LAND USE AND LAND COVER SIMULATION BASED ON INTEGRATION OF ARTIFICIAL NEURAL NETWORKS WITH CELLULAR AUTOMATA-MARKOV CHAIN MODELS APPLIED TO EL-FAYOUM GOVERNORATE

: The paper aims to evaluate the effectiveness of the multi-layer perceptron-Markov chain analysis (MLP-MCA) integrated method in predicting future Land Use and Land Cover (LULC) change scenarios in Fayoum due to rapid urbanization. The study employed machine learning algorithms for image classification using Google Earth Engine (GEE) for classification techniques to derive LULC maps from Landsat imagery taken in 2001, 2011, and 2021. The 2001 and 2011 LULC maps were used to predict the LULC scenario for 2021 using MLP-MCA, and the predicted result was validated against the observed 2021 LULC map using Area under the curve (AUC) that was derived from the Receiver Operating Characteristics (ROC). Subsequently, the study predicted future LULC changes for 2031 using two sub-models; sub-Agri and sub-built. The results show that a rapid growth in both built and agricultural area. The findings of this study highlight the potential of the MLP-MCA method in predicting future LULC changes due to urbanization.


INTRODUCTION
Land cover change is the term used to describe physical alteration of the Earth's surface, including changes in water sources, soil, and air pollution.While the land use alteration is often associated with how human behavior affects physical changes on the Earth's surface.Land use change typically has some impact on land cover change (Aburas et al., 2018).LULC Sustainability is largely concerned with changes in the environment.Human interference in the environment usually leads to changes in land use, such as deforestation, destruction of pastures, urban growth, and shrinkage of waterways (Kourosh Niya et al., 2020).These changes are leading to a shortage of natural resources, shortages in food production, and social and political impacts (Palmate et al., 2017).
Remote sensing (RS) and geographic information systems (GIS) have demonstrated great potential in studying landscape dynamics and offer provide valuable methods in the broader scope of environmental tracking (Armin et al., 2020).It is critical to simulate LULC to comprehend and recognize the potential future changes in each class of land.Planners of land use, environmentalists, and resource managers can benefit from these forecasts in order to reduce the bad impacts of planning to achieve better manage for agricultural land protection , natural resources and urban areas (Kourosh Niya et al., 2020).
In recent years, a various spatial models involved RS and GIS have been developed to predict LULC in future, such as Artificial Neural Network (ANN) model (Abbas et al., 2021), cellular automata (CA) model (Deep and Saklani, 2014), and the Conversion of Land Use and its Effects (CLUE) model (Jafarpour Ghalehteimouri et al., 2022), as a substitute, each model has drawbacks (Triantakonstantis and Mountrakis, 2012).To overcome the drawback of existing models, integration of several models had been adopted.CA-ANN is created by integrating a CA with ANN model (Abbas et al., 2021).The models provide resources for discovering spatial variation in LULC (Munthali et al., 2020).In comparison with the other models, CA-Markov Model (CA-MM) has the highest accuracy, because it can be combined many factors which effect on LULC prediction such as; biophysical, socio-economic, and remotely sensed and geospatial data (Naboureh et al., 2017).CA-MM can succefuly simulate complex patterns (Hyandye and Martz, 2017).Regmi et al. (2017) compared the CA-Markov model with the GEO-MOD model for LULC future simulation and concluded that the CA-MM was more successful in future prediction than GEO-MOD.Therefore, we adopted it for our prediction scenario in El-Fayoum Governorate.
The hybrid Cellular Automata-Markov Chain (CA-MC) model is a commonly used and effective method for modeling the spatiotemporal changes in LULC (Fitawok et al., 2020).The CA-MC model combines the stochastic model of CA and the technique of Markov Chain (MC) to simulate multi-directional LULC changes analysis and provide ways for projecting different future scenarios (Keshtkar and Voigt, 2015).However, Variables with explanatory power such as socioeconomic and environmental data needs to be taken into account when modeling LULC changes (Mas et al., 2014).The integration of the CA-MC model with other models, such as the Multi-Layer Perceptron Neural Network (MLP-NN), has helped to improve its prediction capability (Gharaibeh et al., 2020).MLP-NN is an effective approach that introduces a better understanding of the land change process and provides more accurate results depending on artificial intelligence.Three layers make up MLP-NN, including input, hidden layers, and output, and predicts geospatial changes with the aid of previous changes (Simwanda et al., 2021).MLP-NN has the highest capability to generalize the transition potential through the backpropagation algorithm (Dey et al., 2021).The integration of MLP-NN and CA-MC (MLP-CA-MC) takes advantage of both models (Gharaibeh et al., 2020;Lamchin et al., 2022;Toma et al., 2023).MLP-NN facilitates the automatic calibration of the CA-MC model and provides a more accurate future change scenario than CA-MC alone.
In this study, the LULC trends have been refined and explained in a way that can assist and guide policymakers, land-use planners, and hydrologists for the future.Hydrology, for example, is affected differently by land use and climate change than by the separate factors, and this limits the ability to develop future water management strategies and irrigation plans.The rest of the paper is organized as follows: Section 2 describes the study area, Section 3 shows our adopted methodology to predict and simulate LULC for EL-Fayoum Governorate.Section 4 introduces the result and discussions, finally Section 5 draws the conclusion for the study.

STUDY AREA AND MATERIALS
El-Fayoum is in a depression in the western desert, a remarkable area.It is far from the Nile River, about 25 kilometers away, and 90 kilometers southwest of the capital of the Arab Republic of Egypt, its green land inside the desert, it draws water from the Nile via the Yusuf Sea. Figure 1 shows the location of the study area.
In this study, different LULC classes in Fayoum were mapped using digital remote sensing data from the Landsat series satellites.On GEE, the LULC maps were processed and categorized (https://code.earthengine.google.com/).Digital elevation models (DEMs) and road networks are two additional data sets used in this study.The USGS website's ASTER DEM of 30-m spatial resolution data was downloaded and used to produce slope.Google Earth image is used to extract the shape files for the roads and canals.

LULC Image Classification and Assessment
Using imagery from Landsat, LULC classification maps were created.Figure 2 displays the adopted flowchart for image classification.GEE employed three different classifiers named: support Vector Machine (SVM), Random Forest (RF), and Classification And Regression Tree (CART) as shown in Figure 2. Using the cloud mask method offered by GEE, contaminated pixels brought on by cloud cover were removed from all images (Rahman et al., 2020;Atef et al., 2023).
Four major LULC classes -agricultural land, water, bare land, and built-up -were identified and categorized in this study.By using accuracy assessment, the effectiveness and usability of classified images from 2001, 2011, and 2021 were estimated.After classification techniques were performed, an accuracy assessment was conducted to determine their accuracy.There were two types of training datasets: training and validation.Seventy percent were used for training, and thirty percent of the trained data were used for testing using a confusion matrix (Hamud et al., 2021).A built-in algorithm in GEE, called a confusion matrix, validates and rates the classification accuracy of the images using the kappa coefficient (κ) and overall accuracy (OA) (Congalton and Green, 2019).

Analysis of LULC changes
The Landsat TM/ETM+/OLI images from the years 2001, 2011, and 2021, respectively, served as the basis for the LULC changes analysis used in this study.Using the integrated software environment known as the land change modeler (LCM) for ecological sustainability, the LULC changes analysis, simulation, and LULC changes prediction were carried out (Leta et al., 2021).The LULC maps from two dates are compared to create the Markov chain matrix that serves as the foundation for the change model.Quantifying LULC changes over the periods of 2001-2011, 2011-2021, and 2001-2021  as well as their contributions to the overall net change, were also examined in graphical form.

Transition potentials
For two sub-models that included specific LULC dynamics in accordance with the observed processes of change.SUB-Built is a sub model included the transition from agricultural and bare lands to built-up.SUB-Agri is a sub model included the transition from bare land to agricultural land (land reclamation).Five explanatory variables are detected form both sub models and used for transition potential map, as shown in Table 1 (Mishra and Rai, 2016).
The transition potential maps were produced using MLP.This method can generate transition potentials across non-linear function simulation algorithms (Saha et al., 2022).The MLP parameters were employed in this study: 10,000 iterations, 50% training and 50% testing samples, 0.001 learning rate, 0.01 momentum factor, and 10 hidden layers.The effectiveness of the transition potentials in both sub-models was evaluated using the accuracy rate and skill measure outcomes of the training methods in the MLP.

Future Simulation and assessment
The MC procedure from the LCM was used to model potential future scenarios, as shown in Figure 3.   2001 and 2021 by 8.96%.

LULC changes analysis using LCM
By using LCM's change analysis tool, the LULC maps of 2001, 2011, and 2021 were also analyzed for periods 1, 2, and 3. Different classes' losses and gains were evaluated when evaluating LULC changes.Both gains and losses are shown in Figure 6. Figure 7 shows the change maps from 2001-2021.During period 1, the bare land has lost 76.

MLP-NN skill measure
MLP-NN is a machine learning techniques that has the ability to effectively model sophisticated patterns and conduct, as previously demonstrated by (Gharaibeh et al., 2020).In this study, two sub-models were developed using MLP-NN, each consisting of one input layer, one hidden layer, and one output layer Table 6 .The input layer contained seven driver variables, while the hidden layer contained ten nodes for both SUB-Built and SUB-Agri.The output layer for SUB-Built and SUB-Agri contained four and two nodes, respectively.The nodes in each layer were interconnected through varying weights.To assess the impact and order of influence of each driver variable, 1,000 iterations were performed, and three sensitivity analyzation techniques were utilized.As detailed in the subsequent subsections.The goal of network training was to determine the optimal weights for the connections between the input and hidden layers, as well as between the hidden and output layers, to enable accurate classification of unknown pixels.
For the generation of transition potential maps (TPM), variables previously tested were applied for both sub-models.Thus, for SUB-built and SUB-Agri, all the variables were applied.With this, the accuracy rate increased to 71.00% and 86.78%, for SUB-built and SUB-Agri, respectively as shown in Figure 8.In this study, the results indicated an acceptable accuracy, despite the lowest value detected for SUB-built.Islam and Ahmed    (2012) suggest that values of accuracy for neural network processes should reach up to 70%; however, this result will depend to the influence of used variables.The accuracies for the submodels of this study were considered effective.TMP revealed that bare land had the highest probability of persistence, and that the LULC class with the lowest probability of persistence was built up Table 4.

Projected of LULC for 2031
After forecasting the LULC 2021 (i.e., 10 years from 2011), based on TPM for the lapse of 2001 (time 1) and 2011 (time 2), as well as the transition potentials of each generated submodel (SUB-Built and SUB-Agri), The ROC result showed that the sub-model AUC was greater in every scene than the random baseline value, indicating successful projections (Figure 9).SUB-Built performance was the best even though it was created with the fewest explanatory variables (AUC = 0.686); however, SUB-Agri performance was significantly worse than those of SUB-Built even though they included more explanatory variables.Then we used 2011 as time 1 and 2021 as time 2 to forecast 2031 based on TPM from 2011-2021 (as shown in Table 5).
When we compared the two sub-models in 2031, using the   changes from 2021, the predicted scenarios revealed the findings in LULC (shown in Table 6).SUB-Agri predicted that agriculture's surface would increase by 4.44%.The class of Bare Land showed the biggest loss (1.84%).SUB-built modeling revealed that agriculture will reduce its surface by 6.47%.

CONCLUSIONS
The Triantakonstantis, D., Mountrakis, G., 2012.Urban growth prediction: a review of computational models and human perceptions.

Figure 2 .
Figure 2. Adopted methodology of image classification.
Figure 3. Adopted Flowchart for future prediction.

Figure 8 .
Figure 8.Each sub-model accuracy rate and skill measure.

Figure 9 .
Figure 9. ROC curve and AUC for two sub-models.
study determined how LULC patterns altered in the Fayoum, Egypt, study area between 2001 and 2021.Agricultural encroachment and population growth are two of the biggest factors affecting land use changes.Recognizing past harmful trends is crucial to improving sustainable urban management and planning.Crop yields and temperatures can be negatively affected by changing LULC patterns.A variety of classifiers on GEE were applied to the LULC classification problem, including support vector machines (SVM), random forests (RF), and classification and regression trees (Cart).SVMs were capable of capturing land cover alterations and trends accurately.In terms of the change rate in agricultural lands from 2001 to 2021, we observe a total change rate of 4.21%.Agricultural land conversion and urban settlement are the primary drivers of land-use change in Fayoum between 2011 and 2021.However, land reclamation in the desert made up for this shortfall.There was a change rate of 50.78% in built-up from 2001 to 2021.Bare land classes showed minimal change rates.Two sub-models (SUB-Agri and SUB-Built) were simulated by CA-ANN in 2031.The SUB-Agri model predicted an increase in agricultural land and a decline in bare land.According to the SUB-Built model, built-up land increased and agricultural land declined.It is especially concerning that urbanization and fragmentation of agriculture may result in sharp changes in LULC that could place the environment, natural resources, and food security at risk.Rapid built-up will significantly affect agricultural production.The results of the spatiotemporal and potential LULC simulation may therefore give decision-makers a greater awareness of how socioeconomic factors affect changes in LULC intensity as well as a more solid foundation for environmental preservation and a more sustainable future.Future research should examine the impact of climatic changes caused by the rapid rate of built-up expansion and the scale of economic gains indicated by built-up expansion and land use permitting.Simwanda, M., Murayama, Y., Phiri, D., Nyirenda, V. R., Ranagalage, M., 2021.Simulating Scenarios of Future Intra-Urban Land-Use Expansion Based on the Neural Network-Markov Model: A Case Study of Lusaka, Zambia.Toma, M. B., Belete, M. D., Ulsido, M. D., 2023.Historical and future dynamics of land use land cover and its drivers in Ajora-Woybo watershed, Omo-Gibe basin, Ethiopia.Natural Resource Modeling, 36(1), e12353.
was done using cross-tabulation analysis.The LULC classes' gains and losses, Table 1.Explanatory variables for sub-models.

Table 2 .
Change Rates of LULC from 2001 to 2021.
There has been a loss of 71.87 km 2 and a gain of 171.10 km 2 in the built-up area with a net gain of 99.23 km 2 .With a net gain of 28.15 km 2 , agricultural land has gained 136.67 km 2 while losing 108.52 km 2 .For period 3, the bare land has lost 188.76 km 2 and increased by 47.62 km 2 , resulting in a net loss of 141.14 km 2 .The net loss of water category is 111 km 2 , with 41.63km 2 lost and 10.23 km 2 gained.There has been a loss of 71.87 km 2 and a gain of 171.10 km 2 in the built-up area with a net gain of 99.23 km 2 .With a net gain of 61.54 km 2 , agricultural land has gained 167.02 km 2 while losing 105.48 km 2 .Agricultural land has gained 167.02 km 2 while losing 105.48 km 2 with net gain of 61.54 km 2 .
26 km 2 and increased by 45.99 km 2 , resulting in a net loss of 30.27 km 2 .The net loss of water category is 14.89 km 2 , with 22.31 km 2 lost and 7.42 km 2 gained.There has been a loss of 90.07 km 2 and a gain of 101.84 km 2 in the built-up area with a net gain of 11.77 km 2 .With a net gain of 33.39 km 2 , agricultural land has gained 122.05 km 2 while losing 88.66 km 2 .While in period 2, the bare land has lost 148.28 km 2 and increased by 37.41 km 2 , resulting in a net loss of 110.87km 2 .The net loss of water category is 16.52 km 2 , with 24.16 km 2 lost and 7.64 km 2 gained.

Table 3 .
Explanatory variables for sub-models.

Table 4 .
TPM of LULC changes from 2001 to 2011.

Table 5 .
TPM of LULC changes from 2011 to 2021.

Table 6 .
LULC Projected for 2031 for Sub-Agriculture and Sub-Building.