Next Article in Journal
The Association between Outside Directors’ Compensation and ESG Performance: Evidence from Korean Firms
Previous Article in Journal
Identification of Soil Erosion-Based Degraded Land Areas by Employing a Geographic Information System—A Case Study of Pakistan for 1990–2020
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Sustainability
Volume 14
Issue 19
10.3390/su141911873
sustainability-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Developing a Spatiotemporal Model to Forecast Land Surface Temperature: A Way Forward for Better Town Planning

by
Umer Khalil
1,
Umar Azam
2,
Bilal Aslam
3,
Israr Ullah
4,
Aqil Tariq
5,6,*,
Qingting Li
7,* and
Linlin Lu
8
1
ITC Faculty of Geo-Information Science and Earth Observation, University of Twente, Hengelosestraat, 997514 AE Enschede, The Netherlands
2
Department of Computer Science, COMSATS University Islamabad, Wah Campus, Wah Cantt, Islamabad 47040, Pakistan
3
School of Informatics, Computing, and Cyber Systems, Northern Arizona University, Flagstaff, AZ 86011, USA
4
Division of Earth Sciences and Geography, RWTH Aachen University, 52062 Aachen, Germany
5
Department of Wildlife, Fisheries and Aquaculture, Mississippi State University, 775 Stone Boulevard, Starkville, MS 39762, USA
6
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430072, China
7
Airborne Remote Sensing Center, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
8
Key Laboratory of Digital Earth Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
*
Authors to whom correspondence should be addressed.
Sustainability 2022, 14(19), 11873; https://doi.org/10.3390/su141911873
Submission received: 10 August 2022 / Revised: 10 September 2022 / Accepted: 19 September 2022 / Published: 21 September 2022
Browse Figures
Versions Notes

Abstract

:
The change in the local climate is attributed primarily to rapid urbanization, and this change has a strong influence on the adjacent areas. Lahore is one of the fast-growing metropolises in Pakistan, representing a swiftly urbanizing cluster. Anthropogenic materials sweep the usual land surfaces owing to the rapid urbanization, which adversely influences the environment causing the Surface Urban Heat Island (SUHI) effect. For the analysis of the SUHI effect, the parameter of utmost importance is the Land Surface Temperature (LST). The current research aimed to develop a model to forecast the LST to evaluate the SUHI effect on the surface of the Lahore district. For LST prediction, remote sensing data from Advanced Spaceborne Thermal Emission and the Reflection Radiometer Global Digital Elevation Model and Moderate-Resolution Imaging Spectroradiometer sensor are exploited. Different parameters are used to develop the Long Short-Term Memory (LSTM) model. In the present investigation, for the prediction of LST, the input parameters to the model included 10 years of LST data (2009 to 2019) and the Enhanced Vegetation Index (EVI), road density, and elevation. Data for the year 2020 are used to validate the outcomes of the LSTM model. An assessment of the measured and model-forecasted LST specified that the extent of mean absolute error is 0.27 K for both periods. In contrast, the mean absolute percentage error fluctuated from 0.12 to 0.14%. The functioning of the model is also assessed through the number of pixels of the research area, classified based on the error in the forecasting of LST. The LSTM model is contrasted with the Artificial Neural Network (ANN) model to evaluate the skill score factor of the LSTM model in relation to the ANN model. The skill scores computed for both periods expressed absolute values, which distinctly illustrated the efficiency of the LSTM model for better LST prediction compared to the ANN model. Thus, the LST prediction for evaluating the SUHI effect by the LSTM model is practically acceptable.

1. Introduction

The Earth’s surface undergoes changes in structures and materials due to human settlements. Variations in the composition of the atmosphere and the steadiness of the surface’s energy related to the nearby natural topographies are primarily caused by these changes [1,2]. These changes ensure an exceptional regional environment of anthropoid communities, named the urban environment, and the change progression is termed urbanization [3]. Urban areas are turning into bigger clusters, and diverse land uses are being altered due to the usage of anthropogenic materials due to the increase in the urban population [4]. The movement of people to urban areas from rural areas has been noted to cause an upsurge in the extent of metropolises worldwide. More than 50% of the Earth’s inhabitants reside in metropolises or towns, which is expected to increase in the upcoming years [5,6].
On Earth, urbanization is considered one of the utmost apparent human impressions that cause land-use variations (e.g., building and road construction) and greenhouse gas discharges. Together, these variations can, in turn, influence the daily mean surface temperature [7,8]. Several other factors also influence the surface temperature, but the changes in land use and greenhouse gases are regarded as the most important. Currently, due to urbanization, urban metropolises are enduring the challenge of the Urban Heat Island (UHI) effect where built-up regions express comparatively greater air and surface temperatures in contrast to rural areas because of the surface alterations (impervious surfaces), causing changes in the thermal climate in built-up regions [9]. The additional temperateness of the built-up area or atmosphere compared to non-urbanized settings is denoted as UHI. The UHI is termed Surface UHI (SUHI); when analyzing its effect, the Land Surface Temperature (LST) is utilized [3]. One of the parameters of utmost significance for evaluating UHIs is the skin temperature of the Earth’s surface, the LST [10,11].
In addition to the extent of the city, UHI primarily hinges on the water bodies, built-up configuration, vegetation coverage, building concentration, anthropogenic warmth produced by human activities, building and road geometry, surface materials, canyon geometry, deficiency of evaporation in the metropolis, and thermal and ocular characteristics of utilized resources in the city hedge [4,12]. Human health in built-up areas is at risk because of heat stress. Due to the upsurge in swift urbanization by the growth of metropolises and built-up enlargement, besides global warming, it is likely to deteriorate in the future [13,14,15]. In localities with inadequate vegetation, an elevated proportion of impermeable surfaces, and low albedo values, surface UHI has been noted to be increased. The surface temperature drops by 1.3 °C, with an upsurge in the proportion of vegetation cover by 10% [16]. Varied air and surface temperature trends and escalation in mean temperatures because of weather variation also impact the built-up environment causing the UHI effect [17]. Regional meteorology is affected by UHI in such a manner that it increases humidity, changes local wind patterns, changes the precipitation rate, and forms clouds and fog [8].
A diverse kind of LST-associated SUHI investigations has been executed for diverse metropolises, which became possible due to the availability of LST data from several sensors supported by diverse satellites that subdue a few complications of in situ quantification [18,19,20]. UHI intensity is used to express the magnitude of UHI, which is the variation among rural and urban temperatures [21,22]. SUHI signifies a positive difference, which means that the urban area is hotter than the rural area, whereas SUHS (Surface Urban Heat Sink) has been noted when the difference is negative. Temporal variability, both diurnal and seasonal, is associated with SUHS and SUHI [21]. Qiao et al. [23] observed the strongest UHI in the Beijing metropolitan area owing to the influence of built-up land during the daytime in summer. The overall built-up population influences the spatial extent of the UHI, and the built-up population concentration compared to the total population more intently influences the magnitude of the UHI [5].
In developing countries, significant population growth is the foremost reason for urbanization. Pakistan has witnessed exceptional expansion recently, and many cities in the country have expanded swiftly [24]. The Lahore metropolitan area is among the fastest developing metropolises in the world. There has been an upsurge in the urbanization rate in the city from 32.52% in 1996 to 36.38% in 2016 [25]. Recently, the rate has soared further, and the expanding inclination is causing massive land-use alterations. The city’s existing infrastructure is under an immense burden because of the extreme growth rate; thus, extensive development is being carried out for the increasing population. Consequently, in the built-up zones, air pollution and smog emergence are boosted due to a high rate of heat discharge [26]. The increasing city size tends to increase the UHI intensity. Dissimilarities in land cover, i.e., the variance in the configuration of the Earth’s surface in urban and rural settings, account for the UHI effect [27]. The altered urban surface generally exhibits non-transpiring and reduced evaporating properties. Therefore, an improved perception of the heat island impact and its association with the built-up topographical factors is essential for more effective city planning.
Several researchers have employed multiple techniques to predict LST for the evaluation of SUHI effect. Mathew et al. [28] predicted LST for the exploration of the UHI effect in the Indian city of Ahmedabad, exercising the Linear Time Series model. Another investigation conducted by Mathew et al. [29] predicted LST for SUHI valuation across Chandigarh city, India, utilizing the Support Vector Regression model. Moreover, several UHI prediction investigations are centered on Neural Networks (NNs) [30,31]. Utilizing the air temperature of 28 locations in Seoul, South Korea, for one year, Lee, Kim, and Yun [31] formulated an NN architecture to forecast the UHI. Centered on the NN design, Kolokotroni et al. [32] determined air temperature data inside the city, as well as data from meteorological stations, as the input to forecast the UHI concentration in London. Khalil et al. [33] used multiple deep learning algorithms to predict LST for the Lahore district and found that the Convolutional Neural Network (CNN) outperformed the other methods used. Yu et al. [34] also used a model based on CNN to predict the sea surface temperature and found that the model maintained very good prediction results.
Moreover, as with CNN, Recurrent Neural Networks (RNNs) have also gained much attention recently due to their better performance. The RNNs are the deepest of all NNs and can produce and address memories of sequences of input patterns [35]. Long Short-Term Memory (LSTM) is a special kind of RNN that has been found to demonstrate promising results in several fields. Chen et al. [36] used an LSTM-based hybrid model to predict the temperature of seasonally frozen subgrades and achieved improved predictions. Tsokov et al. [37] developed a hybrid spatiotemporal model based on CNN and LSTM for the prediction of air pollution and obtained optimal results. Zhang et al. [38] predicted the water table depth in agricultural areas based on an LSTM model and concluded that LSTM is better than the conventional Feed-Forward Neural Network. Garg and Jindal [39] used multiple models, including two deep learning models, namely LSTM and CNN-1D, and evaluated time series forecasting of PM2.5 and discovered that LSTM outperformed all other models.
Therefore, in the current investigation, an LSTM-based model is used to forecast LST for the analysis of SUHI. The investigation aimed at developing an LSTM model and forecast LST utilizing diverse urbanization and vegetation parameters to assess the SUHI effect (temporal and spatial patterns of SUHI) for a designated research area of the Lahore district in Punjab, Pakistan. The research examines the SUHI effect over the Lahore district from the predicted LST, mainly focusing on the urban area and the SUHI concentration over the pondered area for diverse periods as it incorporates diverse factors that directly affect the LST.

2. Study Area

The district of Lahore is designated for this investigation. The district of Lahore is situated in the Punjab province of Pakistan, primarily comprising the metropolitan area of Lahore. Lahore is the capital of Punjab province and the second utmost significant metropolitan of Pakistan, which lies between latitude 31°32′59″ N and longitude 74°20′37″ E [33]. According to the data published by the Pakistan Bureau of Statistics, the district of Lahore encompasses an area of 1772 km2 (684 sq mi) with a total population of 11,126,285 as of the 2017 census [33,40]. Compared to the minimum fixed world standard of 25–30% of built-up area compared to green exposed space, Lahore has only 3% of its territory for green spaces. According to the Pakistan Bureau of Statistics Lahore, in terms of the built-up population globally, it ranked 56th in 1975, 38th in 2007, and 24th in 2025 [33,40]. Regarding the urbanized landscape of the explorative area, which is distinguished from the rest of the area by an urban boundary, the cooling effects are imparted by the two significant canals, the Lahore canal and the river Ravi, which flow from within the city. In recent times, the metropolitan area of Lahore has been subjected to an enormous land-use change. The inhabitants account for varying government plans, insignificant land use planning, and prompt development for these changes [25]. The map of the study area is shown in Figure 1. The urban boundary in the study area was marked based on the study by Rana and Bhatti [40], in which the authors highlighted the demographics and administrative division of the Lahore district.

3. Data and Methodology

3.1. Indicators Selection

To alleviate the SUHI effect, vegetation cover is the central focus for intrusion approaches and one of the most promptly governable variables in built-up landscapes [21]. The vegetated areas exhibit lower surface temperatures than the waterproof ones [6]. The Normalized Difference Vegetation Index (NDVI) has substantial negative associations with the LST [23,41,42,43,44,45,46], whereas impervious surface area is related positively to LST [43,47,48,49]. Similar to NDVI, EVI is also a vegetation index, but it is different from the former due to its utilization of factors to diminish atmospheric impacts [50]. As for the advancements of EVI over NDVI, it is essential to discover its execution for UHI investigations. Moreover, the effect of vegetation is not only accountable for the spatial variation of NDVI but also the slope, elevation, obtainability of solar radiation, topography, and additional features [51,52].
One of the foremost observed contributors to temperature intensification, accounting for 70% of the overall alteration in LST, is the Impervious Surface Area [2,53]. The temporal difference and spatial dissemination of built-up thermal configurations and related land-use land-cover (LULC) conditions can quantitively be highlighted by LST and the Normalized Difference Bareness Index (NDBI), together with percent ISA [54].
Consequently, based on the importance of vegetation cover and impervious surfaces, this study chooses EVI, elevation, and Road Density (RD) as the indicators to forecast LST. The impacts of urbanization and vegetation on LST are deliberated utilizing RD as an urbanization parameter with EVI as a vegetation parameter. The reason behind using elevation as an indicator is that it influences the intensity of vegetation.

3.2. Data Acquisition

MODIS products MYD11A2 (8-day land surface temperature and emissivity) and MYD13A2 (16-day vegetation indices), with a resolution of 926.6 m and overlapping dates, were utilized for this investigation. The details of the used remote sensing data in this study are presented in Table 1. The 10-year LST (2009–2019) of two fixed time intervals (January, 009–016 days and May, 137–144 days) is adopted in this research. The reason for choosing these two different time periods is that they represent winter and summer seasons, which will help in understanding the seasonal variations of LST over the area. MODIS products are not only open access but also ready to use, which is a big positive. The study utilizes the MYD11A2 product with a quality flag to comprise only decent-quality pixels (LST error limited to 2.5 K) for the investigation.
The MODIS EVI outcome is computed from atmospherically altered bi-facial surface reflectance screened for clouds, hefty misters, water, and haze shades. The canopy background distinctions are reduced by EVI and preserve improved sensitivity over concentrated vegetation situations. Using the blue band, EVI also excludes the lingering atmosphere adulteration caused by sub-pixel thin clouds and smoke.
The downloaded MODIS data are in HDF-EOS format and the Sinusoidal Projection System. The Earth-gridded tile area covers approximately 1100 km × 1100 km in each MODIS image. The MODIS Re-projection Tool (MRT) was used to pre-process MODIS images. MRT was utilized for the sub-setting of the data to a reduced region. The data were also re-projected from the Sinusoidal projection to the UTM Zone 43 N projection system with WGS 1984 datum and reformatted from HDF-EOS to GeoTIFF format.
The Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM) product was used to extract the elevation of the investigative area. The ASTER GDEM map of 2011 was used to obtain the elevation of the investigative area because the changes in the elevation of an area are not frequent. ASTER GDEM standard data products were produced with 30 m resolution and have Z accuracies, usually among 10 m and 25 m Root Mean Square Error (RMSE). To produce scene-based DEMs, automated processing, including stereo-correlation of the ASTER GDEM scenes, was performed. Outliers and residual bad values were detached. The residual anomalies were then corrected by averaging the chosen data to create the final pixel value. The data were partitioned into 1 × 1 degree tiles to match the resolution of the MODIS products.

3.3. Long Short-Term Memory

LSTM model is an RNN model originally introduced by Hochreiter and Schmidhuber [55]. The model, however, was modified several times afterward [56,57]. LSTM is a fixed layout of RNN, which is a planned layout for constructing time-based sequences. This RNN model has a memory that remembers the values from the previous stages. Cells are the elements where the memory is stored. Each memory cell contains gates: An input and an output gate. The memory flows through these gates, and the input and output are activated. The internal architecture of these cells includes a forget gate, which becomes an input as a self-recurrent connection. LSTM has the unique quality of long-range dependency, making it more precise than any conventional RNNs. Backpropagation causes error backflow within the RNN architecture. The chain rule is the foundation behind the reason for introducing a backpropagation algorithm. When addressing the search for the minimum of the loss function, Stochastic Gradient Descent (SGD) is a competent algorithm [56,57,58].
Upon gaining the gradients of the present inner nodes of the computational graph, the gradients of the said node can be achieved with a point of commencement through the calculation of the point of gradient and the traversing into the course of the negative gradient. This is the quickest technique to hunt for a local minimum. The most elementary form of LSTM is called a memory cell, which is useful for calling back and cultivating unit outputs clear in various time phases. To call data for temporal contexts again, the memory cell of LSTM uses the cell steps. The LSTM has various gates, which are all utilized for the flow and control of information between the various steps [59,60,61].
For this experiment, neural networks concerning LSTM were engaged to predict radiation derived from the time series weather data. The vanishing gradient problem pops up because of the challenge of finding the long-term dependencies mathematically for the neural networks. Concerning the length of the input sequences, it results in being too hectic to note the impact of the starting phases. The gradients concerning the various initial inputs disappear and are reset to zero. The operating function of LSTM is highly regarded as the identity function with a derivative of 1.0 due to its repeated nature. So, the gradient through backpropagation cannot vanish or remain zero but remains unchanged.
The node’s activation function states the result of the particular node. Activation functions that are mainly used for the LSTM network are specifically sigmoid and tangential (tanh). The layout of LSTM is integrated with the sigmoid function through the forget gate and input gate and with the tanh function for the candidate vector that replaces the cell state vector. These activation functions of LSTM are calculated for the Input gate It, Output gate Ot, Forget gate Ft, Candidate vector C′t, Cell state Ct, and Hidden state ht, using the following formulae:
I t = s i g m o i d   W i X t , h t − 1 + b i
F t = s i g m o i d   W f h t − 1 , X t + b f
O t = s i g m o i d   W O h t − 1 , X t + b O
C t ′ = t a n h   W c h t − 1 , X t + b c
C t = F t ∗ C t − 1 + I t ∗ C t ′
h t = O t ∗ t a n h C t
X(t) is the input, ht−1 is the previous state, W is the weight, and b is the bias for each gate.
Figure 2 represents a single LSTM cell. The arrow in red shows the recurrent neural tanh activation function. As discussed above, there are three gates, namely input, output, and hidden gates. In Figure 2, the green box represents the hidden gate, the orange box represents the input gate, and the blue box represents the output gate. Figure 2 is a representation of the interaction between hidden and cell states.

3.4. Time-Series Forecasting Using LSTM

The procedure of performing LSTM requires pre-installation, which includes libraries such as TensorFlow, Keras, Sklearn, Matplotlib, and Pandas for the data frames. After installing these libraries in python, the next step is to load the Jupyter notebook and import the CSV data. After obtaining the values in the data frame, these values are aligned in the matrix to perform further procedures.
After the data are imported and set into the dataset matrix, MinMaxScaler is used to scale each feature into a given range; this feature translates each feature individually in the given range, then the dataset is transformed to fit the scaled features within the given range. Afterward, the dataset is split into training and testing data. The standard and optimized splitting of data is 80%–20%, whereas, in this study, it is performed as 90%–10%.
Following this, a recurrent model on the input sequence of length L, for example, input = (x(1), x(2), …, x(`)), is trained. For the simplicity of presentation, L = 3 is assumed, and an output sequence of length one is desired.
The graphical representation of the training of the LSTM model with the sequence L = 3 is represented in Figure 3. The M function takes the input, and the initial two outputs are ignored. Here, the blue line represents the hidden state (H is the hidden state at the previous timestep t−1, which is also the short-term memory), while the green shows the cell state (C is the cell state at the previous timestep t−1, which is also the long-term memory). The red dashed lines show the route taken during the backpropagation to update the weights inside S (the S represents different gates). Subsequently, the model values are predicted, and the test and predicted data are plotted.
According to the results, the epoch and number of hidden layers are adjusted accordingly. After the model is trained, validation of the model is performed to check the results with the test data. The representation of data is performed using Matplotlib and sometimes Seaborn, which shows the result of the data in graphical form as the model is tuned perfectly according to the required output. The data are reshaped and stored in CSV format.

3.5. Model Validation

To evaluate the function of the intended model, statistical parameters, specifically the Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), Mean Absolute Error (MAE), and R2, are computed based on Khalil, Aslam, Azam, and Khalid [33], Mathew, Sreekumar, Khandelwalm, and Kumar [29], and Mathew, Sreekumar, Khandelwal, Kaul, and Kumar [28]. Moreover, the skill metrics and skill scores are also computed to establish the performance of the LSTM model based on Wheatcroft [62].

4. Results and Discussions

To illustrate the distribution of LST over the study area, the LST images of the Lahore district for two periods of 2019 are exhibited in Figure 4. By probing the urban area (bounded by the built-up borderline) and the rural area (outside the built-up borderline), the influence of the built-up concentration of the Lahore district on LST can be distinctly comprehended. The administrative border of the city of Lahore is used to delineate the borderline between urban and rural areas. In Figure 4, the red color indicates the high-temperature pixels in the middle region of the district. In contrast, the light blue color indicates the low temperature at the outer fringe of the investigative area. For the area inside the built-up boundary, the LST is greater in contrast to the rural area. The region enclosed by the built-up periphery comprises the maximum number of pixels signifying an elevated temperature. Due to anthropogenic materials and impermeable surfaces, urban areas show higher-temperature pixels than rural areas.
In contrast to the LST of the built-up part, the LST of the rural area is lower, as can be seen from Figure 4. The temperature is low in rural areas because of the low urbanization and vegetation. However, for the month of May, the agricultural land becomes empty due to the harvesting of the crop [62], which is why some of the high-temperature pixels can also be seen in the rural area. Throughout the two periods in 2019, the configuration of LST does not vary considerably over the whole investigative area. Moreover, while analyzing the data used for building the model, it is observed that from 2009 to 2019, there was an increasing tendency of mean LST for both January and May. Thus, it can be said that the SUHI intensity, signified by the maximum and minimum LST distribution of an image conforming to a specific period, shows periodic variations from 2009 to 2019.
The EVI images of the Lahore district for the year 2019 are shown in Figure 5 for the purpose of showing the distribution of the vegetation over the study area. However, it is important to note that the EVI values vary noticeably from 2009 to 2019 over the study area, but the trend remains the same. The rural area is seen to have continually more vegetation than the urban area. Lower EVI values are exhibited by inadequately vegetated areas and are signified by purple, whereas the green in Figure 5 signifies amply vegetated areas. The rural area in the research area has very rare purple smudges and adequate green spots of diverse intensities, signifying vegetation growth of differing extent for these two months. The peripheral zones of the explorative area significantly on the northeastern side have much darker green spots and represent substantial EVI values compared to the areas that are adjacent to the enclosed urban boundary. The enclosed urban area seems considerably dreadful and signifies far lower EVI values. A large proportion of this bounded area is developed surface and impermeable. There is awfully inadequate (nearly trivial) vegetation in this area. In the rural area, the EVI values in the figure for the two months are related to a combination of vegetated land and reaped farming land encompassing both lands: Land that has a fraction of the plant left after crop reaping and bare soil. During the summer season, a low supply of water and high temperature instigates deprivation in the concentration of vegetation in the research area. Thus, the limit of EVI values is smallest through the summer season. All these factors are responsible for the variation of EVI values in the urban and rural areas for these two months.
Compared to the urban areas, immediate rural areas exhibit lower LSTs and comparatively lesser disparity in LSTs, as observed by Li et al. [63]. The urban land surfaces strongly influence LST and its spatial non-uniformity inside UHI. In the temperature images, the built-up zones are generally well defined from hot spots. Likewise, vegetated and open spaces can be interpreted as cool islands and areas characterized as cold spots [64]. Vegetation due to the cooling and shadow effect affects the surface temperature. However, both the vegetation cover and its concentration in urban areas are low, and typically, it is almost the lowest. In contrast, the built-up concentration reaches the maximum in a city’s central business district (CBD). At and around the CBD, the main reasons for the maximum temperature are the low levels of vegetation and large built-up zones [28,29].
The NDBI has a frailer relationship with LST, while NDVI and NDBI are satisfactorily associated with the deviations of LST [44]. An adverse association with LST is shown by the NDBI, which signifies soil dissemination in built-up areas. Among LST and NDBI, a positive association has been witnessed in several investigations [43,44,65,66]. Population changes, built-up stretching, and variations in LULC have led to a substantial disparity in the spatiotemporal configurations of the UHIs owing to the devastation of vegetated surfaces and diverse water features [67]. Severe and swift development demonstrates human-caused LULC alterations, which have deteriorated the enduring influences on the weather structure by altering the thermal configuration [68].
The road network map of the Lahore district is shown in Figure 6. Each road, minor or major, inside the Lahore district is digitized to produce the road network map. In ArcGIS software, Road Density (RD) is computed from the represented road network map by exercising the line density tool in the spatial analyst toolbox. The RD is calculated at the pixel level to ensure it matches the spatial resolution of LST data derived from MODIS. RD in this study represents the kilometers of road network per square kilometer area. The road network map signifies the effect of urbanization, and it can be observed that the built-up area has a massive road network. In comparison to the urban area, the measurement of roads in the rural area is exceptionally smaller. The RD of the Lahore district varies from 0 to 35.09 Km/ Km2. As illustrated in the RD map of the Lahore district, the built-up area has higher RD in contrast to the surrounding rural areas.
The DEM of the Lahore district formulated from ASTER GDEM data at a 24.8 m resolution is portrayed in Figure 7. Correspondingly, the altitude is undulant, with the lowermost and uppermost elevations of 170 m and 257 m. The resampled ASTER GDEM image (with the same resolution as the LST image) obtained to compare elevation values with LST is also portrayed in Figure 7. Subsequently, the resampling resulted in the reduction of the uppermost elevation from 257 m to 232 m and the amplification of the lowermost elevation from 170 m to 192 m due to the averaging of neighborhood pixels. Due to the fact that pixels of changed images signify identical areas on the ground, the DEM image has also been snapped with the LST image.
Several parameters, for instance, elevation, EVI, and RD, which are self-governing but associated with each other to some degree, are utilized for the current investigation of LST prediction. Still, each parameter in itself influences the LST distinctively, regardless of the connections they have. Therefore, it is vital to verify the collective influence of these factors on LST. Hence, the LSTM model is fostered for forecasting LST at any position, conforming to a group of contributing factors (elevation, EVI, and RD). For 2009 to 2019, 10 years of EVI and LST data are utilized to formulate the LSTM model, which is then used for the forecast of LST for 2020. The image format for the output of the LSTM model, formerly in discrete data format, is obtained using the ArcGIS software. Utilizing the ArcGIS software, the LST images resulting from MODIS data and conforming to LST images obtained from the LSTM model data for two periods of 2020, January (009–016 days) and May (137–144 days), are presented in Figure 8 and Figure 9, respectively. These figures show the spatial variation of observed and model-predicted LST values over the study area. As evident from the figures inside the built-up margin, a comparable spatial configuration is present. A similar spatial location is shown by the pixels exhibiting the top LST range, and the pixels just outside the built-up edge also have a comparable configuration. The variation in LST extent in the rural areas for the month of May compared to the month of January is due to the absence of crops in the fields.
An assessment of the noted and forecasted minimum, mean, and maximum LST values is presented in Table 2. All the noted and model-forecasted LST values are relatively closely related to each other, and for one of the periods, the standard deviation is also very close, as can be seen in Table 2. Thus, the forecasted LST values relate closely to the noted values for the two periods, as can be comprehended from Figure 8 and Figure 9, and Table 2. This indicates that the intended model has fairly effective competency in computing the disparities in the LST values with diverse pixels for the deliberated area. On the other hand, the model predictions for some of the pixels are not appropriate (Figure 8 and Figure 9), and there are some small-scale variations. However, it is certain from the current investigation that most of the LST values generated by the LSTM model are closely related to the detected values; therefore, it can be established that the model can be exercised for the forecast of LST with enhancements and modifications to achieve superior performance.
For the considered area, the model-predicted LST values are plotted along with the observed values for the 4081 overall pixels in order to relate the temporal disparity of the model-predicted and observed LST values. The observed and model-predicted LST values for all 4081 pixels for January (009–016 days) and May (137–144 days) are exhibited in Figure 10a,b, respectively. The illustration shows that close conformity exists between the model-forecasted and detected LST values for a considerable number of pixels. The pixel-wise comparison in Figure 10a,b also indicates that the predicted and actual values do not match some of the pixels. For some of the pixels, there are small-scale discrepancies in the prediction of LST values in urban areas, and there is a difference of over 2 Kelvin in urban pockets. Furthermore, the model forecasts both the elevated and minor LST values with considerable precision for both periods, as can be seen from Figure 10a,b.
A comparison was made between the predicted LST exercising the developed model (LSTM) and the observed LST data for the error computation with a reference model (ANN). The difference between detected and forecasted LST values for each pixel is obtained, and the absolute error in the LST forecasting from the practiced model is computed for the proposed model. The MSE and R2 values for the proposed and reference models are computed for comparison purposes.
For the two periods of 2020 (January and May), the MAE and MAPE of the proposed model are represented in Table 3. The MAPE for the two periods is practically the same, as evident from Table 3. The observed MAE is only 0.27 K for the two periods, which specifies that there are fewer disagreements and good agreement among observed and model-predicted values for the LSTM model. The extreme MAPE is merely 0.14%, thus demonstrating that the model is competent in forecasting the LST with elevated precision. Moreover, the MSE and R2 values for the two models are also listed in Table 3. The LSTM model has lower MSE for both periods than the ANN model. Furthermore, a high correlation exists between the LSTM predicted model and observed LST with R2 values of 0.95 and 0.94 for January and May, respectively. The correlation between the ANN model for the two periods is also relatively high.
The number of pixels of the investigative area, classified based on the error in the prediction of LST, for LSTM and ANN models are presented in Table 4 and Table 5, respectively. The implementation of the model has also been assessed utilizing these values. A positive error signifies that the noted value is greater than the model-predicted LST value, while the observed value is lower than predicted when the error is negative. The model forecasted the LST in diverse zones of the investigative area very precisely, as can be witnessed in Table 4. The error in the prediction of LST for approximately 43% of pixels in the investigated area is between −1.0 and 0.0 K for January, whereas, for May, 38% of the pixels in the investigated area have an error between 0.0 and 1.0 K. The proportion of pixels showing error past 1.0 K (a positive error showing that the model-predicted values are less than the observed LST values) is 27% and is only observed for May. However, for the ANN model, not only do the pixel proportions vary but also the ranges of negative and positive errors are more than in the LSTM model, as can be observed from Table 5. There can be a number of reasons for these different ranges of errors. One of the possible reasons might be the upscaling of DEM and RD of the area. The downscaling might have resulted in the loss of data. Another reason can be the coarse resolution of the MODIS data. Moreover, in this study, the urban area is only considered a road network function. These factors might have hampered the functioning of the model.
For the two periods, the spatial disparity of error in forecasting LST by the two models is exhibited in Figure 11 and Figure 12. In the prediction of LST, those pixels with errors are disseminated over the entire intended area. For January, as perceived from the Figure, the error among model-predicted and observed LST values is between 0.0 to 0.5 K for fewer pixels of the investigative area that specifically lie inside or near the built-up boundary. In contrast, for May, the error between the model-forecasted and noted LST is between 1.0 to 2.5 K for several pixels, which is more extensive than the first period. These pixels predominantly fall outside the urban boundary. These noted error values are due to the fact that for the urban area, the EVI values have lesser variation related to the EVI of rural areas. There is a significant change in the EVI of the rural area through diverse spells. Besides this, owing to the disparities in rainfall concentration, it also experiences annual changes. For elevated error in the rural area, another reason is perhaps the RD. One type of RD data is utilized for the current investigation. There is a possibility that the rural area has experienced some growth over 10 years through which several roads might have been built, consequently causing variations in the RD. On the other hand, in the urban area, this problem is not likely, as road broadening is achieved through later years of development despite constructing new roads on the already present roads in the area. It can also be perceived that the error in prediction of more pixels of the rural area in May is between 1.0 and 2.5 K. This difference can be because of the influence of seasonal variation as May is during the summer spell, and it is likely that extra weather variations are witnessed during this period in the rural area due to other factors such as bare areas, vegetation, etc. It is also evident from Figure 11a,b that the error in forecasting at a maximum of the explorative area for both periods is within the low to moderate range of the observed values. Moreover, a similar spatial trend of error distribution was observed for the ANN model but with a different error range and extent, as can be observed in Figure 12a,b.
Additionally, the MSE is used to calculate skill metrics and the related skill score centered on MSE for both the proposed and reference models. The ANN model’s MSE values for the January and May periods are 0.264 and 0.294, respectively (Table 3). There will be a precise forecast when the observed and the predicted values are undeniably equal and have 0.0 MSE, thus signifying a 0.0 value of the forecast skill metric, whereas the resulting skill score will be 1.0 If there is the same MSE for both the reference and forecast models and they are of equivalent skill, then the skill metric will be 1.0, whereas otherwise there will be a skill score of 0.0. None of the intended models in this situation can be stated as good or bad; both of them are of equal quality. The forecast model will be less skillful than the preferred model if the skill metric is more than 1.0, i.e., the MSE of the model is lower than the MSE of the forecast model, and there will be a negative value of the skill score factor. In this instance, there might be limitless skill score values to a negative extent.
For the two periods, the skill metric and skill score values for LSTM and ANN models are computed and are represented in Table 6. The skill metric falls among 0 and 1 for both periods, leading to absolute skill score values, i.e., for both of the models. However, the LSTM model is noted as relatively better than the ANN model. The skill metric values of 0.474 and 0.684 for the two periods for the LSTM model are higher than the ANN model. Similarly, skill score values of 0.426 and 0.228 for the periods for the LSTM model are also higher than the skill score values for the two periods for the ANN model. The better performance of LSTM over other deep learning and machine learning techniques has also been documented by several studies [35,36,37,38,69,70] in the past. Consequently, from previous years’ LST images, LST can be predicted using the LSTM model with reasonably acceptable results, given that adjustments and the differences in other factors, such as the climatic classification of the area, are considered.

5. Conclusions

In the current investigation, the LSTM model was used to predict LST by utilizing 10 years (2009–2019) of LST and EVI values alongside elevation and RD as input parameters to the LSTM model. The current investigation uses LST data from MODIS, and to include decent-quality data, the quality flag offered with these data is also utilized. The extreme error that is restricted by using the quality flag is ±2.5 K, and at no place in the study area does it exceed this value. It can be established from Figure 8, Figure 9 and Figure 10 that a suitable correlation exists among the noted and model-forecasted LST values. The MAE of the model is the same (0.27 k) for both periods, whereas the MAPE of the model varies from 0.12 % to 0.14 %. Most of the area expresses an error between −1.0 and 0.0 K for January, whereas, for May, it is between 0.0 and 1.0 k, which specifies a reasonably close correlation between observed and model-predicted LST values. Thus, it can be established that the predicted LST by the model is acceptable, and it can be furthered for SUHI analysis. The LSTM model has also been related to the ANN model for comparison purposes and accuracy assessment. The MSE and R2 values of the LSTM model are comparatively more acceptable than the ANN model. The skill metric and skill score values of the LSTM model are also computed. The computed skill metric and score values for both periods exhibit positive values, representing the LSTM model’s efficacy for better LST prediction.

5.1. Limitations

Several other indicators also impact LST and can be included to predict LST, for example, the population density and the division of urban functional areas (commercial areas/residential areas). However, the current study only considered three indicators, namely, EVI, RD, and elevation, to predict LST. This was because the study approached the objective from a remote sensing point of view rather than using any census or governmental data. The remote sensing data used are freely available, whereas this is not the case for data acquired from any type of governmental body. Moreover, the current study only captures the connection between LST and surface parameters. However, the atmospheric forcing environment is also critical in determining land–air energy and heat exchange and controlling LST.

5.2. Implications and Future Research

LST can be advantageous for governing SUHI influence. The predicted LST information can be utilized effectually for the strategic expansion of metropolitan Lahore to control the SUHI effect. The research could assist in the appraisal of the influence of the factors accountable for SUHI. The information available on SUHI, besides the various principles of planning, can also be exploited through the development plans, and it might permit city developers to propose the physical urban layout in such a way as to help the containment of SUHI related to the development. Moreover, in current times, when the whole world is lurching under the effects of climate variations, investigations of this extent may be beneficial in providing significant information regarding the SUHI trend, which can be utilized in formulating policies for alleviating its influences on the global climate.
Further studies can also consider using Landsat data with a 30 m resolution to derive LST instead of MODIS data of a 1 km resolution. Moreover, in addition to the indicators used in this study, future research can also focus on exploring the influence of other LST-related indicators, as mentioned in the limitations section.

Author Contributions

Conceptualization, U.K. and B.A.; methodology, U.K., U.A., L.L., and B.A.; software, L.L., I.U., and B.A.; validation, U.K., I.U., U.A., A.T. and B.A.; formal analysis, A.T.; investigation, L.L., Q.L. and A.T.; resources, A.T., L.L.,Q.L., B.A. and I.U.; data curation, B.A., A.T. and I.U.; writing—original draft preparation, U.K. and B.A.; writing—review and editing, U.K., B.A., I.U., Q.L., L.L. and A.T.; visualization, L.L. and A.T.; supervision A.T.; project administration, Q.L.; funding acquisition, Q.L. All authors have read and agreed to the published version of the manuscript.

Funding

This work is supported by the China high-resolution earth observation system (grant no. 03-Y30F03-9001-20/22) and National Natural Science Foundation of China (grant no. 42071321).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data presented in this study are available on request from the first or corresponding authors.

Acknowledgments

The authors wish to thank the editors and anonymous reviewers for their valuable comments and helpful suggestions. The authors would like to thank Stephen C. McClure for his enthusiastic support and valuable suggestions during the review of the manuscript.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Sharifi, A.; Mahdipour, H.; Moradi, E.; Tariq, A. Agricultural Field Extraction with Deep Learning Algorithm and Satellite Imagery. J. Indian Soc. Remote Sens. 2022, 50, 417–423. [Google Scholar] [CrossRef]
  2. Bera, D.; Das Chatterjee, N.; Mumtaz, F.; Dinda, S.; Ghosh, S.; Zhao, N.; Bera, S.; Tariq, A. Integrated Influencing Mechanism of Potential Drivers on Seasonal Variability of LST in Kolkata Municipal Corporation, India. Land 2022, 11, 1461. [Google Scholar] [CrossRef]
  3. Ullah, I.; Aslam, B.; Shah, S.H.I.A.; Tariq, A.; Qin, S.; Majeed, M.; Havenith, H.-B. An Integrated Approach of Machine Learning, Remote Sensing, and GIS Data for the Landslide Susceptibility Mapping. Land 2022, 11, 1265. [Google Scholar] [CrossRef]
  4. Fu, C.; Cheng, L.; Qin, S.; Tariq, A.; Liu, P.; Zou, K.; Chang, L. Timely Plastic-Mulched Cropland Extraction Method from Complex Mixed Surfaces in Arid Regions. Remote Sens. 2022, 14, 4051. [Google Scholar] [CrossRef]
  5. Haq, S.M.; Tariq, A.; Li, Q.; Yaqoob, U.; Majeed, M.; Hassan, M.; Fatima, S.; Kumar, M.; Bussmann, R.W.; Moazzam, M.F.U.; et al. Influence of Edaphic Properties in Determining Forest Community Patterns of the Zabarwan Mountain Range in the Kashmir Himalayas. Forests 2022, 13, 1214. [Google Scholar] [CrossRef]
  6. Majeed, M.; Lu, L.; Haq, S.M.; Waheed, M.; Sahito, H.A.; Fatima, S.; Aziz, R.; Bussmann, R.W.; Tariq, A.; Ullah, I.; et al. Spatiotemporal Distribution Patterns of Climbers along an Abiotic Gradient in Jhelum District, Punjab, Pakistan. Forests 2022, 13, 1244. [Google Scholar] [CrossRef]
  7. Tariq, A.; Riaz, I.; Ahmad, Z. Land surface temperature relation with normalized satellite indices for the estimation of spatio-temporal trends in temperature among various land use land cover classes of an arid Potohar region using Landsat data. Environ. Earth Sci. 2020, 79, 40. [Google Scholar] [CrossRef]
  8. Tariq, A.; Shu, H. CA-Markov chain analysis of seasonal land surface temperature and land use landcover change using optical multi-temporal satellite data of Faisalabad, Pakistan. Remote Sens. 2020, 12, 3402. [Google Scholar] [CrossRef]
  9. Siddiqui, S.; Ali Safi, M.W.; Rehman, N.U.; Tariq, A. Impact of Climate Change on Land use/Land cover of Chakwal District. Int. J. Econ. Environ. Geol. 2020, 11, 65–68. [Google Scholar] [CrossRef]
  10. Tariq, A.; Shu, H.; Li, Q.; Altan, O.; Khan, M.R.; Baqa, M.F.; Lu, L. Quantitative Analysis of Forest Fires in Southeastern Australia Using SAR Data. Remote Sens. 2021, 13, 2386. [Google Scholar] [CrossRef]
  11. Majeed, M.; Tariq, A.; Anwar, M.M.; Khan, A.M.; Arshad, F.; Mumtaz, F.; Farhan, M.; Zhang, L.; Zafar, A.; Aziz, M.; et al. Monitoring of land use–Land cover change and potential causal factors of climate change in Jhelum district, Punjab, Pakistan, through GIS and multi-temporal satellite data. Land 2021, 10, 1026. [Google Scholar] [CrossRef]
  12. Mumtaz, F.; Arshad, A.; Mirchi, A.; Tariq, A.; Dilawar, A.; Hussain, S.; Shi, S.; Noor, R.; Noor, R.; Daccache, A.; et al. Impacts of reduced deposition of atmospheric nitrogen on coastal marine eco-system during substantial shift in human activities in the twenty-first century. Geomat. Nat. Hazards Risk 2021, 12, 2023–2047. [Google Scholar] [CrossRef]
  13. Waqas, H.; Lu, L.; Tariq, A.; Li, Q.; Baqa, M.F.; Xing, J.; Sajjad, A. Flash flood susceptibility assessment and zonation using an integrating analytic hierarchy process and frequency ratio model for the chitral district, khyber pakhtunkhwa, pakistan. Water 2021, 13, 1650. [Google Scholar] [CrossRef]
  14. Tariq, A.; Shu, H.; Siddiqui, S.; Mousa, B.G.; Munir, I.; Nasri, A.; Waqas, H.; Lu, L.; Baqa, M.F. Forest fire monitoring using spatial-statistical and Geo-spatial analysis of factors determining forest fire in Margalla Hills, Islamabad, Pakistan. Geomat. Nat. Hazards Risk 2021, 12, 1212–1233. [Google Scholar] [CrossRef]
  15. Hamza, S.; Khan, I.; Lu, L.; Liu, H.; Burke, F.; Nawaz-ul-Huda, S.; Baqa, M.F.; Tariq, A. The Relationship between Neighborhood Characteristics and Homicide in Karachi, Pakistan. Sustainability 2021, 13, 5520. [Google Scholar] [CrossRef]
  16. Tariq, A.; Shu, H.; Kuriqi, A.; Siddiqui, S.; Gagnon, A.S.; Lu, L.; Linh, N.T.T.; Pham, Q.B. Characterization of the 2014 Indus River Flood Using Hydraulic Simulations and Satellite Images. Remote Sens. 2021, 13, 2053. [Google Scholar] [CrossRef]
  17. Felegari, S.; Sharifi, A.; Moravej, K.; Amin, M.; Golchin, A.; Muzirafuti, A.; Tariq, A.; Zhao, N. Integration of Sentinel 1 and Sentinel 2 Satellite Images for Crop Mapping. Appl. Sci. 2021, 11, 10104. [Google Scholar] [CrossRef]
  18. Baqa, M.F.; Chen, F.; Lu, L.; Qureshi, S.; Tariq, A.; Wang, S.; Jing, L.; Hamza, S.; Li, Q. Monitoring and modeling the patterns and trends of urban growth using urban sprawl matrix and CA-Markov model: A case study of Karachi, Pakistan. Land 2021, 10, 700. [Google Scholar] [CrossRef]
  19. Hu, P.; Sharifi, A.; Tahir, M.N.; Tariq, A.; Zhang, L.; Mumtaz, F.; Shah, S.H.I.A. Evaluation of vegetation indices and phenological metrics using time-series modis data for monitoring vegetation change in Punjab, Pakistan. Water 2021, 13, 2550. [Google Scholar] [CrossRef]
  20. Tariq, A.; Yan, J.; Gagnon, A.S.; Riaz Khan, M.; Mumtaz, F. Mapping of cropland, cropping patterns and crop types by combining optical remote sensing images with decision tree classifier and random forest. Geo-spatial Inf. Sci. 2022, 1–19. [Google Scholar] [CrossRef]
  21. Tariq, A.; Mumtaz, F.; Zeng, X.; Baloch, M.Y.J.; Moazzam, M.F.U. Spatio-temporal variation of seasonal heat islands mapping of Pakistan during 2000–2019, using day-time and night-time land surface temperatures MODIS and meteorological stations data. Remote Sens. Appl. Soc. Environ. 2022, 27, 100779. [Google Scholar] [CrossRef]
  22. Schwarz, N.; Lautenbach, S.; Seppelt, R. Exploring indicators for quantifying surface urban heat islands of European cities with MODIS land surface temperatures. Remote Sens. Environ. 2011, 115, 3175–3186. [Google Scholar] [CrossRef]
  23. Qiao, Z.; Tian, G.; Xiao, L. Diurnal and seasonal impacts of urbanization on the urban thermal environment: A case study of Beijing using MODIS data. ISPRS J. Photogramm. Remote Sens. 2013, 85, 93–101. [Google Scholar] [CrossRef]
  24. Imran, M.; Mehmood, A. Analysis and mapping of present and future drivers of local urban climate using remote sensing: A case of Lahore, Pakistan. Arab. J. Geosci. 2020, 13, 278. [Google Scholar] [CrossRef]
  25. Shirazi, S.A.; Kazmi, J.H. Analysis of socio-environmental impacts of the loss of urban trees and vegetation in Lahore, Pakistan: A review of public perception. Ecol. Processes 2016, 5, 5. [Google Scholar] [CrossRef]
  26. Levermore, G.; Parkinson, J.; Lee, K.; Laycock, P.; Lindley, S. The increasing trend of the urban heat island intensity. Urban Climate 2018, 24, 360–368. [Google Scholar] [CrossRef]
  27. Debbage, N.; Shepherd, J.M. The urban heat island effect and city contiguity. Comput. Environ. Urban Syst. 2015, 54, 181–194. [Google Scholar] [CrossRef]
  28. Mathew, A.; Sreekumar, S.; Khandelwal, S.; Kaul, N.; Kumar, R. Prediction of surface temperatures for the assessment of urban heat island effect over Ahmedabad city using linear time series model. Energy Build. 2016, 128, 605–616. [Google Scholar] [CrossRef]
  29. Mathew, A.; Sreekumar, S.; Khandelwal, S.; Kumar, R. Prediction of land surface temperatures for surface urban heat island assessment over Chandigarh city using support vector regression model. Sol. Energy 2019, 186, 404–415. [Google Scholar] [CrossRef]
  30. Gobakis, K.; Kolokotsa, D.; Synnefa, A.; Saliari, M.; Giannopoulou, K.; Santamouris, M. Development of a model for urban heat island prediction using neural network techniques. Sustain. Cities Soc. 2011, 1, 104–115. [Google Scholar] [CrossRef]
  31. Lee, Y.Y.; Kim, J.T.; Yun, G.Y. The neural network predictive model for heat island intensity in Seoul. Energy Build. 2016, 110, 353–361. [Google Scholar] [CrossRef]
  32. Kolokotroni, M.; Davies, M.; Croxford, B.; Bhuiyan, S.; Mavrogianni, A. A validated methodology for the prediction of heating and cooling energy demand for buildings within the Urban Heat Island: Case-study of London. Sol. Energy 2010, 84, 2246–2255. [Google Scholar] [CrossRef]
  33. Khalil, U.; Aslam, B.; Azam, U.; Khalid, H.M.D. Time Series Analysis of Land Surface Temperature and Drivers of Urban Heat Island Effect Based on Remotely Sensed Data to Develop a Prediction Model. Appl. Artif. Intell. 2021, 35, 1803–1828. [Google Scholar] [CrossRef]
  34. Yu, X.; Shi, S.; Xu, L.; Liu, Y.; Miao, Q.; Sun, M. A novel method for sea surface temperature prediction based on deep learning. Math. Probl. Eng. 2020, 2020, 1–9. [Google Scholar] [CrossRef]
  35. Zhao, R.; Yan, R.; Chen, Z.; Mao, K.; Wang, P.; Gao, R.X. Deep learning and its applications to machine health monitoring. Mech. Syst. Signal Processing 2019, 115, 213–237. [Google Scholar] [CrossRef]
  36. Chen, L.; Liu, X.; Zeng, C.; He, X.; Chen, F.; Zhu, B. Temperature Prediction of Seasonal Frozen Subgrades Based on CEEMDAN-LSTM Hybrid Model. Sensors 2022, 22, 5742. [Google Scholar] [CrossRef]
  37. Tsokov, S.; Lazarova, M.; Aleksieva-Petrova, A. A Hybrid Spatiotemporal Deep Model Based on CNN and LSTM for Air Pollution Prediction. Sustainability 2022, 14, 5104. [Google Scholar] [CrossRef]
  38. Zhang, J.; Zhu, Y.; Zhang, X.; Ye, M.; Yang, J. Developing a Long Short-Term Memory (LSTM) based model for predicting water table depth in agricultural areas. J. Hydrol. 2018, 561, 918–929. [Google Scholar] [CrossRef]
  39. Garg, S.; Jindal, H. Evaluation of time series forecasting models for estimation of PM2.5 levels in air. In Proceedings of the 2021 6th International Conference for Convergence in Technology (I2CT), Maharashtra, India, 2–4 April 2021; pp. 1–8. [Google Scholar]
  40. Rana, I.A.; Bhatti, S.S. Lahore, Pakistan–Urbanization challenges and opportunities. Cities 2018, 72, 348–355. [Google Scholar] [CrossRef]
  41. Liang, B.; Weng, Q. Assessing urban environmental quality change of Indianapolis, United States, by the remote sensing and GIS integration. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2010, 4, 43–55. [Google Scholar] [CrossRef]
  42. Mackey, C.W.; Lee, X.; Smith, R.B. Remotely sensing the cooling effects of city scale efforts to reduce urban heat island. Build. Environ. 2012, 49, 348–358. [Google Scholar] [CrossRef]
  43. Deng, C.; Wu, C. Examining the impacts of urban biophysical compositions on surface urban heat island: A spectral unmixing and thermal mixing approach. Remote Sens. Environ. 2013, 131, 262–274. [Google Scholar] [CrossRef]
  44. Guo, G.; Wu, Z.; Xiao, R.; Chen, Y.; Liu, X.; Zhang, X. Impacts of urban biophysical composition on land surface temperature in urban heat island clusters. Landsc. Urban Plan. 2015, 135, 1–10. [Google Scholar] [CrossRef]
  45. Mathew, A.; Khandelwal, S.; Kaul, N. Investigating spatial and seasonal variations of urban heat island effect over Jaipur city and its relationship with vegetation, urbanization and elevation parameters. Sustain. Cities Soc. 2017, 35, 157–177. [Google Scholar] [CrossRef]
  46. Alshaikh, A.Y. Space applications for drought assessment in Wadi-Dama (West Tabouk), KSA. Egypt. J. Remote Sens. Space Sci. 2015, 18, S43–S53. [Google Scholar] [CrossRef]
  47. Zhang, K.; Wang, R.; Shen, C.; Da, L. Temporal and spatial characteristics of the urban heat island during rapid urbanization in Shanghai, China. Environ. Monit. Assess. 2010, 169, 101–112. [Google Scholar] [CrossRef]
  48. Lazzarini, M.; Marpu, P.R.; Ghedira, H. Temperature-land cover interactions: The inversion of urban heat island phenomenon in desert city areas. Remote Sens. Environ. 2013, 130, 136–152. [Google Scholar] [CrossRef]
  49. Mathew, A.; Khandelwal, S.; Kaul, N. Spatial and temporal variations of urban heat island effect and the effect of percentage impervious surface area and elevation on land surface temperature: Study of Chandigarh city, India. Sustain. Cities Soc. 2016, 26, 264–277. [Google Scholar] [CrossRef]
  50. Huete, A.; Justice, C.; Van Leeuwen, W. MODIS vegetation index (MOD13). In Algorithm Theoretical Basis Document; University of Arizona: Tucson, AZ, USA, 1999; Version 3. [Google Scholar]
  51. Birhane, E.; Ashfare, H.; Fenta, A.A.; Hishe, H.; Gebremedhin, M.A.; Solomon, N. Land use land cover changes along topographic gradients in Hugumburda national forest priority area, Northern Ethiopia. Remote Sens. Appl. Soc. Environ. 2019, 13, 61–68. [Google Scholar] [CrossRef]
  52. Zhao, X.; Pu, J.; Wang, X.; Chen, J.; Yang, L.E.; Gu, Z. Land-use spatio-temporal change and its driving factors in an artificial forest area in Southwest China. Sustainability 2018, 10, 4066. [Google Scholar] [CrossRef] [Green Version]
  53. Imhoff, M.L.; Zhang, P.; Wolfe, R.E.; Bounoua, L. Remote sensing of the urban heat island effect across biomes in the continental USA. Remote Sens. Environ. 2010, 114, 504–513. [Google Scholar] [CrossRef]
  54. Zhang, Y.; Odeh, I.O.; Han, C. Bi-temporal characterization of land surface temperature in relation to impervious surface area, NDVI and NDBI, using a sub-pixel image analysis. Int. J. Appl. Earth Obs. Geoinf. 2009, 11, 256–264. [Google Scholar] [CrossRef]
  55. Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735–1780. [Google Scholar] [CrossRef] [PubMed]
  56. Gers, F.A.; Schmidhuber, J.; Cummins, F. Learning to forget: Continual prediction with LSTM. Neural Comput. 2000, 12, 2451–2471. [Google Scholar] [CrossRef]
  57. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv 2014, arXiv:1412.3555. [Google Scholar] [CrossRef]
  58. Gers, F.A.; Schraudolph, N.N.; Schmidhuber, J. Learning precise timing with LSTM recurrent networks. J. Mach. Learn. Res. 2002, 3, 115–143. [Google Scholar]
  59. Graves, A.; Mohamed, A.-R.; Hinton, G. Speech recognition with deep recurrent neural networks. In Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 26–31 May 2013; pp. 6645–6649. [Google Scholar]
  60. Graves, A. Long short-term memory. In Supervised Sequence Labelling with Recurrent Neural Networks; Springer: Berlin/Heidelberg, Germany, 2012; pp. 37–45. [Google Scholar]
  61. Sak, H.; Senior, A.W.; Beaufays, F. Long short-term memory recurrent neural network architectures for large scale acoustic modeling. In Proceedings of the INTERSPEECH 2014, 15th Annual Conference of the International Speech Communication Association, Singapore, 14–18 September 2014. [Google Scholar]
  62. Wheatcroft, E. Interpreting the skill score form of forecast performance metrics. Int. J. Forecast. 2019, 35, 573–579. [Google Scholar] [CrossRef]
  63. Li, Y.-Y.; Zhang, H.; Kainz, W. Monitoring patterns of urban heat islands of the fast-growing Shanghai metropolis, China: Using time-series of Landsat TM/ETM+ data. Int. J. Appl. Earth Obs. Geoinf. 2012, 19, 127–138. [Google Scholar] [CrossRef]
  64. Tiangco, M.; Lagmay, A.; Argete, J. ASTER-based study of the night-time urban heat island effect in Metro Manila. International J. Remote Sens. 2008, 29, 2799–2818. [Google Scholar] [CrossRef]
  65. Mathew, A.; Chaudhary, R.; Gupta, N.; Khandelwal, S.; Kaul, N. Study of urban heat island effect on Ahmedabad City and its relationship with urbanization and vegetation parameters. Int. J. Comput. Math. Sci. 2015, 4, 126–135. [Google Scholar]
  66. Pal, S.; Ziaul, S. Detection of land use and land cover change and land surface temperature in English Bazar urban centre. Egypt. J. Remote Sens. Space Sci. 2017, 20, 125–145. [Google Scholar] [CrossRef]
  67. Zhang, H.; Qi, Z.-F.; Ye, X.-Y.; Cai, Y.-B.; Ma, W.-C.; Chen, M.-N. Analysis of land use/land cover change, population shift, and their effects on spatiotemporal patterns of urban heat islands in metropolitan Shanghai, China. Appl. Geogr. 2013, 44, 121–133. [Google Scholar] [CrossRef]
  68. Tariq, A.; Siddiqui, S.; Sharifi, A.; Hassan, S.; Ahmad, I. Impact of spatio—Temporal land surface temperature on cropping pattern and land use and land cover changes using satellite imagery, Hafizabad District, Punjab, Province of Pakistan. Arab. J. Geosci. 2022, 15, 1045. [Google Scholar] [CrossRef]
  69. Le, X.-H.; Ho, H.V.; Lee, G.; Jung, S. Application of long short-term memory (LSTM) neural network for flood forecasting. Water 2019, 11, 1387. [Google Scholar] [CrossRef]
  70. Liu, Y.; Li, D.; Wan, S.; Wang, F.; Dou, W.; Xu, X.; Li, S.; Ma, R.; Qi, L. A long short-term memory-based model for greenhouse climate prediction. Int. J. Intell. Syst. 2022, 37, 135–151. [Google Scholar] [CrossRef]
Sustainability 14 11873 g001 550
Figure 1. Map of the study area showing the elevation of the area and extent of urban boundary.
Figure 1. Map of the study area showing the elevation of the area and extent of urban boundary.
Sustainability 14 11873 g001
Sustainability 14 11873 g002 550
Figure 2. Description of single LSTM cell.
Figure 2. Description of single LSTM cell.
Sustainability 14 11873 g002
Sustainability 14 11873 g003 550
Figure 3. Graphical representation of the training model.
Figure 3. Graphical representation of the training model.
Sustainability 14 11873 g003
Sustainability 14 11873 g004 550
Figure 4. Spatial and temporal variation of LST over the Lahore district for January (a) and May (b) 2019.
Figure 4. Spatial and temporal variation of LST over the Lahore district for January (a) and May (b) 2019.
Sustainability 14 11873 g004
Sustainability 14 11873 g005 550
Figure 5. EVI images of the Lahore district for January (a) and May (b) for the year 2019.
Figure 5. EVI images of the Lahore district for January (a) and May (b) for the year 2019.
Sustainability 14 11873 g005
Sustainability 14 11873 g006 550
Figure 6. Road network (a) and Road Density (RD) Map (b) of the Lahore District.
Figure 6. Road network (a) and Road Density (RD) Map (b) of the Lahore District.
Sustainability 14 11873 g006
Sustainability 14 11873 g007 550
Figure 7. DEM (a) and resampled DEM (b) of the Lahore District.
Figure 7. DEM (a) and resampled DEM (b) of the Lahore District.
Sustainability 14 11873 g007
Sustainability 14 11873 g008 550
Figure 8. Spatial variation of observed (a) and model-predicted (b) LST for January.
Figure 8. Spatial variation of observed (a) and model-predicted (b) LST for January.
Sustainability 14 11873 g008
Sustainability 14 11873 g009 550
Figure 9. Spatial variation of observed (a) and model-predicted (b) LST for May.
Figure 9. Spatial variation of observed (a) and model-predicted (b) LST for May.
Sustainability 14 11873 g009
Sustainability 14 11873 g010 550
Figure 10. Comparison of observed and model-predicted LST for all the pixels of the investigative area for January (a) and May (b).
Figure 10. Comparison of observed and model-predicted LST for all the pixels of the investigative area for January (a) and May (b).
Sustainability 14 11873 g010
Sustainability 14 11873 g011 550
Figure 11. Spatial variation of error between observed LST and LSTM model-predicted LST for January (a) and May (b).
Figure 11. Spatial variation of error between observed LST and LSTM model-predicted LST for January (a) and May (b).
Sustainability 14 11873 g011
Sustainability 14 11873 g012 550
Figure 12. Spatial variation of error between observed LST and ANN model-predicted LST for January (a) and May (b).
Figure 12. Spatial variation of error between observed LST and ANN model-predicted LST for January (a) and May (b).
Sustainability 14 11873 g012
Table
Table 1. Specifics of the used remote sensing data in this study.
Table 1. Specifics of the used remote sensing data in this study.
Remote Sensing ProductShort NameSensorPlatformTemporal ResolutionSpatial Resolution (m)
Land Surface Temperature and EmissivityMYD11A2MODISAqua8 Day926.626
Vegetation IndicesMYD13A2MODISAqua16 Day926.626
Land Cover TypeMCD12Q1MODISCombined Aqua and TeraYearly463.3
Global Digital Elevation ModelASTGTMASTERTerra_24.8
Table
Table 2. Comparison of 2020 maximum, minimum, and mean noted and forecasted LST values.
Table 2. Comparison of 2020 maximum, minimum, and mean noted and forecasted LST values.
Days No.TypeMaximumMinimumMeanStandard Deviation
009–016Actual293.18286.86290.021.427
Predicted293.51288.94291.221.348
137–144Actual317.38307.14312.261.452
Predicted315.41308.92312.161.416
Table
Table 3. MAE, MAPE, and MSE for both of the months.
Table 3. MAE, MAPE, and MSE for both of the months.
ModelMeasuresJanuary (009–016)May (137–144)
MAE (K)0.270.27
LSTM (Forecasted)MAPE (%)0.120.14
MSE
R2		0.242
0.95		0.253
0.94
ANN (Referred)MSE0.2640.296
R20.930.93
Table
Table 4. Number of pixels (% of pixels) of the investigative area in different error ranges for LSTM model.
Table 4. Number of pixels (% of pixels) of the investigative area in different error ranges for LSTM model.
Time PeriodBetween
−3.0 to −4.0 K		Between
−2.0 to −3.0 K		Between
−1.0 to −2.0 K		Between
−1.0 to 0.0 K		Between
0.0 to 0.5 K
January (009–016)204 (5%)449 (11%)1551 (38%)1755 (43%)122 (3%)
Between
−2.0 to −2.5 K		Between
−1.0 to −2.0 K		Between
−1.0 to 0.0 K		Between
0.0 to 1.0 K		Between
1.0 to 2.5 K
May (137–144)286 (7%)122 (3%)1020 (25%)1551 (38%)1102 (27%)
Table
Table 5. Number of pixels (% of pixels) of the investigative area in different error ranges for ANN model.
Table 5. Number of pixels (% of pixels) of the investigative area in different error ranges for ANN model.
Time PeriodBetween
−5.0 to −3.0 K		Between
−3.0 to −1.0 K		Between
−1.0 to 0.0 K		Between
0.0 to 1.0 K		Between
1.0 to 0.2 K
January (009–016)219 (5%)543 (13%)1461 (36%)1583 (39%)275 (7%)
Between
−2.0 to −1.0 K		Between
−1.0 to 0.0 K		Between
0.0 to 2.0 K		Between
2.0 to 3.0 K		Between
3.0 to 5.0 K
May (137–144)687 (17%)623 (15%)892 (22%)1243 (30%)636 (16%)
Table
Table 6. Skill metric and skill score values for the two periods.
Table 6. Skill metric and skill score values for the two periods.
ModelDays No.Skill MetricSkill Score
LSTM90.4740.426
730.6840.228
ANN90.4120.365
730.6250.185
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Khalil, U.; Azam, U.; Aslam, B.; Ullah, I.; Tariq, A.; Li, Q.; Lu, L. Developing a Spatiotemporal Model to Forecast Land Surface Temperature: A Way Forward for Better Town Planning. Sustainability 2022, 14, 11873. https://doi.org/10.3390/su141911873

AMA Style

Khalil U, Azam U, Aslam B, Ullah I, Tariq A, Li Q, Lu L. Developing a Spatiotemporal Model to Forecast Land Surface Temperature: A Way Forward for Better Town Planning. Sustainability. 2022; 14(19):11873. https://doi.org/10.3390/su141911873

Chicago/Turabian Style

Khalil, Umer, Umar Azam, Bilal Aslam, Israr Ullah, Aqil Tariq, Qingting Li, and Linlin Lu. 2022. "Developing a Spatiotemporal Model to Forecast Land Surface Temperature: A Way Forward for Better Town Planning" Sustainability 14, no. 19: 11873. https://doi.org/10.3390/su141911873

APA Style

Khalil, U., Azam, U., Aslam, B., Ullah, I., Tariq, A., Li, Q., & Lu, L. (2022). Developing a Spatiotemporal Model to Forecast Land Surface Temperature: A Way Forward for Better Town Planning. Sustainability, 14(19), 11873. https://doi.org/10.3390/su141911873

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop