1 Introduction

Land use land cover changes (LULCC) have a big impact on several important aspects of the global environment, including hydrology, ecosystem processes, biodiversity, climate, and biogeochemical cycles (Eyelade et al. 2022). There is a growing understanding of the significance of land use as a critical factor influencing environmental change (Turner et al. 2007). Anthropogenic activities often influence LULC and lead to changes that have an impact on the ecosystem (Muhammad et al. 2015), biodiversity (Sharma et al. 2018), landscape (Kamwi and Mbidzo 2022), urban sprawl (Lu et al. 2018), the net productivity of vegetation (Wang and Wang 2023), and water changes (Assaf et al. 2021; Hasanuzzaman et al. 2022; Mohammed et al. 2022). The utilizing of geographic information system (GIS) methods and remote sensing (RS) data has been recognized as suitable techniques for managing LC and other natural resources (Zanaty et al. 2023). A variety of spectral, spatial, and temporal resolution images from RS data can be utilized to distinguish alterations on the earth’s surface at a lower cost and faster rate than conventional field survey techniques (Koko et al. 2020).

Accurate and up-to-date LULC data are essential for tracking and analyzing transitions. It has been demonstrated in numerous studies that this information can be used to map and analyze changes (Elhag and Boteva 2016). There are several classification algorithms that were applied to satellites image for accurate LULC maps such as Maximum-likelihood classifier (MLC) (Rogan et al. 2002), logistic regression (Colkesen and Kavzoglu 2017), clustering (Lemenkova 2021), neural networks (Mas and Flores 2008), decision trees (Xu et al. 2005), random forest (RF) (Mahmoud et al. 2022), and support vector machines (SVM) (Singh et al. 2014). In a study by Khatami et al. (2016), various image classification methods including SVM, neural networks, RF, and MLC were evaluated. The results indicated that support vector machines achieved the highest accuracy, followed by neural network methods. Additionally, the random forest classifier outperformed the traditional decision tree classifier by a significant margin. However, maximum likelihood classifiers, which are commonly used as benchmark algorithms, showed low accuracy in this evaluation. Wieland and Pittore (2014) evaluated the performance of four machine learning algorithms in recognizing urban patterns from multi-spectral satellite images. It was concluded that SVM and Random Trees (RT) demonstrated the highest classification performance across all image types, meeting the predefined performance criteria. On the other hand, KNN and Normal Bayes (NB) demonstrated satisfactory performance in certain setups, but they struggled with generalization and exhibited unstable behavior when subjected to different training–testing scenarios. Kranjčić et al. (2019) examined the mapping of green urban areas through the utilization of satellite imagery from the sentinel-2 multispectral instrument. They employed four machine learning methods, namely support vector machine, random forest, artificial neural network, and naïve Bayes classifier. The outcomes of their research revealed that support vector machines exhibited superior classification accuracy and performance time when compared to random forest, artificial neural networks, and the naïve Bayes classifier.

Several land use models (LULC) have been developed and applied to analyze the consequences of changes in land use on the Earth system and to predict future land use in different regions and scales (Aburas et al. 2017; Gollnow et al. 2018; Kisamba and Li 2023; Weslati et al. 2023). Precisely monitoring LULC is crucial in assessing the environmental implications of any changes (Cao et al. 2020). Integrating simulations into LULC can support future developments and improve understanding of their potential (Omar et al. 2014). There are several models commonly used in the literature to predict LULC, including statistical models (Somvanshi et al. 2020), Markov chain (MC) models (Ahmad et al. 2017), cellular automata (CA) models (Khan et al. 2022), logistic regression (Buya et al. 2022), and artificial neural networks (ANN) (Aksoy and Kaptan 2022; Sayl et al. 2022). However, no individual model can take into account all change agents, urbanization processes, and physiographic features at various scales (Verburg et al. 2008). These models have been shown to be effective tools for observing and modeling LC in environmental planning as well as management, which is crucial for sustainable urban growth (Yang 2002). Multiple researchers have employed hybrid models, namely the CA-ANN model (Asori and Adu 2023), and CA-Markov model (Abijith and Saravanan 2022), to mitigate the limitations associated with individual models.

The integrated cellular automata (CA) and Markov chain model (CA-Markov) are two widely used models with high accuracy (Munthali et al. 2020; Sibanda and Ahmed 2021). CA-Markov models are robust and offer a convenient way to model the spatiotemporal dynamics of LULC changes in complex systems (Hyandye and Martz 2017; Memarian et al. 2012). Geospatial, remote sensing, biophysical, and socioeconomic data can be integrated with the model to allow for more comprehensive simulations of LULC changes (Weng 2002). LULC Changes in Egypt's northwest coastal desert (Halmy et al. 2015) were investigated, and future alterations were predicted using the CA-Markov. Furthermore, a CA-Markov model was employed to analyze LULCC in Damietta Governorate, Egypt (El-Hamid et al. 2022). The CA-Markov model is well-known and has been utilized in numerous studies (Firozjaei et al. 2019; Fu et al. 2022; Zhang et al. 2021). When comparing the capabilities of CA-Markov and Clue-S models in simulating land-use changes, it becomes evident that CA-Markov has certain advantages over Clue-S. In particular, CA-Markov is capable of simulating a significantly larger area, which gives it a greater scope for analysis and prediction. Additionally, CA-Markov exhibits relative advantages over Clue-S due to its ability to calculate the unpredictability of land-use changes in urban areas. On the other hand, Clue-S lacks this capability, making it less effective in assessing and forecasting changes in city areas (Chotchaiwong and Wijitkosum 2019; Zheng and Hu 2018). Regmi et al. (2017) compared the CA-Markov model to the GEOMOD model for predicting future LULC shifts in the Phewa Lake Watershed, Nepal, and discovered that combining the CA and Markov models was more effective than GEOMOD in forecasting future LULC scenarios.

This study aims to investigate and monitor LULC changes in El-Fayoum, Egypt by considering the socioeconomic transition period between 2010 and 2020. The primary objective is to predict future scenarios of transition between different LULC classes within the region. Google Earth Engine is implemented to apply SVM for improved classification of landsat satellite images. The CA–Markov Model and the multi-criteria analysis model based on geographic information systems (GIS-MCA)—have been integrated into this study using an innovative approach for future prediction of El-Fayoum 2030. The findings of this study contribute to the sustainable development of ecosystem services, which depends on El-Fayoum Governorate's efficient use, management, and preservation of its land resources.

2 Study area, datasets, and software

The governorate of El-Fayoum is situated in the Western Desert, bordered by the desert on all sides except the southeast where it connects with Beni-Suef governorate. Its geographical coordinates range from approximately 28° 55′ N to 29° 40′ N latitude and 29° 55′ E to 31° 5′ E longitude (El-Zeinya and Effatb 2017) (Fig. 1). This region is known for its rich green oasis within the desert terrain. The governorate’s topography is homogenous, and the elevation altered at mean sea level from − 45 to 125 m (Allam et al. 2019). El-Fayoum’s natural environment is a mix of agricultural land, deserts, and urban. The study area experiences moderate winter and hot summer seasons (Khalifa and El-Khateeb 2011), with the lowest and highest temperatures recorded at 14.5 °C and 31 °C, respectively, and an annual rainfall not exceeding 7.2 mm (El-Zeinya and Effatb 2017).

Fig. 1
figure 1

El-Fayoum Geographical location

Full size image

Four different remote sensing datasets are used in the monitoring of El-Fayoum governorate. The first datasets are Landsat 5 and 8 that are described in Table 1, which provide satellite images with medium resolution and data are available for free download through USGS website (http://earthexplorer.usgs.gov). El-Fayoum is covered by two Landsat sensors (TM and OLI) with two different scenes: 177–39 and 177–40. The images were processed by filtering the date, masking clouds with less than 10% coverage, and converting them to GeoTiff format with WGS 84 datum and UTM 36N projection.

Table 1 Specifications of Satellite sensors
Full size table

The second dataset is AW3D30 digital elevation model (DEM). AW3D30 is a DEM with spatial resolution of 30 m that was produced from the Japanese Aerospace Exploration Agency (JAXA) and released in May 2015 (https://portal.opentopography.org). It was used to calculate slope and aspect after filling any depressions. The last two datasets are Google Earth and Open Street Map (www.openstreetmap.org) that are used to digitize roads and canals then they are imported into TerrSet (2020) to calculate the distance from roads and canals.

To accelerate the time spent for downloading and processing of remote sensing data, Google Earth Engine (GEE) platform is used for manipulating images for 2000, 2010, 2012 from Landsat 5™, and 2020 from Landsat 8(OLI). ArcGIS Pro 2.8.3 is conducted to apply SVM to the selected datasets, while Terrset 2020 (IDRISI selva) program environment is used in conjunction with CA-Markov to generate the necessary transition maps, TPM, and TAM from 2010 to 2020 in order to simulate the LULC in the year 2030.

3 Methodology

The spatio-temporal changes in El-Fayoum, Egypt, were analyzed using Landsat images. The proposed flowchart is shown in Fig. 2. Our proposed flowchart has three main phases: Image Classification, Suitability Analysis, and Future LULC Prediction. The following sections provide a detailed discussion of each phase of the flowchart.

Fig. 2
figure 2

The methodology flowchart which applied in this research

Full size image

3.1 Image processing and classification

The Landsat images obtained through Google Earth Engine (GEE) were processed by applying filtering and calibration techniques to enhance the classification results. Standard image processing steps, including geo-referencing, training the classification algorithm, supervised classification, and accuracy assessment, were performed to classify the images. A visual analysis of the images and LULC classes was conducted to gain insights into their spectral and spatial profiles, which were then utilized to train the classification algorithm using the SVM classifier on ArcGIS Pro 2.8.3. Four distinct classes, namely agricultural land (AGRI), water bodies (WTR), built-up areas (BLT), and desert (DSRT), were extracted using SVM classification algorithm across four different time slots: 2000, 2010, 2012, and 2020. A signature polygon was created to validate the detected features from high-resolution satellite images. Furthermore, an error matrix was calculated to evaluate the performance, utilizing equations [Eqs. (1)–(4)] (Naeem et al. 2022).

$$UA=({C}_{ii}/{C}_{i+}) \times 100$$
(1)
$$PA=({C}_{ii}/{C}_{+j}) \times 100$$
(2)
$$OA=({C}_{c}/{C}_{t}) \times 100$$
(3)
$$K=\frac{N{\sum }_{i=1}^{n}{C}_{ii}-{\sum }_{i=1}^{n}{(C}_{i+}{C}_{+J})}{{N}^{2}-{\sum }_{i=1}^{n}{(C}_{i+}{C}_{+J})}$$
(4)

where Ct represents the overall pixel count and Ct represents the correctly classified pixels (diagonal ones). N is the number of validated points, and the symbols Ci+ and C+i stand for the marginal sums of rows and columns, respectively.

3.2 LULC prediction

The CA Markov model is a hybrid model that combines two models with different concepts: the Markov chain and the CA filter. The Markov chain model is utilized as a stochastic process to examine the probability of transitioning from one state to another. This analysis involves predicting the state of a system at a specific time (t) based on its state at a previous time (t − 1). However, it is important to note that the Markov chain model operates independently of the neighboring states of the observed cell. Consequently, relying solely on Markov modeling is inadequate for analyzing the problem because it neglects the spatial distribution of each category. Although the Markov chain model can accurately forecast the magnitude of change, it fails to determine the appropriate direction. To address this limitation and consider the spatial characteristics of the data, the cellular automata model is implemented. By incorporating the spatial nature of the problem, this model provides insights into both the magnitude and direction of change (Ghosh et al. 2017). Markov model could be represented as described in Eqs. 5 and 6, while CA (Hou et al. 2004) is described by Eq. (7).

$${S}_{t}={P}_{ij}*{S}_{t-1}$$
(5)

where Pij is the transition probability matrix (TPM) for a state. It is calculated by Eq. (6) (Wan et al. 2015),

$${\text{P}} = {\text{P}}_{{{\text{ij}}}} = \left[ {\begin{array}{*{20}c} {{\text{p}}_{11} } & {\quad {\text{p}}_{12} } & {\quad {\text{p}}_{{1{\text{n}}}} } \\ {{\text{p}}_{21} } & {\quad {\text{p}}_{22} } & {\quad {\text{p}}_{{2{\text{n}}}} } \\ {{\text{p}}_{{{\text{n}}1}} } & {\quad {\text{p}}_{{{\text{n}}2}} } & {\quad {\text{p}}_{{{\text{nn}}}} } \\ \end{array} } \right],{ }\,\mathop \sum \limits_{{{\text{i}} = 1}}^{{\text{n}}} {\text{P}}_{{{\text{ij}}}} = 1$$
(6)

where P is the matrix of Markov transitions, (i and j) are the categories of LULC for initial and successive timeframes, respectively, n is the number of LULC classes, and Pij is the likelihood that a given type of land will transition from one LULC type to another.

$${S}_{(t-1,t)}=f({S}_{\left(t-1\right)},N)$$
(7)

where t − 1 and to denote the various times, S is the set of finite and separate cellular states, N is the cellular field, and f is the rule for transforming cellular states in local space.

3.3 CA-Markov model validation

Model validation for future LULC changes is a critical component, and the model’s accuracy is frequently evaluated in studies using the kappa index. (TerrSet) IDRISI’s validate modules were employed to estimate the correlation of Kappa indexes between the simulated and actual maps. In case that two maps have a perfect correlation, the correlation value will be close to one. If the simulated map deviates significantly from the actual map, the correlation value will be close to zero. The correlation value reveals how similar or dissimilar the two maps are (Gebresellase et al. 2023). The value of kappa has different indications: one refers to complete agreement, 0.75–1 indicates high agreement, 0.50–0.75 indicates moderate agreement, and kappa less than 0.50 indicates low agreement (Abd El-Hamid et al. 2021). To evaluate the accuracy of the CA-Markov model in El-Fayoum, LULC for 2020 will be employed for validation purposes using TerrSet 2020 software.

The summary statistics’ Eqs. (8), (9), and (10) are as follows (Matlhodi et al. 2021):

$${k}_{(no)}=\frac{({M}_{m}{N}_{n)}}{({P}_{m}-{N}_{n})}$$
(8)
$${k}_{(Location)}=\frac{({M}_{m}{N}_{n)}}{({P}_{m}-{N}_{n})}$$
(9)
$${k}_{(standard)}=\frac{({M}_{m}{M}_{n)}}{({P}_{n}-{N}_{n})}$$
(10)

where Grid cells without information are N(n), grid cells with medium information are M(m), and grid cells with perfect information are P(p).

3.4 Suitability analysis

A combination of factors and constraints to create transition suitability maps specific to each (LULC) category are employed. This process involves conducting a LULC-based multicriteria evaluation (MCE) (BEHERA et al. 2012; Sameer et al. 2021), where the interactions and impacts of various factors are comprehended. Boolean images are used to represent constraints (0 to 1) that limit variation in LULC. A value of “1” indicates a suitable area for suitability analysis, whereas a value of “0” indicates areas controlled by the suitability analysis (e.g., water bodies) (Matlhodi et al. 2021). Factors are usually based on distance, which is applicable to regional variations (Gashaw et al. 2017), for instance, distance to rivers, roads, or urban areas (Jana et al. 2022; Matlhodi et al. 2021; Singh et al. 2022). Furthermore, slope and elevation are factors that determine the usefulness of a land to humans (Ruben et al. 2020). The factors responsible for LULC alter and distribution were used (slope, digital elevation model, canal distance, and road distance) to generate suitability maps. As shown in Fig. 2, the followed steps for suitability maps are illustrated, while Table 2 illustrate the constraints and factors used in this study. In addition the control points collected from previous research (Alimi et al. 2016; Jana et al. 2022; Owusu et al. 2017; Ruben et al. 2020).

Table 2 The criteria used for suitability maps preparation
Full size table

Using the fuzzy approach described by Zadeh (1965), we matched the scale and standardization of sub-criteria. There is a membership function associated with the fuzzy set that indicates the grade of membership. Depending on the function type, such as decrement or increment, the membership function can be sigmoidal, J-shaped, linear, or user-defined (Malczewski and Rinner 2015). Because the sub-criteria were scaled to illustrate the change potential of an area, a linear membership function was employed to normalize all layers between (0 and 1), corresponding to minimum and maximum utility.

3.5 Weighting method

We used the MCE process to compare criteria of different importance. The weights are determined by assigning a number to each factor for for agricultural and classes built-up. The AHP is used for a variety of land suitability assessments to assign weights to the different criteria. By first creating a pairwise comparison matrix (PCM) based on Saaty (1980) guidelines and using a numerical scale of 1–9 to identify the relative relevance of the criteria, AHP was utilized in this work to generate weights for the reclassified and standardized criteria (Saaty 2008). Tables 3, 4 present the paired matrices created for the study, highlighting the relative weights of each agricultural and built-up criteria.

Table 3 Pairwise comparison approach to derive the factor weights for agricultural suitability maps
Full size table
Table 4 Pairwise comparison approach to derive the factor weights for built-up suitability maps
Full size table

The weights calculated using the AHP methods are comparison weights for each individual criterion in relation to the ranks that were assigned to them. It is possible to spot and fix logical inconsistencies in the PCM created using experience or expert opinion by calculating the consistency ratio (CR) using Eq. (11). The judgement may be considered to be inaccurate when the CR value is greater than 0.1 since it may be too near to randomness and has to be modified (Yalew et al. 2016). The consistency index (CI), which serves as an input in calculating the CR, is provided by Eq. (12).

$$CR=\frac{CI}{RI}$$
(11)
$$CI=\frac{\left({\lambda }_{max}\right)-n}{(n-1)}$$
(12)

where, n is the number of criteria or count, and λ max is maximum eigenvalue.

3.6 Future LULC prediction

CA-Markov model is utilized to project the future LULC of 2030 by several steps. Firstly, LULC data in ASCII format was entered into TerrSet 2020, and the CA-Markov model was utilized to generate maps displaying appropriate transitions, the Transition Probability Matrix (TPM), and the Transition Area Matrix (TAM) for the period between 2000 and 2010. These outputs were then used to simulate the LULC for the year 2020. The CA-Markov was implemented based on the LULC maps from 2000 and 2010, and the TPM generated in the previous step. To simulate LULC for 2020, four cycles were calculated considering the number of years to be predicted. An error rate of 0.15 was calculated using the default filter, 5 × 5. Additionally, suitability maps were prepared using MCE. A comparison was made between the predicted and actual LULC maps for 2020, with the validation tool “validate” in TerrSet 2020 used to assess simulation accuracy. This verification process ensured that the simulations for 2030 would yield consistent results. To simulate the LULC in 2030, the TerrSet2020 (IDRISI Selva) program environment was used in conjunction with CA-Markov, generating necessary transition maps, TPM, and TAM from 2010 to 2020. The CA-Markov model was employed to predict the LULC for 2030, considering the number of years to be predicted. A default error rate of 0.15 was established using the default filter (5 × 5).

4 Results and discussion

4.1 Classification of LULC assessment

All three images (2000, 2012, and 2020) were classified using the SVM-supervised technique as shown in Fig. 3. Testing the accuracy and validity of the classification was performed using the tracking data gathered by Google Earth for the relevant years. The accuracy of each map was validated using a confusion matrix by comparing the LULC classes of 370 randomly chosen sites with the referenced data for the selected years. Similar strategies were implemented to validate the classified LULC, and prior research revealed acceptable accuracy (Atef et al. 2023; Khwarahm et al. 2021). Table 5 illustrates (OA) and (K) of the various LULC classes in El-Fayoum governorate land cover maps for various periods. These results seem acceptable when examining the OA and K coefficients.

Fig. 3
figure 3

LULC classification for years 2000, 2012 and 2020

Full size image
Table 5 Accuracy assessment
Full size table

4.2 LULC dynamics

In the current study, two various time periods of 2000 to 2012 and 2012 to 2020 were used to monitor LC changes. Table 6 and Fig. 4 depict the variations in LULC classes. All four classes experienced significant changes, as shown by the change detection analysis. During the first time period, BLT and AGRI were increased while DSRT and WTR were decreased. The BLT area increased from 200.67 to 266.93 km2, and the AGRI increased from 1463.92 to 1570.83 km2. For the decrease classes, DSRT was decreased from 3849.55 to 3665.54 km2, and WTR decreased from 337.73 to 348.57 km2. During the second time period, the change detection found considerable alteration in all classes. The BLT area had increased from 266.93 to 350.34 km2, while the AGRI decreased from 1570.83 to1557.42 km2. DSRT decreased from 3665.54 to 3628.86 km2, while WTR decreased from 348.57 to 315.42 km2. The increasing BLT class occurred due to encroachment on agricultural land. The increase in AGRI class in the first period occurred due to land reclamation in desert areas, but in the second period, there was a decline in AGRI due to encroachment on agricultural land. There were decline in DSRT class due to urban expansion and land reclamation. Finally, the WTR consists of Lake Qarun and Wadi Al-Rayyan, in which agricultural drainage water collects, and due to the lack of agricultural land, it led to a shortage of WTR.

Table 6 Analysis of change detection (in km2) for the years 2000, 2012, and 2020
Full size table
Fig. 4
figure 4

The status of LULC in three various periods in km2

Full size image

The AGRI class, representing agricultural lands, exhibited a remarkable expansion, flourishing from 1463.92 km2 in 2000 to 1557.24 km2 in 2020. Witnessing the growth of AGRI by human cultivation and its impact on the landscape. In contrast, the DSRT class, signifying desert regions, underwent a significant decline, with its territories decreasing from 3849.55 km2 in 2000 to 3628.86 km2 in 2020. The BLT class, representing built-up areas. Its domains expanded from 200.67 km2 in 2000  to 350.34 km2 in 2020. As for the WTR class, symbolizing water bodies, it receded from 337.73 km2 in 2000 to 315.42 km2 in 2020. Drawing intriguing comparisons, we compare our study to the studies of Allam et al. (2019), El-Zeiny and Effat (2017). Our research mirrors the past, as we find parallels with Allam et al.’s accounts of AGRI’s rise from 1547.64 km2 in 1984 to 1635.15 km2 in 2016. Similarly, the water (WTR) class had experienced growth, as indicated by El-Zeinya and Effatb’s findings, expanding from 345.15 km2 in 1990 to 345.46 km2 in 2016. Their study also revealed the captivating ascent of the urban (BLT) class, from 59.13 km2 in 1990 to 235.94 km2 in 2016. And like the sands of the desert, the DSRT class underwent a significant decrease from 3836.67 km2 in 1990 to 3579.33 km2 in 2016.

4.3 LULC transition probabilities for the year 2020

In this study we used TerrSet 2020 to calculate CR and weight of factor for each class. The CR for agricultural and built-up were less than 0.1 as shown in Table 7. As the weight increases, the criterion becomes more important (Malczewski 1999). The final weights generated were not applied to the factor images as a whole; rather, they were applied ‘pixel by pixel’ in the order of suitability scores. The final maps of continuous suitability were the result of criteria aggregation using an operation that is said to be exactly halfway between the AND/ OR operations. In this study, the weighted linear combination (WLC) method was used for the aggregation of parameters (Rahman and Szabó 2022). This process carries the lowest possible risk, as the areas considered suitable are those with all criteria fulfilled. The effect of ‘order of weights’ is most easily understood in terms of levels of risk and trade off. It was neither extremely risk-averse nor extremely risk-taking (Myint and Wang 2006). Any factor could compensate for any other according to its factor weight. At both extremes of the continuum, tradeoffs are not possible, but in the middle, there is the potential for full tradeoffs. Here, the suitability of areas was determined with consideration of drivers or factors, i.e., DEM, slope, distance from the road network, and distance from the canal. The suitability map for each LULC class was prepared with different criteria and relative weights (Fig. 5).

Table 7 weights for factors for the agricultural and the built-up
Full size table
Fig. 5
figure 5

Suitability maps for a BLT class; b AGRI class; c DSRT; d WTR class

Full size image

The transition matrix shows how many cells are expected to change from one LULC category to another over time. According to the transition probability, there are moderate chances that the AGRI will change into other classes, but it has an 80.81% chance of staying AGRI. As illustrated in Table 8, DSRT has a probability of 82.70% of staying unchanged. WTR has a probability of 78.95% of staying unchanged. The BLT class has a probability of 55.25% staying unchanged.

Table 8 Transition probabilities matrix for predicted LULC of 2020 using 2000 and 2010
Full size table

4.4 CA-Markov model validation

The CA-Markov model was used to predict future LULC maps. The predicted LULC was predicted and validated using Terrset 2020. The predicted LULC 2020 was compared to the actual LULC 2020 using the validated tool. Table 9 shows the kappa index values. The values are greater than 0.80, indicating a significant level of agreement between the simulated and real LULC 2020. Table 10 shows the difference between the actual LULC 2020 and the simulated LULC 2020 in km2. In the present study, the CA-Markov model was confirmed to be able to simulate future LULC situations.

Table 9 The findings of Kappa index
Full size table
Table 10 Details the different between actual and simulated LULC 2020 in km2
Full size table

4.5 LULC transition probabilities for the year 2030 (Predicted)

According to the statistics in Table 11, the AGRI and BLT have a greater chance of changing into other classes because they have a 79.18% and 62.58% chance of remaining unchanged, respectively. DSRT and WTR classes have almost the same chance to stay unchanged; they have 80.92% and 79.86%, respectively.

Table 11 Transition probabilities matrix for predicted LULC of 2030 using 2010 and 2020 images
Full size table

4.6 Prediction of LULC

Figure 6 represents the simulated LULC in 2030. AGRI, BLT, and WTR classes will increase by 9.68%, 133.62%, and 6.06%, respectively, while DSRT classes will decrease by 17.58% Table 12.

Fig. 6
figure 6

Predicted LULC map for years 2030

Full size image
Table 12 Monitoring changes between actual LULC 2020 and predicted LULC 2030
Full size table

5 Conclusion

Analyzing satellite images is crucial for determining the type, pattern, and intensity of change in different types of LULC, as well as for coming up with different solutions for existing environmental issues. This study focused on monitoring prediction of the spatio-temporal LULC changes over the last two decades considering the transition of socioeconomic factors in El-Fayoum governorate. The satellite images were classified into four major LULC classes (water, desert, built-up, and agricultural) to understand the governorate’s LULC changes and predict the future scenario. Multi-temporal satellite data and GIS techniques were employed to monitor the city’s LULC pattern using the land cover maps of 2000, 2010, 2012, and 2020. The image classification procedures were conducted by SVM algorithm and evaluated its accuracy using kappa statistical parameters. The kappa coefficient ranges from 0.91 to 0.94 which is acceptable results for implementing the classified images through the next procedures. The CA-Markov Model and the multi-criteria analysis model based on geographic information systems (GIS-MCA)—have been integrated into this study using an innovative approach for future prediction of El-Fayoum 2030. The weights of the criteria influencing the spatial suitability map were selected as well using the Analytic Hierarchy Process (AHP). Prediction of LULC for 2020 was conducted and compared to actual LULC, the validation showed the simulation model’s reliability by a strong correlation between the simulated landcover map and satellite-derived map. The CA-Markov model was conducted to predict the future LULC change dynamics in 2030. The results presented in this study showed noticeable changes in the spatial and quantitative distribution of LULC. The result of the LULC change prediction suggests that the urban has extra expansion compared to the continuous increase in AGRI class. The area percentages of AGRI, DSRT, BLT, and WTR in EL- Fayoum in 2030 from the total area of governorate are 29.19%, 51.11%, 13.98%, and 5.72%, respectively. It was detected that DSRT will decrease by 17.58% based on the expansion of BLT (133.62%) and AGRI (9.68%) compared to LULC2020. This reflects the requirements of updated planning for future scenarios of increasing the AGRI class. LULC analysis findings might influence sustainable land use management and planning. Our future objective is to integrate the socioeconomic variables and the significance of the policies that affect urbanization.