed at basin and landscape scales by combining different GIS and ML measurement approaches. Specifically, we proposed an AI-driven workflow framework that combines traditional classification of the satellite images based on clustering, maximal likelihood and ML method using Python's Scikit-Learn library embedded in GRASS GIS to assess the reliability of modelling and mapping. The series of several Landsat-8 OLI/TIRS satellite images processed in this study by ML techniques illustrated the process of intense land cover changes that is related to climate processes, such as rise in temperatures, lack of precipitation and increased evaporation of water over coastal Eritrea, which proves the previous results reported earlier [37-38]. Climate fluctuations over Eritrea in period 2014 to 2024 was found to play a dominant role in the precipitation and evapotranspiration partitioning which resulted in changed land cover types, Figure 5. The algorithms included AI application for image analysis of Python's Scikit-Learn library: GND, DTC and LDA. The statistical characteristics on the initial image properties for four scenes are summarised in Table 1. Table 1: Numerical characteristics of the initial properties of multispectral bands in satellite Landsat scenes Year Stat Band 1 Band 2 Band 3 Band 4 Band 5 Band 6 Band 7 2014 Means 9162.01 9860.14 11384.9 12259.4 13715.7 15723.4 14815 St. Dev. 1665.86 1904.88 2890.35 4024.82 5046.8 6495.05 5919.18 2018 Means 9154.1 9938.56 11577.5 12482.2 13883 15826.9 14922.7 St. Dev. 1797.38 1923.61 2703.44 3791.01 4719.46 6077.47 5511.82 2022 Means 8899.43 9658.08 11188.3 11874.6 13591.5 15141.3 14002.1 Vol. 4 • (2025) • No. 2 Engineering Today FMCE Kraljevo 21 Year Stat Band 1 Band 2 Band 3 Band 4 Band 5 Band 6 Band 7 St. Dev. 1634.65 1801.18 2557.68 3574.09 4676.77 5836.53 5099.32 2024 Means 8785.91 9433.87 10727.5 11383.4 13271.2 14787 13712.2 St. Dev. 1453.00 1651.73 2503.25 3468.44 4725.69 5879.95 5185.96 Figure 6: Classified images Landsat 8 OLI/TIRS using DTC algorithm. Source: author. Processes of environmental changes are notable in the coastal areas due to the distribution of wetland ecosystems, comparing four images on several dates: 2014, 2018, 2022 and 2024. The DTC algorithm (Figure 6) has a high automation in image processing, accuracy in pixel assignments, increased details of classification with more fragmented identified patterns in the landscapes of eastern Eritrea. The detected landscape dynamics enabled to demonstrate the robustness of the AI-based classification method to monitoring seasonal variations using Earth observation data. The drawbacks of the DTC for image processing include the instability to changes and noise in the dataset (even low cloudiness might affect the output). The reported advantages of the DTC include the distinguishability of urban areas from barren and bare land use types, such as beach and coastal cities [39] and crop areas [40]. Besides, the algorithm is based on complex calculations and is computationally expensive with need for computer memory and resources which is demanding for large datasets and limited computer memory. Finally, the classified categories are unbalanced as the algorithm is non-continuous. The results on computed class means for each of the 10 classes and four evaluated years are reported in Table 2. Table 2: Results on computational analysis: means of pixels for each of the 10 classes in Landsat images. Year Class 1 Class 2 Class 3 Class 4 Class 5 Class 6 Class 7 Class 8 Class 9 Class 10 2014 2438 178 71 184 351 568 729 930 789 771 2018 1914 650 66 170 330 578 768 862 908 748 2022 2388 206 125 316 184 544 798 828 808 808 2024 2490 119 198 145 364 497 708 909 844 827