Regional mapping and monitoring land use/land cover changes: a modified approach using an ensemble machine learning and multitemporal Landsat data

Abstract Regional mapping and monitoring of land use/land cover (LULC) still remain a challenge that depend on classifier and remote sensing data selected. This study aims to create precise LULC maps and explore the efficiency of an ensemble machine learning approach that integrates random forest (RF) and support vector machine (SVM). Two sets of remote sensing data were multi-temporal Landsat and a single scene from QuickBird covering the coastal area of the United Arab Emirates (UAE) were used. By training the classifier using samples collected from QuickBird and knowledge-based and optimal parameterization, the overall accuracy was enhanced from 70% to more than 90%. For the proposed approach, the result showed that the F1-score was 0.99. The results exhibited a rapid increase in all classes, accompanied by a significant change in the shoreline. The proposed approach has the potential to be applied to other regions and to produce accurate LULC maps.


Introduction
Currently, about 50% of the world's population survives in the coastal areas where several cities have developed (Elmahdy and Mohamed 2018).These areas have experienced rapid changes due to a significant influx of population and are therefore subject to rapid land use/land cover (LULC) changes (Ban and Yousif 2016).These rapid LULC changes have a variety of effects on the atmosphere and land surface conditions, and they are important in socioeconomic estimation, climate change modelling, groundwater quantity and quantity, air quality, and landslide land subsidence hazards (Bonan et al. 1992;McPherson 2007;Elmahdy et al. 2022a,b).Consequently, mapping and analysing LULC changes, especially in developing countries with rapid urbanization, intensified LULC changes, and population growth, is required for a variety of purposes (Dutta et al. 2019;Hoan et al. 2018;Elmahdy et al. 2022a;Elmahdy & Ali 2022).CONTACT Mohamed M. Mohamed m.mohamed@uaeu.ac.aeDuring the last decade, regional mapping and monitoring of LULC changes have been carried out using satellite images with low and medium spatial resolutions such as Landsat, Satellite Pour l'Observation de la Terre, or Satellite for observation of Earth (SPOT), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Indian Remote Sensing (IRS) sensors, and Moderate Resolution Imaging Spectroradiometer (MODIS) (Stefanov and Netzband 2005;Mas et al. 2017;Toure et al. 2018).
With the wide availability of remote sensing data, the importance of remote sensing has increased in all applications, especially for urban planning purposes (Edgeworth 1908;Elmahdy et al. 2022c).
Locally, a limited number of studies have been done using machine learning and remote sensing for regional mapping and monitoring changes in LULC (Moore et al. 2015).Most of these studies have been conducted based on field surveys, visual interpreting, and manual screen digitizing, which introduce a level of bias and can be subjective (Elmahdy and Ali 2022).Globally, several methods have been developed to map LULC since the launch of Landsat in 1972.These include statistical classifiers such as maximum likelihood (Edgeworth 1908), Mahalanobis distance (Mahalanobis 1936), and parallelepiped (Goodenough and Shlien 1974).These parametric methods are based on a statistical calculation of how surface reflectance should be distributed.Thus, they have been widely used to classify LULC from Landsat images due to their simplicity.After that, a new generation of complex classifiers based on a myriad of various architectures, including hyperplanes such as support vector machine (Cortes and Vapnik 1995), decision tree such as random forest (Breiman 2001), and neural network such as deep learning (Bengio and Courville 2013).
The most commonly used machine learning methods are supervised learning, which includes support vector machine (SVM) (Adam et al. 2014;Pham et al. 2019;Nhu et al. 2020a,b), random forest (RF) (Nhu et al. 2020a), classification and regression tree (CART) (Felic ısimo et al. 2013), radial basis function (RBF) (Chen et al. 2018), alternating decision tree (ADTree) (Bui et al. 2019).These machine learning algorithms have involved solving problems and can efficiently handle regional remote sensing data as well as outperform the parametric classifiers with higher accuracy (Zhong et al. 2019;Elmahdy et al. 2022a,b).
Other studies have been applied to compare the performance of FR and SVM algorithms against other algorithms such as logistic regression (LR), artificial neural network (ANN), and logistic model tree (LMT) (Bui et al. 2018;Pandey et al. 2020).
A comparative assessment of different sensors and a combination of remote sensing data has been achieved to evaluate the impact of spatial and spectral resolution on the classified LULC maps.
Ensembles of machine learning algorithms generate one optimal predictive model and can yield enhanced outputs over a single model by combining their powers (Liu et al. 2019).It provides a set of state-of-the-art classifiers for pixel-based classification that can be used in several applications, especially LULC classification and natural hazard prediction (Pandey et al. 2020).
Though these methods are well documented and widely used in the literature, a limited number of studies have been conducted to explore the efficiency of using ensemble machine learning when classifying satellite images with a time span of 10 years over a regional scale in a hyper-arid region and compare the performance of an ensemble approach against a single algorithm.Hence, this study aims to explore the efficiency of applying ensemble machine learning for LULC mapping and multitemporal images at a regional scale (country level) and compare the efficiency of the ensemble approach and the single algorithm for LULC classification.
The proposed approach, which merges RF and SVM algorithms, was applied to the coastal area of the United Arab Emirates.This study is a combination of exploring and comparing the efficiency of an ensemble approach and a single algorithm, which is a pioneering step in the applications of Landsat images and advanced machine learning algorithms.

Study site
The study site (coastal strip) stretches from longitude 52 12'15"E to 56 01'0"E and latitude 23 51'00"N to 25 60'30"N and has an area of about 7427.89 km 2 and includes the coastal area of the Emirates of Abu Dhabi, Dubai, Sharjah, Ajman, Umm Al Quwain and Ras Al Khaimah (Figure 1).The study site stretches like a strip in shape starting from Ras Al Khaimah in the northeast to Al Sal'a near the border with Saudi Arabia.The study site can be divided into two main parts.The first part is the northern Emirates, which includes the Emirates of Dubai, Sharjah, Ajman, Umm Al Quwain, and Ras Al Khaimah.The second part includes the Emirate of Abu Dhabi, which is comprised of three regions.The first region is Abu Dhabi Region, which includes the nation's capital Abu Dhabi, the Eastern region, where Al Ain City lies, and the Western Region, where Liwa village lies (Figure 1).Here, our interest focuses on the coastal area of the capital of Abu Dhabi and the Western Region (Figure 1).
The climate of the UAE is categorized as hyperarid and the highest temperature and the lowest precipitation rate were reported in the southern and western parts of the UAE, while the highest precipitation rate and lower temperature values were reported in the Emirates of Fujairah and Ras Al Khaimah (Figure 1).Topographically, the UAE is interrupted by the Musandam Peninsula heights in the north and the Oman mountains in the east (Figure 1).Geomorphologically, the study site is comprised of evaporites and salt ponds in the west, alluvial deposits in the middle, and low sand dunes in the east.Most of the agricultural area is limited to alluvial plains and sand dune corridors, while the built-up area is concentrated in the coastal areas.The green areas such as gardens and parks are distributed among built-up areas.About 90% of the built-up area is concentrated along the coastal area.

Datasets and preprocessing
Two remotely sensed data were used in this study.The first dataset was a set of multitemporal Landsat images with a time span of 10 years.This includes Landsat Thematic Mapper (TM) acquired on 23rd August 1990, the Landsat Enhanced Thematic Mapper (ETMþ) acquired on 23rd August 2000 and 19th August 2010, and the Operational Landsat Imager (OLI) Landsat 8 acquired on 15th August 2020.Landsat sensors collect multispectral data with 8-13 bands in the visible (VNIR), and near-infrared and shortwave infrared (SWIR) regions of the spectrum (Table 1), every 16 days.These include three scenes with path/row coordinates 181/24, 181/25, and 181/26 covering the test site region.The sensor characteristics and information about satellite images are described in Table 1.All Landsat images (Table 1) were downloaded for the same month with less than 5% cloud coverage for reducing errors and ensuring better classification accuracy (Foody and Mathur 2006;Elmahdy et al. 2020a).
We used Landsat images from 1990 to 2020 because of the bad quality, low coverage, coarse spatial resolution, and poor availability during the period from 1972 to 1989 (Vogelmann et al. 2016;Elmahdy et al. 2022c).Additionally, these datasets were the most appropriate remote sensing data to map and monitor changes in LULC over thirty years and were used widely in the literature (Vogelmann et al. 2016;Elmahdy and Mohamed 2018;Elmahdy et al. 2022a,b).The second dataset was a single of QuickBird image with a spatial resolution of 0.6 m acquired on 21st August 2020 (Table 1).The latter dataset was employed to collect the training datasets or regions of interest (ROIs) and check visually the accuracy of the LULC map.
As a first step of preprocessing, all layers were stacked, and the images were re-projected to UTM datum WGS zone N 39 and N40 and registered as an image to image with an RMSE of less than 0.6 (Foody and Mathur 2006).
The major issues with the acquired remote sensing data are that there is irregular coverage over the entire region and that there is missing data due to cloud cover and object shadows.The atmospheric correction was performed by Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) implemented in Envi v. 4.6 software.This process consists of radiometric calibration and dark subtraction.In radiometric calibration, beta nought calibration, all DN values were converted into the top of atmosphere (TOA), reflectance.TOA was performed using four parameters, namely calibration type (reflectance), output interleave (BSQ), output data type (float), and scale factor value of 1.In dark object subtraction, TOA was converted into surface reflectance (SR) using band minimum.
For ensuring better classification accuracy, the following steps were performed: i. Conversion of digital numbers (DNs) values to the TOA reflectance values using conversion coefficients in the metadata file.ii.Conversion from TOA reflectance to surface reflectance (SR) using the Simplified Model for Atmospheric Correction (SMAC) (Rahman and Dedieu 1994).iii.Detection of shadows and clouds was achieved using the Fmask algorithm (Zhu and Woodcock 2012).To do this, an application available on the webpage of Google Code (Google Code Archive: Long-term storage for Google Code Project Hosting.) was used.iv.Filling missing pixels due to shadows and clouds and reconstruction of the Landsat images using self-organizing Kohonen maps (SOMs).
The atmospheric correction was performed using Envi v. 4.5 software and the source code acquired from the webpage of multitemp (S eries Temporelles -Time series of satellite images: news, tools, applications (obs-mip.fr).

Collecting of training datasets
Collecting training datasets is an important step in producing a higher quality LULC classification, especially when the classifiers are trained using training datasets collected from satellite images with a higher spatial resolution (Elmahdy and Mohamed 2018).Training samples were performed using a straight random sampling or proportional method (Van Niel et al. 2005;Elmahdy and Mohamed 2018).This method was chosen due to its ability to reduce error and bias during the classification process.This technique divides the population into homogenous groups and generates training sample sizes that are directly related to the scale of the LULC classes.
The frequency of collected training samples was related to the 30 m pixel size of Landsat images and varied according to the spatial distribution and cluster of LULC classes (Van Niel et al. 2005).The number of collected samples was 5-7 per pixel for the area of the high cluster of LULC classes and decreased to 1-3 per pixel for distributed patches of LULC classes (Van Niel et al. 2005;Elmahdy and Mohamed 2018).After that, all training samples were draped over the Landsat images for 1990, 2000, and 2010, followed by subtracting (selecting) the related training samples from the collected training samples (Elmahdy et al. 2022a,b).The positive samples, which agree well with the features in the image were selected, while the negative samples were excluded.The collected training samples were selected and checked using visual interpretation and knowledge of the authors who live in the study area since 1990.

Ensemble machine learning for regional LULC classification
To regional map LULC and solve the problems related classification process, an ensemble approach was proposed (Figure 2).The proposed approach merges two of the most commonly used in the literature; namely random forest (RF) and support vector machine (SVM).Firstly, RF is an ensemble machine learning supervised classifier that conducts a number of decision trees at randomly selected features and predicts the classes of a test instance by voting of the individual trees.It is ensemble machine learning that performs classification and regression tasks creating a precise classification all the time owing to its simplicity (Breiman 2001).The model utilizes a random selection of variables to predict and then divide each node of each tree followed by developing each tree to reduce errors in the classification process.However, this type of selection affects the classification process and thus produces a very unsuitable single-tree classification.
For better accuracy, we tested several values for the RF variables such as the optimal number of variables (mtry) and the number of trees (ntree).The tested values were 200, 400, 600, and 1000, and 1, 2, 3, and 5 variables at each split to create a stable model.The number of trees (mtree) value of 2 combined with a ntree 200 is the optimal value.Both RF and SVM algorithms were then applied to each Landsat image in the R package and PythonTM, respectively.RF code was sourced from Scikit-learn (Abraham et al. 2014).
The support vector machine (SVM) of Vapnik (1979) is a supervised non-parametric statistical learning algorithm.The SVM is a binary linear classifier that assigns a particular test and samples a class from one of the two possible labels (Zhu and Blumberg 2002).The algorithm assigns each pixel as a vector in a multidimensional system, where each class has a pixel to create the class margins (Mountrakis et al. 2011).During the classification process, SVM ensures that it selects the proper hyperplane that is at a maximum distance from the nearest data points in the two regions.Finally, the individual pixel derived from remote sensing data is labelled and represented as a pattern vector for each image band (Cortes and Vapnik 1995).
As mentioned in the introduction section, there are several machine learning algorithms used in the litterateur.We selected the most applied machine learning algorithms to design an ensemble approach for regional mapping LULC such as RF and SVM (Nhu et al. 2020a,b;Breiman 2001).However, every algorithm has its advantages and disadvantages (Mohamed and Elmahdy 2017;Sahin et al. 2018;Elmahdy & Ali 2022).Hence, an ensemble of machine learning, as a technique, can improve regional mapping and produce more accurate results than a single classifier, especially when merged effectively (Liu et al. 2019).For assessing the accuracy of LULC classification, we tested several values for c, P, C and threshold to all Landsat images.The optimal values were 0.05, 1,100 and 0.03, respectively.
The most common methods for combining different machine learning models are (i) stacking, which builds multiple different types of models, and a supervisor model that learns how to best merge the results of the primary models, (ii) bagging, which builds multiple same-type models from different subsamples of the training dataset, and (iii) boosting, which builds multiple same-type of models with each of them learning to fix the errors of a prior model in the chain.
We applied an ensemble model to find the best fit for two weight values, reduce errors and overfitting and increase the overall accuracy.The designed ensemble model can be defined as the following: where w1 and w2 are real values between 0 and 1.This process builds a model on top of the RF and SVM based on the verified dataset, which is implemented in the R package through the 'glmnet'.Finally, to identify LULC class pixels, a threshold of 0.9 was applied to the outputs.
The produced LULC maps consist of five classes: namely residential, industrial, garden/park, farmland, and bare land.

Accuracy assessment and performance evaluation
To assess the accuracy of the produced LULC classification maps, two methods were performed.In the first method, all LULC maps were visually assessed using QuickBird images with a spatial resolution of 0.6 m and LULC maps produced by Esri with a spatial resolution of 10 m processed in the Google Earth Engine (GEE) platform (https://lcviewer.vito.be/download).The visual inspection was performed by comparing the textural features evident from the proposed ensemble approach and those produced by Esri with a spatial resolution of 10 m and determining whether these patterns were different until satisfactory LULC maps were obtained (Elmahdy et al. 2021a(Elmahdy et al. , 2021b;;Amani et al. 2021;Amani et al. 2022).In the second method, a statistical accuracy assessment using a set of confusion matrices was performed.
The confusion matrix represents the ratio of classified and training samples collected from QuickBird images and field observation.After that, a kappa, user's, and producer's accuracy were then calculated and kappa analysis was analyzed.All LULC maps were standardized and compared based on a pixel by pixel producing numerical values for LULC classes commission, classes omission, total incorrect pixels, percentage of incorrect pixels, precession, recall, and F1-score (Raschka and Mirjalili, 2019;Elmahdy et al. 2022c).These parameters were calculated based on true-positive (TP), true-negative (TN), falsepositive (FP), and false-negative (FN).Accuracy, precision, recall, and F1-score were calculated via the following equations: Where po is the observed agreement ratio and pe is the expected agreement where TP is the true positive; FP is the false positive, and FN is the false negative.
All LULC maps were compared with the collected training samples and the results were presented in a confusion matrix.We assessed the accuracy to check whether different classes in the LULC maps were confused with other classes.

Change detection
Maps of LULC produced using the ensemble approach were converted into class images using Envi v.4.5 software.Each pair of different LULC maps (1990-2000, 2000-2010, 2010-2020, and 1990-2020) was compared using the image differencing method, which is used widely in the literature.The method subtracts the DN value of two spatially registered class images of two different dates.The difference between the DV values can be calculated using the following formula: Where Xij k (T1) and Xij k (T2) are the DN value of pixel X located at row i and column j for and k at time T1 and T2.
If the DN values are either negative or positive, it means the change occurs.Here, we used threshold values evenly spaced between (þ1) and (-1) for the simple difference.In simple difference, the initial state image is subtracted from the final state image).The positive change or a positive value represents the first (n/2) classes, while the negative change or a negative value represents the last (n/2) classes.The no-change class ((n/2 þ 1) represents the middle class.The resulting change detection map consists of different colour codes to facilitate visual interpretation.

Regional LULC classification and change detection
The regional maps of LULC classification along the coastal area of the UAE for the years 1990-2020 are shown in Figures 3, 4, 5, and 6, while the total area of each LULC class in km 2 is graphically represented in Figure 7 and summarised in Table 2.After collecting training samples and achieving optimal parametrization, we ran the algorithms on all Landsat images, and LULC classification maps were produced.The five LULC classes detected during the classification process were residential area, industrial area, garden/park, farmland, and bare land.
Compared with 2020 QuickBird images, the proposed approach showed a strong ability to detect all LULC classes and discriminate built-up and vegetation areas.For instance, the proposed approach precisely detected a highly clustered urbanization observed onshore and offshore.Onshore, intensive residential and garden/park areas were observed on both sides of the Khors (creeks) and waterfront areas.This is because these Khors have played a vital role in trading and commercial exchanges, in the UAE since 1972.They are limited to a narrow strip with 424 km length and 20 km width along the coastal area, while farmland is spatially distributed within sand dune corridors (east).Offshore, there are several artificial islands shaped like continents and palm trees.
The maps show that there is a large portion of intensive urbanization (2,778 km 2 ) along the northeast-southwest trending coastal line.Furthermore, vegetation areas (garden/park and farmland) and built-up areas (residential and industrial) in the study area have grown widely during the last 30 years.
The graphical representation and estimates of LULC's total area (in km 2 ) are shown in Figure 7 and Table 2. Residential and industrial areas are the most dominant LULC classes in the northern part of the UAE (Figures 3 and 4) and the Emirate of Abu Dhabi (Figures 5 and 6), while farmland is the least dominant LULC class classified using the proposed approach.They cover about two-thirds of the total area of the LULC area, while the vegetation and industrial areas are the most dominant LULC classes in the coastal area of the western region of the Emirate of Abu Dhabi.
The residential area covers an area of about 909.11 km 2 in the NUAE, 313.35 km 2 in the capital of Abu Dhabi, and 31.82km 2 in the coastal area of the western region of the Emirate of Abu Dhabi (Figures 5 and 6 and Table 2).The garden/park covers an area of about 54.91 km 2 in the NUAE (Figures 3 and 4 The spatial analysis shows that the residential, industrial, and garden/park areas are more prominent and well-discriminated in the LULC maps obtained using the proposed ensemble approach, while they are least prominent in the maps obtained using the SVM.On the other hand, farmland is more prominent in the LULC maps obtained using SVM.Overall, the LULC classes are excellently classified in maps produced using the ensemble approach.The results also show that there is a clear rapid change and distinction between vegetation and built-up areas.But, some portions of roads and built-up areas were misclassified due to the mixed pixels of small houses and their adjoining areas of vegetation.

Change detection and analysis
The detected from four LULC maps, is presented in Figures 8, 9, 10 and 11, while the total area (in km 2 ) of each LULC class changes is graphically presented in Figure 12 and listed in Table 3 to quantify each class's change from 1990 to 2020.
The results show that there are two extensions observed onshore and offshore.In the onshore area, two urban growth areas were observed: the first one involved the desert areas in the northeast-southwest direction, while the second was in the east direction.
Figures 8, 9, 10, and 11 show the overall changes in LULC classes over the last thirty years, as well as specific changes in LULC classes.In accordance with these figures, rapid changes in LULC can be observed among the four LULC change maps.From 1990 to 2000, the residential area increased from 282.39 km 2 (3.8%) to 457.74 km 2 (6.16%), with a positive change of 175.35 km 2 (2.36%), while the garden/park area increased from 27.73 km 2 (0.37%) to 79.09 km 2 (1.06%), with a positive change of 51 km 2 (0.58%).Similarly, farmland area increased from 7.26 km 2 (0.09%) to 17.09 km 2 (0.23%), with a positive change of 9.83 km 2 (0.14%).The assessment of changes in the periods 2000-2010 and 2010-2020 confirmed that net vegetation area (garden/park and farmland) increases were more pronounced from 2010-2020 than from 2000-2010, increasing from 242.95 km 2 (3.26%) to 430.25 km 2 (5.79%), respectively.Overall, all LULC classes showed continuous increases and upward trends across the study area during the period from 1990 to 2020 (Figures 8, 9, 10 and 11 and Table 3).

Accuracy assessment and performance evaluation
The results indicate that the proposed ensemble approach successfully detect all LULC classes with a higher F1-score of more than 0.99 (Figure 13).As can be seen in Table 4, the best overall accuracy obtained was 95.26% (a Kappa coefficient of 0.91) for the ensemble approach.In the 1990 and 2000 confusion matrices, the overall accuracy of the 1990 and 2000 maps was 90.18% (kappa coefficient of 0.81) and 91.33% (kappa coefficient of 0.88), respectively.Conversely, the user's accuracy and the producer's accuracy for some LULC classes were mostly higher than the overall accuracy.For example, the accuracy of the industrial area is higher than the overall accuracy.The overall accuracy of the 2010 and 2020 LULC maps was 93.33% (Kappa coefficient of 0.89) and 95.26% (Kappa coefficient of 0.91), respectively.As observed in Table 4, the 2010 and 2020 LULC maps are well-classified and quite better than those of 1990 and 2000.
In addition to the accuracy assessment, a visual inspection was performed by comparing the textural features evident from the proposed ensemble approach (2020 LULC map) and the 2020 LULC map produced by Esri and determining whether these patterns were different.All of the LULC classes in the obtained LULC maps agreed well with those produced in Google Earth Engine (EEG) using Sentenil-2 with a spatial resolution of 10 m.However, some errors and confusion were reported in the LULC maps obtained using the proposed approach.These include the spectral signature of the building shadows and vegetation and discrimination between the building roofs and the Sabkha (evaporite) area.
Next, we evaluated the performance of the individual algorithms and the ensemble approach using precision, recall, and the F1 score.Figure 13 shows different precision, recall, and F1 values for the SVM (Figure 13a), RF (Figure 13b), and the ensemble approach (Figure 13c).The ensemble approach yielded the highest values for recall (0.96), precision (0.98), and F1-score (0.99), while SVM presented the lowest value (0.6) for F1score.Slight differences in the F1-score between RF and the proposed approach were observed.This difference appears to be due to the difference in precision and recall scores, and this difference is due to the difference in overall incorrect pixels between SVM, RF, and the ensemble approach.The ensemble approach showed a stronger ability to discriminate between all LULC classes compared with RF and SVM.The proposed approach and individual algorithms, however, demonstrated less ability to distinguish cropland and palm trees than visual inspection and field observation.Although the Landsat images have the same spatial resolution, the ASTER and Sentinel-2 images lack the regular time span and geographical coverage of the study area.

Regional LULC classification and change detection
The present study proposes a stacked ensemble machine learning approach for regional mapping and monitoring LULC changes.The proposed ensemble approach was applied to one of the most developed regions in the world.The study starts by collecting training samples from satellite images with a higher resolution, followed by applying straight random sampling.The method reduces errors and bias during the classification process as indicated by the accuracy assessment (Van Niel et al. 2005;Elmahdy and Mohamed 2018).
Testing of several stacking threshold values shows that the best value was 0.9 to detect LULC pixels.The stacking of machine learning was the best and recommended by  Wolpert (1992) and has been extensively used in a variety of applications (Ekbal and Saha 2013;Haralabopoulos et al. 2020;Zhang et al. 2021) due to its simplicity and reliability.
The results of the present study suggested that within the fields of LULC classification and monitoring changes, where built-up and vegetation areas are the major objectives of classification, and it is recommended to use free-of-charge remote sensing data with a moderate spatial resolution (20-30 m) and an ensemble machine learning approach, especially when the research budget is not enough.Unlike traditional studies of manual screen digitizing and visual interpretation, which introduce bias and errors, the proposed approach delivers a better result, especially when trained by datasets collected from QuickBird images and optimally parameterized (Elmahdy and Mohamed 2018;Elmahdy et al. 2020b).
Compared with the spectral incidences (NDVI, NDBI, and NDWI) and individual algorithms, the ensemble machine learning approach has the best-fit classifier for LULC classification over a regional scale with a low-cost and time-consuming manner.(Ramanath et al. 2019;Zhang et al. 2022;Elmahdy and Ali 2022).They concluded that the ensemble machine learning had the highest ability to discriminate between healthy vegetation and weakened vegetation.The use of ensemble machine learning presents a better result for LULC classification with a smaller regression error (Bunting et al. 2018).Thus, it has been widely used in a variety of applications (Ekbal and Saha 2013;Haralabopoulos et al. 2020;Zhang et al. 2021).The proposed approach is innovative as it avoids overfitting, minimizes bias, and fits unseen patterns without affecting its performance.

Accuracy assessments and performance evaluation
Following the accuracy assessment, built-up, farmland, industrial, and bare land classes showed high producer accuracy in all of the confusion matrices.This is explained by the unmixing of pixels being higher than those of lower producer's accuracy, such as garden/park classes.The highest overall accuracy (95.26%) was calculated by applying the proposed approach to atmospherically corrected 2020 Landsat images, while the lowest overall accuracy (90.18%) was obtained by applying the proposed approach to 1990 Landsat images.This variation appears to be due to the technical characteristics of the Landsat sensors and atmospheric variations (Erbek et al. 2004;Elmahdy et al. 2022a,b).Other factors, such as the number of unmixing pixels and the pixel size of training samples, had an important impact on the accuracy of LULC maps (Ballabio and Sterlacchini 2012;Bui et al. 2018;Pandey et al. 2020).
The efficiency of the proposed approach was estimated based on the performance encountered by the recall, precision, and F1-score methods.For the proposed approach, we found that the recall, precision, and F1-score were 0.96, 0.98, and 0.99, respectively (Figure 13).The F1-score of the proposed approach was slightly higher compared to RF and much higher compared to SVM, which meant the proposed approach fit much better than a single classifier.
Our findings concluded that the proposed approach had the most reliability with the spatial resolution of the Landsat images and the reality of the study area.This finding is consistent with the main findings of other studies (Adam et al. 2014;Abedini et al. 2019;Pandey et al. 2020) which compared the performance of the support vector machine (SVM), random forest (RF), Bayesian logistic regression (BLR), kernel logistic regression (KLR), naive bayes tree (NBT), and alternating decision tree (ADTree) and concluded that the SVM and RF algorithms outperformed the other algorithms.
The lowest user accuracy across all four confusion metrices was for farmland appears due be to the difficulty discrimination this class from garden/park class and low sampling number as well as different in sensor sensitivity and characteristics.In other talks, green area was composed mainly of garden/park and farmland was visually similar and some portions (palm trees) still difficult to classify due to similar spectral signatures (Elmahdy and Ali 2022).Indeed, Pandey et al. (2020) and Elmahdy et al. (2022) concluded that the performance of RF was much better than SVM and ensemble approach was the most efficient approach for most applications and outperformed single classifier, including RF and SVM.However, Khatami et al. (2016) found that SVM was the most efficient classifier for most remote sensing applications and outperformed several algorithms, including, RF, decision trees and neural networks.
Our results agree well with Prasad et al. (2022) who concluded that high spatial and spectral resolution of multi-sources and multi-temporal remote sensing are required for a better mapping of LULC.RF and SVM achieved high and similar predictive overall accuracy, especially when the training samples and classified LULC belong to the same region (Diengdoh et al. 2020).On the other, this accuracy is lower and different when training samples and classified LULC belong to different geographic regions.
Despite the high accuracy and performance of the proposed approach, there were a few errors, such as the misclassification of shadows and wetlands as vegetation.Visual inspection shows that there is confusion between building shadows and wetlands and vegetation, as has been observed in the LULC maps.Building shadows and wetlands have similar spectral signatures (Elmahdy et al. 2022b).Another confusion was reported between the spectral signature of the built-up area and that of coastal sabkha (Elmahdy et al. 2022b;Elmahdy and Ali 2022).Small water bodies, such as swimming pools and fountains within the built-up area, were misclassified.

Limitations and recommendations
Although we found relatively high overall LULC classification accuracy, some limitations and uncertainties persisted.Multitemporal Landsat images used in this study with a spatial resolution of 30 m, introduce a source of bias and misclassification in the LULC classification and change analysis.More specifically, Landsat sensors failed to discriminate between palm trees and cropland in the vegetated areas.The use of cloud-free Table 4. Accuracy assessment of the resulting LULC maps using an ensemble approach.multitemporal Landsat images acquired from sensors with different technical characteristics over thirty years introduced another source of error and bias.It was therefore difficult to conduct field observation in remote and inaccessible locations such as the western region of the study area.Some portions of small buildings, such as villas and small road networks, were misclassified due to mixed pixels of intensive vegetation cover within built-up areas.Future studies will be carried out using Sentinel-2 with a spatial resolution of 10 m and synthetic aperture radar (SAR) data, which could increase the overall accuracy of the resultant LULC maps and the analysis of change detection (Elmahdy and Ali 2022).

Conclusion
In this study, we applied an ensemble machine learning approach for regional mapping and monitoring LULC from multi-temporal Landsat images.The use of multitemporal Landsat images has permitted better regional mapping and monitoring of LULC changes over the past thirty years.The proposed approach combines different machine learning algorithms and utilizes a stacking ensemble strategy.A stacked ensemble with a threshold of 0.9, which is widely used in a variety of applications, was found to be the most suitable and effective method.The produced LULC maps had an overall accuracy of more than 90%, and the change detection analysis revealed that the study areas had experienced intensive urbanization since 1990, particularly in the northern emirates and around Abu Dhabi.The highest expansion in the areas of the LULC classes in the northern Emirates was reported for residential areas followed by the industrial area, while the highest expansion in the area of the LULC classes in the capital of Abu Dhabi was observed for residential areas followed by garden/park areas and farmlands.A large portion of the bare land (sand dunes and Sabkhah) was additionally converted into residential, industrial and farmland areas.These results confirm that the proposed approach can successfully be applied for regional mapping and monitoring LULC changes in a similar region.The resultant maps provide extra information on a regional scale that could not be collected using traditional methods of field surveys.The new LULC maps can be used in further studies to investigate the impact of rapid urbanization on groundwater and air quality as well as future conservation efforts.

Figure 1 .
Figure 1.The western coastal area and regions of the United Arab Emirates.The red polygon highlights the study site.

Figure 2 .
Figure 2. Flow chart of the overall methodology.
) and 130.47 km 2 in the capital of Abu Dhabi.In general, built-up and vegetated areas are the dominant LULC classes in the study area.They gradually decrease in the directions from the Emirates of Dubai and Abu Dhabi to the NNE-SSW.Between them, there are two gaps.The first gap is between the Emirates of Dubai and Abu Dhabi, with the second gap extending from the capital of Abu Dhabi to the western region of the Emirate of Abu Dhabi.

Figure 3 .
Figure 3. LULC maps produced using the ensemble approach for the years from 1990-2020 along the coastal area of the northern Emirates (from Umm Al Quwain to Ras Al Khaimah).

Figure 4 .
Figure 4. LULC maps produced using the ensemble approach for the years 1990-2020 along the coastal area of the northern Emirates (from Dubai to Umm Al Quwain).

Figure 5 .
Figure 5. LULC maps produced using the ensemble approach for the years 1990-2020 along the coastal area of the capital of Abu Dhabi.

Figure 6 .
Figure 6.LULC maps produced using the ensemble approach for years from 1990-220 along the coastal area of the western region of the Emirate of Abu Dhabi.

Figure 7 .
Figure 7.The graphical representation of the total area (in km 2 ) of built-up area (residential and industrial) and vegetation area (garden/park and farmland) in the northern Emirates (a), the capital of Abu Dhabi (b), and the western region of the Emirate of Abu Dhabi during the period from 1990 to 2020.

Figure 8 .
Figure 8.The overall change in the areas of built-up (a), industrial (b), and vegetation such as gardens/parks (c) along the coastal area of the northern Emirates (from Umm Al Quwain to Ras Al Khaimah) during the period from 1990 to 2020.

Figure 9 .
Figure 9.The overall change in the areas of built-up (a), industrial (b), and vegetation such as gardens/parks and farmland (c) along the coastal area of the northern Emirates (from Dubai to Umm Al Quwain) during the period from 1990 to 2020.

Figure 10 .
Figure 10.The overall change in the areas of built-up (a), industrial (b), and vegetation such as gardens/parks and farmland (c) along the coastal area of the capital of Abu Dhabi during the period from 1990 to 2020.

Figure 11 .
Figure 11.The overall change in the areas of built-up (a), industrial (b), and vegetation such as gardens/parks and farmland (c) along the coastal area of the western region of the Emirate of Abu Dhabi during the period from 1990 to 2020.

Figure 12 .
Figure 12.The graphical representation of the total area (in km 2 ) of built-up (residential and industrial), and vegetated areas (garden/parks and farmland) changes (a), and their differences from 1990 to 2020 in the NUAE (a), the capital of Abu Dhabi (b), and the western region (c).

Table 1 .
Characteristics of Landsat images used in the study area.

Table 2 .
Estimates of LU/LC class's total area of each region from 1990 to 2020.

Table 3 .
Estimates of LU/LC class's total area and percentage and changes from 1990 to 2020.