Introduction

The dynamic transition in land use and land cover, particularly within and around urban areas, the results of the increasing global populations and urbanization trends at the extent of the vegetated regions, altering ecological ecosystems and microclimates (Basu et al., 2023; Negesse et al., 2024; Patel et al., 2024a). Modeling LULC changes is for monitoring ecological quality, environmental sustainability, and uncontrolled development at various spatiotemporal dimensions (Aboelnour & Engel, 2018; Balogun & Ishola, 2017). As the urban population and their needs and environmental impacts continue to increase, a comprehensive analysis of the past and future LULC scenarios becomes important for a sustainable urban environment (Jawarneh et al., 2015; Faisal et al., 2021). Even though urbanization indicates civilian advancements, it brings short- and long-term consequences in different aspects (Maimaitiyiming et al., 2014). The rapid transformation into urbanized regions is vastly linked with the change of LST, which could interrupt the ecologic and climate balance in the foreseeable future (Kafy et al., 2021; Maimaitiyiming et al., 2014; Mallick et al., 2008).

Rising temperature is a direct impact of changing climate, which in turn associated with consuming green areas by urbanization; urban areas’s temperatures are estimated to be between 2 and 4 °C higher temperatures than those in rural areas (Mishra & Rai, 2016; Pal & Ziaul, 2017; Zhou et al., 2014; Zine El Abidine et al., 2014). Land Surface Temperature (LST) is an indication employed to evaluate the environmental health of a particular system (Kalnay & Cai, 2003; Patel et al., 2024a). However, understanding LST and its attributions, especially in urban areas, has been under investigation for a long time. Several remotely-sensed based indices (i.e., the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Built-up Index (NDBI), the Normalized Difference Water Index (NDWI), and the Normalized Difference Bareness Index (NDBaI)) have been utilized in an attempt to understand the interplay between LULC and LST (Chen et al., 2006; Patel et al., 2024b; Weng & Lu, 2008).

The advancement in deep learning has improved the understanding and acquired higher accuracy of the relationships between past and future land variations (Zhang et al., 2023). Deep learning algorithms are becoming more robust for providing precise future simulations of the landscapes, as ground features could be covered and represented on time (Alkaraki & Hazaymeh, 2023). Multiple classified models (i.e., deep learning) have been widely applied to predict the LULC and LST. In the realm of LULC prediction, numerous models (i.e., Markov-Cellular Automation) have gained prominent use in analyzing LULC dynamics across various scales like (Jawarneh, 2021; Abijith & Saravanan, 2022; Atef et al., 2024; Beroho et al., 2023; Dos Santos et al., 2024; Jawarneh et al., 2024; Taloor et al., 2024). The CA-SLEUTH and Artificial Neural Network (ANN-CA) models also gained good reputation in modeling LULC changes (Ansari & Alam, 2024; Dolui & Sarkar, 2024; Kannapiran & Bhaskar, 2024; Kumar & Agrawal, 2022; Rienow, 2024; Satari et al., 2024; Varquez et al., 2023; Yagci & Iscan, 2024; Yue et al., 2024; Zhang et al., 2023). The CA Markov model reveals a major constraint in its capacity to integrate spatial data, limiting the effectiveness of incorporating geographical information. Although transition probabilities could be accurate for individual categories, there is an ambiguity in assessing the spatial distribution of occurrences within each land use classification, as noted by Ye and Bai (2008). The ANN-CA approach integrates artificial neural networks (ANN) and cellular automata (CA) to generate simulating LULC outcomes. The ANN is an approach of deep learning that demonstrates the ability to adapt the changes in LULC based on a sequence of LULC-classified images and spatial variables (Abbas et al., 2021;Yatoo et al., 2022; Baig et al., 2022). This technique has acquired viral employment to comprehensively understand potential urbanized areas for future and past scenarios.

Previous literature has utilized ANN-CA as a proxy for forecasting changes in the LULC. Likewise, several countries in the Middle East, including Jordan (Gharaibeh et al., 2020), Iran (Mahdavi Estalkhsari et al., 2023), and Saudi Arabia (Bindajam et al., 2021), have implemented the approach of ANN to predict LULC scenarios. The model has also been adopted in various South Asian countries, including China (Wang et al., 2023; Yue et al., 2024; Zhang et al., 2023), Bangladesh (Kafy et al., 2021; Rahman et al., 2017), Malaysia (Baig et al., 2022), India (Ansari & Alam, 2024; Dolui & Sarkar, 2024; Kannapiran & Bhaskar, 2024), and Indonesia (Saputra & Lee, 2019).

To predict LST, several models have been widespread across various scales, including Markov MC (Tran et al., 2017); MLP-MC (Nurwanda & Honjo, 2020); CA-Markov (Feng et al., 2018; Hussain et al., 2024); ANN (Al Kafy et al., 2021; Gupta & Aithal, 2024); ANN-CA (Faisal et al., 2021); Long Short-Term Memory (LSTM) (Badugu et al., 2024; Qi et al., 2024; Yang & Li, 2024; Zhang et al., 2023); and the XGB Regression (Mohammad et al., 2022). Among those models, the LSTM has a more comprehensive use in predicting the LST due to its accuracy on longer pattern changes, primarily to address the difficulties of gradient disappearance and gradient explosion during long sequence training. Hence, the LST simulations avoid the likelihood of over-fitting and the consequent loss of data and precision forecasting (Zhang et al., 2023).

The increasing interest in analyzing past, present, and future changes in LST and LULC is primarily due to their impact on energy plans, environmental sustainability, and resource management, especially in those countries with limited resources. Jordan is a developing country in the Middle East with very limited natural resources. 80% of its land is arid, with most of its population concentrated in the small northwestern region of the country, where the semi-arid climate is dominant (Jawarneh et al., 2024). The country has been experiencing dramatic land use and land cover changes due to the increasing population of refugees from neighboring countries (Jawarneh & Biradar, 2017). The Greater Amman Municipality (GAM) received the most significant number of refugees, mainly from Iraq and Syria (Khawaldah, 2016). In 2014, GAM population was 2,584,600, and in 2015, the number doubled to more than 4,000,000 inhabitants, representing 58% of the total population of Jordan (DOS 2016a; DOS 2016b). The GAM is a crucial area in the Amman governorate since it hosts the primary government agencies and essential services. Furthermore, the region is often regarded as a significant hub for individuals seeking opportunities for investment and employment, which has led to significant alterations in land utilization.

The main objectives of this research were to analyze the trends of change in LULC and LST in the GAM historically and in the future from 1980 to 2023, as well as employ two deep learning-based prediction techniques to predict the LULC changes (i.e., ANN-CA) and LST changes (i.e., LSTM) for the GAM for 2024. The study used various methods that have been frequently utilized to compare their results and evaluate to what extent these techniques can provide sustainable reliability. In addition, the study assessed the effect of change in LULC on LST for the study period. These analyses aimed to facilitate discussion and assist policymakers in planning for the potentially adverse consequences of urban expansion, population growth, and land transformation by forecasting forthcoming scenarios for LULC and LST in the major urban area of Jordan.

Study area

Greater Amman Municipality (GAM) is the major urban hub in Jordan as it compasses the capital city of Amman with an area of ~ 800 km2 (Al-Kofahi et al., 2018; Al-Saad et al., 2023) (Fig. 1). The area of Amman Governorate is ~ 7579 km2, covering about 8.5% of the total area of Jordan (DOS, 2023). GAM’s climate is semi-arid with hot and dry conditions in the summer between May and September and moderate to cold and rainy conditions during the winter season from December to March (Jaber, 2018). Topographically, the terrain significantly varies between 57 and 1102 m.a.s.l. The population in Jordan at the end of 2022 was estimated to be about 11,302,000. GAM is devided into five main districts (e.g., Amman Qasabah, Marka, Quaismeh, Al-Jami’ah, and Wadi Essier district) (DOS, 2023).

Fig. 1
figure 1

The study area location

Full size image

Methodology

A three-step methodology was implemented to predict the changes in LULC and LST and to investigate the changes of the LULC and their impact on the LST in the GAM: (i) data collection and preprocessing; (ii) prediction of the future LULC for 2030; and (iii) prediction of the future LST for 2030.

Data collection and preprocessing

Different data sources were used to collect the variables used to model LULC and LST in the study area (Table 1). The Jordan National Land Cover Dataset (JNLCD) encompasses decadal LULC maps for the whole of Jordan at 30 m spatial resolution for 1980, 1990, 2000, and 2015 (Jawarneh & Biradar, 2017). It was developed from Landsat satellite imageries (Table 2) using a combined classification technique (ISODATA clustering and rule-based). The original JNLCD LULC maps contain nine classes. However, for this study, all agricultural-related classes (rainfed crops, orchards, and irrigated) were aggregated into one category. Following the same classification techniques, we developed the 2023 LULC map using the Google Earth Engine (GEE) platform (https://earthengine.google.com/). Landsat 8-OLS (collection 2, level 2) scene that was acquired in April 2023 to accurately classify vegetated areas during the peak of the growing season (Alkaraki & Hazaymeh, 2023; Hazaymeh & Hassan, 2017). We divided the study period based on data availability, analyzing the years 1980, 1990, and 2000 initially, followed by 2015 and 2023. The intervals between the first set of years were ten years, but due to the lack of a 2010 Landsat image, we used 2015 instead, resulting in a fifteen-year gap between 2000 and 2015. Finally, we included 2023 as the most recent year in our analysis. For calculating LST, we utilized the same Landsat scenes used to develop the LULC maps.

Table 1 A summary of the data types and sources used in this study
Full size table
Table 2 Landsat Images that have been analyzed in this study
Full size table

The NDVI and NDBI indices were calculated using the GEE platform for the same years. These indices are calculated by Eqs. (1 and 2) (Rouse et al., 1974; Waqar et al., 2012):

$${\text{NDVI}} = \frac{NIR - Red}{{NIR + Red}}$$
(1)
$${\text{NDBI}} = \frac{SWIR 1 - NIR}{{SWIR 1 + NIR}}$$
(2)

where the (Red) is red, (NIR) is the near-infrared, and (SWIR1) is the shortwave infrared bands centered at ~ 1.64 µm.

The terrain properties were derived from the digital elevation model (DEM) from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) satellite with a spatial resolution of 30 m. Third, the roads and the maps of major settlements were obtained from the Open Street Map (OSM) at https://extract.bbbike.org/. All the spatial layers were unified to the same coordinate system (UTM Zone 36 N) and rasterized at spatial resolution (30 m).

LULC modeling

The Modules for Land-Use Change Simulation (MOLUSCE) plugin was employed by QGIS software version 2.8 to perform simulation and prediction for the LULC in the GAM (Fig. 2). This plugin was developed by Asia Air Survey and NextGIS (https://github.com/nextgis/molusce). First, the MOLUSCE model validation was conducted by generating a simulated map for LULC for 2023 and comparing it with the observed image of 2023 to check the model’s performance. Then, the Pearson correlation coefficient (R) was performed to ascertain the relationship between the driving factors (i.e., Slope, distance to roads) and LULC maps (i.e., 2000 and 2015). The ANN algorithm was then employed to generate Transition Potential Modeling (TPM), which is a process that determines the types that will become in the future (Parsamehr et al., 2020). The attributes for TPM were set as follows: the neighborhood was set to 1 pixel, the learning rate was set to 0.001, the maximum number of iterations was set to 100, the number of hidden layers was set to 10, and the momentum was set to 0.025. After that, the adoption of CA aimed to replicate LULC changes during 2023. After ensuring modeling quality, the prediction LULC map for 2030 was generated using a below steps. First, the LULC maps of 2015 and 2023 were seated as the initial and final images, respectively, with the spatial variables for 2023. After that, the CA algorithm was performed to generate the predicted LULC of 2030, which was determined by the TPM results for producing the image.

Fig. 2
figure 2

Schematic diagram for the process in MOLUSCE plugin to check the model’s validity and predict the LULC map for 2030

Full size image

LST calculation and modeling

The LST images for the study period were calculated in Google Earth Engine. These LST maps were then used as inputs in the LSTM model to simulate and predict the LST in GAM using MATLAB software (Fig. 3). To calibrate and validate the model’s performance, we first simulated the 2023 LST image and compared it to the observed 2023 LST image. Simulating 2023 LST image was based on setting the 2000 LST image as the initial image and the 2015 image as the final image, as well as setting the spatial variables data (i.e., NDVI, NDBI, and LULC of 2015) that have an impact on LST, as reported by Faisal et al. (2021); Al Kafy et al. (2021); and Shatnawi and Abu Qdais (2019). These variables correlate highly with LST, impacting the LST simulation and prediction process.

Fig. 3
figure 3

Schematic diagram for the process in MATLAB to check the model’s validity and predict the LST map for 2030

Full size image

The simulation and prediction in the LSTM work in several steps (Zhang et al., 2018). The first step is to confirm that all entered data (Cell values) has been taken and that is the cell values in these data will be discarded or retained in the model (Old data), as well as detect the past LST prediction and their relationship among the NDVI, NDBI, and LULC in a time. This can be determined using forget gate layer (\(f\)), which it divides the results into two classes, (i.e., 0 means fully forgot and 1 means completely keep) using the following Eq. (3):

$$ft = \sigma \left( {Wf.\left[ {ht - 1, xt} \right] + bf} \right)$$
(3)

where the (\(f\)) is the forget gate layer in time, \(\sigma\) is the results of the (\(f\)) (i.e., 0 and 1), \(Wf\) is the weight matrix for (\(f\)), \(xt\) is the NDVI, NDWI, and LULC in the period time, \(ht - 1\) is the past LST predictions and their relationship with NDVI, NDBI, and LULC (i.e., 2015), and \(df\) is the bias vector for (\(f\)).

The second step is to determine what the data that will be stored in each cell value by creating a new layer called the input gate layer (\(i)\) that contains the data that will be used in the simulation from data (i.e., include the present NDVI, NDBI, and LULC and in our case for validation, the 2015 is a present year) (Eq. 4). Then it creates a new layer called tanh (°Ct) that has the potential cell value (\(C\)) based on the input date layer and the output of these steps as new data (Eq. 5).

$$it = \sigma \left( {Wi.\left[ {ht - 1, xt} \right] + bi} \right)$$
(4)
$$^\circ {\text{C}}t = \tanh \left( {Wc.\left[ {ht - 1, xt} \right] + bc} \right)$$
(5)

where the (\(it)\) is the input gate layer in time, \(Wf\) is the weight matrix for (\(i\)), \(di\) is the bias vector for (\(i\)), \(^\circ {\text{C}}t\) is the tanh layer that has potential cell values, \(Wc\) is the weight matrix for (\(C\)), and \(bc\) is the bias vector for (\(C\)).

The third step involves updating the cell values (\(Ct\)) based on companies the old data with new data using the following Eq. (6):

$$Ct = ft \times Ct - 1 + it \times^\circ {\text{C}}t$$
(6)

where the \(Ct\) updates the cell value, and \(Ct - 1\) is an old cell value (Old data). The fourth step is to create the output gate layer \(\left( o \right)\), which is the cell values that should be output to simulate the LST in 2015 using the following Eq. (7):

$$ot = \sigma \left( {Wo.\left[ {ht - 1, xt} \right] + bo} \right)$$
(7)

where the \(ot\) is the output gate layer in a time, \(Wo\) is the weight matrix for \(o\), \(ht - 1\) is the past hidden values, and \(bo\) is the bias vector \(\left( o \right)\) However, the output of these layers is new hidden values (\(ht)\) based on \(o\) and \(Ct\) that are used to predict LST image for the year 2030, using Eq. (8):

$$ht = ot \times \tanh \left( {Ct} \right)$$
(8)

The simulated and observed images were compared to check the performance of the LSTM results using regression analysis. The root mean squared error (RMSE), coefficient of determination (R2), and mean absolute error (MAE) were calculated using 319 random points that were generated using the polynomial sampling (Eq. 9) (Al Shogoor et al., 2022):

$${\text{N}} = \frac{{Z^{2} \left( p \right)\left( q \right)}}{{E^{2} }}$$
(9)

where N is the number of samples, p is the expected percent accuracy of the entire map (i.e., 85%), q = 100–p (i.e., 15%), E is the allowable error (i.e., 4%), and Z = 2. Then, the RMSE, R2, and MAE were calculated using the following Eqs. (10, 11, and 12) (Georganos et al., 2021):

$${\text{RMSE}} = \sqrt {\frac{{\mathop \sum \nolimits_{i = 1}^{N} \left( {Yi - Xi} \right)^{2} }}{N}}$$
(10)
$${\text{R}}^{2} = 1 - \frac{{\sum {\left( {Xi - Xj \to } \right)^{2} } }}{{\sum {\left( {Yi - Yj \to } \right)^{2} } }}$$
(11)
$${\text{MAE}} = \frac{{\mathop \sum \nolimits_{i = 1}^{N} Yi - Xi}}{N}$$
(12)

where the \(Xi\) is the observed values, \(Yi\) is the simulated values, \(Xi \to\) is the observed mean value, \(Yi \to\) is the simulated mean value, and N is the number of points.

Results

Quality of the LULC maps and MOLUSCE modeling

Table 3 shows the accuracy assessment for the LULC maps that have been used in this study. The overall accuracy ranged between 85 and 94%, while the kappa coefficient ranged from 0.78 to 0.91. The user’s accuracy ranged from 40 to 100%, while the producer’s accuracy ranged from 54 to 100%. Based on these evaluation parameters, the LULC was highly accurate in studying the trends of changes in the study area.

Table 3 A confusion matrix for the classified images in the studied years based on Google Earth’s image (n = 319)
Full size table

The kappa value for TPM was 0.92, and the minimum validation overall error was 0.0045 (Table 4), which means the transition potential was highly accurate in predicting the potential changes in LULC types. The results for validation of the model indicated that the simulated results exhibit high levels of agreement and correction by obtaining the percentage of correctness, kappa (overall), Kappa (histo), and Kappa (loc) (i.e., 93%, 0.89, 0.92, and 0.98, respectively) (Table 4).

Table 4 The accuracy of the results for the LULC modeling in the MOLUSCE model
Full size table

Spatio-temporal of LULC dynamics in GAM (1980–2023)

The GAM LULC maps for 1980, 1990, 2000, 2015, and 2023 (Fig. 4) show that urban areas increased in all years. Their coverage increased from 6.6% of the total land in 1980 to 12.2% in 1990. It continued to rise, reaching 17.3% in 2000, 40.76% in 2015, and 45.3% in 2023, with net changes of about 35.8% between 1980 and 2023 (Table 6). Agricultural areas covered about 18.5% in 1980 and increased in 1990 and 2000 to 30.6% and 31.8%, respectively. After that, it decreased in all later years to 20.54% and 17% in 2015 and 2023, respectively, with a net decline of about 1.5% between 1980 and 2023. Barren areas decreased in all years from 6.6% in 1980 to 6% in 1990, to 5.1% in 2000, to 4.11% in 2015, and 4.22% in 2023, with a net decline of 2.4% between 1980 and 2023. Forested areas slightly decreased in all years, from 0.56% of the total area in 1980 to 0.53% in 2023. Rangelands/sparsely vegetated areas covered about 67.7% in 1980 and decreased to 50.6% in 1990, 45.2% in 2000, 34% in 2015, and 32.9% in 2023, with a net decline of 34.8%.

Fig. 4
figure 4

The land use and land cover (LULC) maps for the GAM from 1980 to 2023

Full size image

Figure 5 shows the spatial patterns for LULC types in GAM using the gradient direction method that divided the direction into 16 directions. The urban areas are concentrated in the middle of the study area and extend into the northwest, northeast, and southwest directions of the GAM area. 2015 and 2023 experienced the most urban expansion in all directions in the GAM. This expansion took place at the expense of the agricultural lands, especially those located south, southwest, and west of the GAM area. Barren areas were also converted to urban, especially in the eastern and NW parts of GAM.

Fig. 5
figure 5

The spatial pattern of each LULC type in the GAM (1980 to 2030). a Urban areas, b Agriculture, c Rangelands/sparsely vegetated, d Forests, e Barren

Full size image

Forecasting LULC to 2030

The simulation results indicated that our model had high precision in predicting the LULC maps in 2030 (Table 4). The observed and simulated 2023 LULC map comparison showed high agreement (Table 5, Fig. 6). According to the predicted results, the urban areas are expected to continue expanding to cover approximately 48.9% by 2030, which could represent a net change of 42.9% from 1980. The other LULC types are expected to decline with the growth in the urban area. For instance, the coverage of agricultural land, barren lands, forests, rangelands, and sparsely vegetated areas is expected to cover ~ 13.6%, 4.14%, 0.42, and 32.9% by 2030, respectively. Table 6 summarizes the areas, percentages, and total area change for each LULC type in the GAM from 1980 to 2030.

Table 5 Comparison between the simulated versus observed LULC 2023 for each LULC type
Full size table
Fig. 6
figure 6

The LULC maps, simulated and observed images for 2023, as well as predicted image for 2030

Full size image
Table 6 The changes between LULC types between 1980 and 2030
Full size table

Spatio-temporal patterns of LST in the GAM (1980–2023)

The LST timeline (Fig. 7) revealed that the northwest and central areas of the GAM experienced low LST values, whereas the northeast and southeast regions encountered high LST values. The eastern part of the study area showed median LST values. In 2023, there was a noticeable increase in the LST values, with the minimum, mean, and maximum values of 24.44 °C, 38.72 °C, and 50.27 °C, respectively (Fig. 8). In contrast, the LST values in 2015 were lower, with the minimum, mean, and maximum values at 20.39 °C, 33.97 °C, and 43.59 °C, respectively. In 2000, the LST values were lower than those observed in 2015; the minimum, mean, and maximum values of LST in 2000 were 15.59 °C, 30.8 °C, and 35.16 °C, respectively. The values were lower in 1980, at a minimum of 20.49 °C, a mean of 29.7 °C, and a maximum of 38.96 °C. The spatial pattern of LST variation showed a concentration of the lowest LST values in the middle, north, and northwest parts of the GAM. In contrast, the highest values were evident in the study area’s southwest, northeast, east, southwest, and west parts.

Fig. 7
figure 7

The LST distributions in the GAM from 1980 to 2030

Full size image
Fig. 8
figure 8

The minimum, mean, and maximum LST values in the GAM from 1980 to 2030

Full size image

Forecasting LST to 2030

The validation of the LSTM model shows reliable prediction capability with an R2value of 0.9 (Fig. 8). The predicted LST for 2030 (Fig. 7) showed that the maximum LST recorded was 53.13 °C, which increased from the LST recorded in 2023. The minimum LST recorded was 28.29 °C, while the mean LST recorded was 41.57 °C. The prediction for 2030 indicated a projected increase in LST compared to previous years (Fig. 9).

Fig. 9
figure 9

Validation for the LSTM model. (n = 319)

Full size image

The relationship between the LULC and LST in GAM (1980–2030):

Figure 10 illustrates the mean, minimum, maximum, and standard deviation of LST values across different LULC types. In urban areas, the mean LST increased in 1980 and 1990, reaching 28.35 °C and 28.57 °C, respectively, with a slight deviation of approximately 0.22 °C. This trend reversed in 2000, dropping to 24.31 °C, but then escalated significantly in subsequent years, reaching 32.53 °C in 2015, 36.81 °C in 2023, and is projected to be 39.76 °C by 2030, with a deviation of around 7.23 °C. For barren areas, the mean LST also increased in 1980 and 1990, reaching 29.67 °C and 30.25 °C, respectively, with a deviation of about 0.58 °C. This value decreased to 26.89 °C in 2000, followed by a rise to 34.03 °C in 2015, 39.69 °C in 2023, and a projected 42.56 °C in 2030, with a deviation of approximately 8.53 °C.

Fig. 10
figure 10

The variation of the mean LST over different LULC types in GAM from 1980 to 2030

Full size image

Forested areas exhibited a similar pattern, with mean LST values increasing in 1980 and 1990–28.42 °C and 28.52 °C, respectively, with a minimal deviation of 0.1 °C. This trend saw a decline to 23.15 °C in 2000 but later increased to 31.19 °C in 2015, 34.16 °C in 2023, and is anticipated to be 36.87 °C by 2030, with a deviation of about 5.68 °C. Agricultural lands showed an increase in mean LST in 1980 and 1990, reaching 29.39 °C and 30.72 °C, respectively, with an increase of approximately 1.33 °C. This decreased to 25.69 °C in 2000 but increased again to 34.98 °C in 2015, 38.06 °C in 2023, and is expected to be 40.87 °C by 2030, with a deviation of around 5.89 °C. Rangelands/sparsely vegetated areas experienced an increase in mean LST in 1980 and 1990, reaching 30 °C and 31.57 °C, respectively, with a rise of 1.58 °C. This value decreased to 27.33 °C in 2000 but increased again to 35.26 °C in 2015, 41.79 °C in 2023, and is projected to be 44.65 °C by 2030, with a deviation of approximately 9.39 °C.

The results indicate that in 1980, the LST values for urban, forest, agricultural, barren, and rangelands-sparsely vegetated areas were in ascending order. Similarly, in 1990, the same ascending order was observed. For the years 2015, 2023, and 2030, the LST values followed a different ascending order: forest, urban, barren, agriculture, and rangelands/sparsely vegetated. Notably, forests consistently exhibited the lowest LST values across all years except 1980, where urban areas followed closely. Agricultural areas consistently ranked third in terms of LST values. Barren and rangelands/sparsely vegetated areas consistently recorded the highest LST values throughout the observed years.

Discussion

The study investigated the spatial patterns and temporal trends of LULC and LST in the Greater Amman Municipality. The analysis revealed significant growth in urban regions, the most dynamically changing unit within the study area, at the expense of other LULC types. Projections indicate a potential 42.4% increase in urban areas from 1980 to 2030. In contrast, agricultural lands, barren lands, forests, and sparsely vegetated rangelands are expected to decline by approximately 4.9%, 2.5%, 0.14%, and 35%, respectively. The urban expansion in GAM was attributed to a significant increase in population density, which surged from 2,469 capita/km2 in 2003 to 4,453 capita/km2 in 2015—an 80% increase. This population inflation has played a crucial role in the observed urban growth in Amman.

Furthermore, the economic circumstances of the urban locality, characterized by limited resources, have exacerbated the encroachment of urban development onto agricultural lands. Khawaldeh (2016) identified several factors influencing urban expansion in the study area, including the 1991 Gulf War, which prompted a significant return of Gulf nationals to Amman, and the 2003 Iraq invasion, which led to a substantial migration of the Jordanian population to the capital. As of 2007, the study area hosted approximately 40% of the country’s population. According to Al Shogoor et al. (2022), the protracted Syrian conflict since 2013 has also significantly contributed to urban expansion. Hazaymeh et al. (2022) noted that these factors have negatively impacted vegetation cover, reducing its coverage area. Additionally, the high cost of land and apartments in central GAM has driven many individuals to purchase property in rural areas, which are comparatively more affordable and accessible. This shift has contributed to an increase in the rural populace and extensive urban expansion at the expense of other landforms.

The analysis reveals a significant increase in Land Surface Temperature within the Greater Amman Municipality (GAM) during the study period, indicating considerable spatiotemporal changes. An upward trend in temperature is evident across all land use categories, except for the year 2000, where a distinct decrease was observed. Alkaraki and Hazaymeh (2023) attribute this anomaly to the unusually high precipitation levels in 2000, which had a cooling effect on the region by reducing heat emissions, as corroborated by Jensen (2016). The overall temperature rise can be attributed to several factors, including urbanization, rapid and uncontrolled urban development, rural-to-urban migration, climate change, and global warming (Al Kafy et al., 2021; Al Shogoor et al., 2022). The study revealed a confirmed LST mean exists for every LULC type and years concerning the association between LULC types and LST. The LULC types of urban areas and forests have displayed the least amount of LST compared to the other types of LULC.

Conversely, agricultural lands and barren areas have demonstrated a moderate LST. Furthermore, the rangelands/sparsely vegetated areas exhibit elevated LST values. Nonetheless, the outcomes obtained from our research were incongruent with prior investigations conducted by Zhang et al. (2023), Al Kafy et al. (2021), Faisal et al. (2021), and Chaudhuri and Mishra (2016). The findings of these studies indicate that urban regions have the highest temperatures, whereas barren lands demonstrate moderate temperatures, followed by agricultural lands with the lowest temperatures. Zhang et al. (2023) stated a significant difference in LST between urban and agricultural areas in Wuhan, China.

Furthermore, Al Kafy et al. (2021) findings indicate that urban areas in Rajshahi, Bangladesh, exhibit high temperatures, while barren lands show moderate temperatures and agricultural lands display relatively lower temperatures. Faisal et al. (2021) reported similar observations for the Dhaka Metropolitan Area, where urban regions have elevated temperatures while barren and agricultural lands have moderate and low temperatures, respectively. Chaudhuri and Mishra (2016) also noted variations in land surface temperatures across different land use types in India and Bangladesh, with urban areas showing high temperatures and agricultural and forest lands exhibiting moderate to low temperatures. These distinctions can be attributed to various factors, including the timing of image capture, which directly affects thermal readings (Jensen, 2016). In our study area, Landsat 5 captured images at 7:59 am and Landsat 8 at 8:10 am, whereas in other countries, such as China, images were captured at 2:43 pm for Landsat 5 and 2:55 pm for Landsat 8. In Bangladesh and India, images were captured at 4:18 pm for Landsat 5 and 4:29 pm for Landsat 8.

The early morning image captures in Jordan resulted in relatively low temperatures due to differential sunlight exposure across various terrains. Rangelands, sparsely vegetated areas, barren lands, and agricultural lands received direct sunlight, while urban areas and forests experienced partial obstruction of the sun’s rays due to buildings and trees. This phenomenon is due to the horizontal position of the sun at this time of day. Conversely, in China and Bangladesh, the images captured at midday resulted in higher temperatures, as the sun’s rays exposed all surfaces to intense solar energy. Additionally, buildings in urban areas absorb and store significant amounts of energy at noon, which is why urban regions consistently appear to have the highest temperatures. Satellite images captured by MODIS and Sentinel may provide more accurate LST readings in Jordan, as the Landsat satellites have shown limitations in effectively capturing these variations.

Limitations

Despite the valuable insights gained, this study is subject to several limitations. Firstly, our findings may lack generalizability compared to other studies analyzing mean temperatures across different LULC types. This is primarily because the Landsat satellite captured images of the study area, Jordan, during morning hours. During this time, a temperature disparity exists between urban and forested regions and rangelands, sparsely vegetated areas, barren lands, and agricultural lands. Urban and forested regions tend to have lower temperatures due to the presence of buildings and trees. In contrast, the more open rangelands, sparsely vegetated areas, barren lands, and agricultural lands experience higher temperatures. Various satellites, such as Sentinel that collect data at noon are crucial for accurate LST analysis. However, for this study, we utilized Landsat primarily to address LULC data. Additionally, incorporating historical temperature data from weather stations would enhance the study. Unfortunately, many of these stations lack complete records. Moreover, employing a range of deep learning algorithms to generate future predictions of LULC and LST is essential. Identifying the algorithm that provides the highest accuracy in a simulation would improve the robustness of the predictions. These methodological enhancements and data sources could significantly refine the analysis and provide more comprehensive insights.

Conclusions and future outlook

This study investigated the spatial patterns and temporal trends of land use and land cover types and Land Surface Temperature (LST) in the Greater Amman Municipality by projecting changes up to 2030. The results reveal a significant rise in urbanization accompanied by a reduction in other LULC types. The study also found an increase in maximum LST from 38.96 °C in 1980 to 53.13 °C in 2030, with a deviation of approximately 14.17 °C. The findings of this study provide crucial insights into the dynamic spatial patterns of LULC and LST in the GAM area, carrying significant policy implications for managing urbanization, preserving natural resources, and mitigating climate change impacts. With the notable increase in urban areas at the expense of agricultural, barren, and forest lands, there is an urgent need for comprehensive urban expansion management strategies. Local authorities are urged to consider implementing zoning regulations, land use planning frameworks, and sustainable development practices to harmonize urban growth with preserving vital ecosystems.

The observed correlation between population density increases and urban expansion underscores the importance of regulating population growth. Policy interventions promoting equitable population distribution and affordable housing initiatives can alleviate pressure on urban areas and prevent encroachment on agricultural lands. These strategies are essential for achieving a balanced approach to development that safeguards the Environment and the population’s well-being. Recognizing the impact of rapid urbanization on vegetation cover and land surface temperature, policymakers should mandate comprehensive Environmental Impact Assessments (EIAs) for all urban development projects. The rise in LST accentuates the need for climate-responsive urban planning, urging authorities to integrate sustainable design principles, including green infrastructure, heat-resistant materials, and increased green spaces, to counteract the urban heat island effect and enhance overall urban livability. These implications stress the urgency of adopting sustainable urban development practices, regulating population density, and leveraging advanced technologies for effective monitoring. Implementation of these recommendations can guide policymakers in achieving a harmonious balance between urban growth and environmental conservation in the GAM area.

Future research activities in the study area will investigate the underlying factors driving the significant rise in urbanization and its impact on other land use and land cover types. Detailed studies on the socioeconomic and policy factors influencing urban expansion can provide insights for more effective urban planning and management strategies. More advanced predictive modeling techniques will be explored to enhance the accuracy of LULC and LST projections, thereby developing more precise future scenarios and mitigation strategies. Another significant area will emphasize the effectiveness of various urban green infrastructure solutions in mitigating the urban heat island effect and improving urban livability. This will allow for long-term monitoring of these interventions, providing valuable data on their performance and sustainability in a highly dynamic urban environment.