Spatial Variability of Land Surface Temperature of a Coal Mining Region Using a Geographically Weighted Regression Model: A Case Study

Abstract

Coal accounts for over half of India’s energy needs currently. However, it has resulted in significant environmental impacts such as altering land cover and land surface temperatures. This study quantifies the land surface temperature (LST) of Dhanbad City (India)—home to India’s largest coal reserves. It uses the Landsat 8 image data to evaluate urban and rural temperature variations across different land use–land cover (LULC) classes. Using a Geographically Weighted Regression Model (GWR), we examined the spatial heterogeneity of the LST using key environmental indices, such as the Normalized Difference Vegetation Index (NDVI), Modified Normalized Difference Water Index (MNDWI), Normalized Difference Moisture Index (NDMI), and Normalized Difference Barren Index (NDBI). The seasonal LST variations revealed significant urban–rural area temperature disparities, with rural regions exhibiting stronger correlations with the key indices above. The GWR model accounted for 78.31% of the spatial variability in LST, with unexplained heterogeneity in urban areas linked to anomalies identified in the coal mining area fire map. These findings underscore the necessity of targeted mitigation strategies to reduce high LST values in coal fire-affected regions, with localized spatial measures in mining areas.

1. Introduction

The increase in population across the world has resulted in rapid urbanization and industrialization that have not only resulted in changes in landscape, urban geometry, and climate change worldwide [1,2,3] but have also heightened the demand for fuel, energy, and raw materials, leading to expanded mining activities evident across the globe [4]. Coal remains a critical energy source in India and worldwide. The European Union also added coking coal as a raw material of high economic importance in 2014, ranking second right after Tungsten [5]. The International Energy Agency’s 2023 report [4] estimated that India’s coal consumption reached 105 mt due to rapid economic growth and reduced hydropower outputs.

Fuel mineral exploitation activities result in substantial environmental impacts, including changes in land use and land cover (LULC), soil quality, and air and water resources [6]. The coal industry serves as a substantial contributor to greenhouse gas (GHG) emissions throughout its entire lifecycle, from extraction to utilization, thereby contributing to global warming and climate change [7,8,9]. Additionally, coal’s self-heating and spontaneous combustion characteristics present a critical challenge in mining regions, leading to elevated land surface temperatures (LSTs) and potential disruptions to local ecosystems and climate change dynamics [10,11]. Jaiswal et al. [12] investigated the impact of coal mining in a region for over two decades and observed increased temperatures in mining and adjacent urban areas. They attributed these to the expansion of mining operations and deforestation. Li et al. [13] conducted an Ecological Cumulative Effect (ECE) assessment in a mining region and determined that increasing mining activities can significantly influence the LST. However, measures such as ecological restoration and green mine construction can mitigate and stabilize the rate of LST change. Similarly, Chaddad et al. [14] investigated the impact of mining-induced deforestation on soil surface temperature in mining regions of the Amazon rainforest and reported a 10 °C increase in LST over a 15-year period. Other studies have also linked land use changes to variations in LST in mining-adjacent areas [12,15,16].

The study of underlying factors influencing the LST has become essential in urban climate research, with remote sensing playing a pivotal role in such analyses [6]. Climatic parameters such as ambient temperature, wind direction, wind speed, humidity, incoming solar radiation, and time of day influence the LST of the urban microclimate. Additionally, various spatial and anthropogenic factors, including population density, building height, building density, surface material, air pollution, distance from roads, urban green spaces, and proximity to water bodies, also contribute to variations in LST [1,17,18,19]. Spectral indices derived from satellite imagery, such as the Normalized Difference Vegetation Index (NDVI), Modified Normalized Difference Water Index (MNDWI), Normalized Difference Barren Index (NDBI), and Normalized Difference Moisture Index (NDMI), have been extensively employed to investigate LST variations [16,20,21,22,23]. Research consistently shows that green spaces and water bodies mitigate elevated LST impacts, whereas built-up areas and barren soils exacerbate LST impacts [24,25]. Urban green spaces have been shown to significantly reduce the LST in hot and dry regions [25,26,27]. However, their cooling effect is comparatively weaker in hot and humid regions [25,28]. LST studies over mining areas have also found a high correlation between the LST, land use type, vegetation coverage, coal mining activities, coal waste dumps, deforestation, and coal fires [11,14,29,30,31].

Researchers mostly rely on conventional statistical analyses such as Ordinary Least Square (OLS) and simple linear regression to explore these relationships [32,33,34,35]. However, these have limited effectiveness when spatial data exhibit high multicollinearity among independent variables or when local factors can influence the LST outcome [36]. Recent studies have integrated vector-based 3D analysis to examine local climate zones (LCZs) and the underlying factors influencing microclimatic conditions [37,38]. Zhang et al. [39] found significant improvement in LCZ microclimate analysis using ENVI-met over microclimate analysis using remote sensing-based LSTs. However, most of these studies were conducted at the microscale, and the development of large-scale 3D models can be resource-intensive and computationally intensive, particularly in cases where high-resolution secondary data are not readily available. The Geographically Weighted Regression (GWR) model has emerged as a promising alternative for studying spatial heterogeneity and local relationships [32]. GWR accounts for spatial heterogeneity by modeling variations between dependent and independent variables, incorporating neighborhood- and distance-based factors. Unlike conventional statistics like OLS and simple linear regression, which work on global regression instead of local regression and do not take spatial relationships between dependent and independent variables into account, GWR can model the spatial relationships between dependent and independent variables, thus giving better results where local factors are important [32]. Numerous studies have demonstrated that GWR provides a better fit and yields better localized insights compared to traditional regression models like OLS, making it a valuable tool for explicit LST analyses [15,39,40,41]. Siqi et al. [15] found that Geographically Weighted Regression (GWR) exhibited a stronger correlation between land surface temperature (LST) and underlying variables compared to Ordinary Least Squares (OLS) in the city of Hong Kong. Similarly, using GWR and OLS regression, Kashki et al. [42] analyzed the relationship between LST and underlying factors in Shiraz, Iran. Their findings indicated that while OLS achieved an R² of only 0.29, GWR significantly outperformed it with an R² of 0.95. Additionally, Jia et al. [42] employed a combination of deep learning and GWR for LST prediction and found that GWR provided more accurate predictions for impervious surfaces and water bodies compared to vegetation. With most studies on the application of GWR and LST focusing on large urban centers and there being limited such research in coal mining areas [30] involving different spectral indices and fires near mining regions, there is an identified need for such research in this area [15,31,39,41,43].

This research addresses the above gap by employing the GWR model to analyze LST anomalies in the coal mining region of Dhanbad City, India, and adjoining rural areas. The Jharia coal field (JCF) is one of the major coal fire-affected mining areas and has been on fire for more than 100 years [44]. This continuous burning has not only impacted the environment through the continuous emission of harmful gases from coal fires but has also taken numerous lives [45]. These regions have had minimal attention in remote sensing-based LST dynamics studies and offer unique challenges with the prevalence of coal mine fires. This research analyzes the spatial distribution of NDVI, MNDWI, NDMI, NDBI, and LST across various land classes in urban and rural regions, with a seasonal comparison of LST distribution between summer and winter. Additionally, coal fire-affected areas were delineated using LST data, and a comparative analysis was conducted to examine the relationship between the LST and individual spectral indices in urban and rural settings. Furthermore, this study explores the potential of Geographically Weighted Regression (GWR) in assessing the influence of local variables on the LST and identifying anomalies attributed to these factors using remotely sensed data in the mining region. The study utilizes the Landsat 8 OLI dataset with a high thermal resolution of 100 m. This study models spatial heterogeneity and identifies factors that can better explain LST variations in mining-affected areas. In this context, this study contributes to the broader understanding of mining-induced thermal anomalies while providing insights that can help in future city planning and mitigating heat islands in fire-affected coal mine regions.

2. Description of Case Study Area

This study surrounds Dhanbad City in Jharkhand, India, located near extensive open-cast coal mines, which lie just north of the mining operations (Figure 1). It is geographically bounded between latitudes 23°45′00″ N and 24°00′00″ N and longitudes 86°20′00″ E and 86°30′00″ E in the WGS84 coordinate system, corresponding to Indian topographic sheet number F45C5. Commonly known as the “Coal Capital of India,” it holds significant economic importance for the country due to extensive coal reserves and contributes 41% of the state’s total revenue from mining and quarrying activities [46]. The climate has distinct seasonal variations. During the winter (December to February), the mean maximum and minimum temperatures are 25.4 °C and 10.5 °C, respectively. During the summer (March to early June), the average temperatures are higher, with mean maximum and minimum values of 38.1 °C and 23.4 °C. The average annual rainfall is 1300.5 mm, with the majority (approximately 84%) occurring during the monsoon season from June to September. Minor rainfall events also occur in January and February, influenced by local meteorological conditions [47].

The area encompasses about 47,670 ha with an average elevation of 227 m above mean sea level. The study area was divided into two distinct zones: (a) the Southern Zone (urban region), which predominantly includes mining zones and the urban expansion of Dhanbad City, and (b) the Northern Zone (rural region), which is characterized by open landscapes such as grasslands and agricultural lands, with the Tundi Forest Reserve located around the extreme northern boundary. The study area is dynamic, with land cover patterns continually evolving due to mining operations, mine closures, land reclamation activities, and changes in urban settlements and vegetation cover. The northern portion features scattered rural villages, while the southern portion includes densely populated urban areas adjacent to active open-cast coal mines.

3. Data and Methodology

Landsat 8 Operational Land Imager (OLI) data for the year 2020 were utilized for the calculation of various indices (

Table 1. Satellite data for research.

Satellite	Sensor	Collection and Level	Year	Season, Date and Time *		Band Name	Wavelength (Micrometers)	Spatial Resolution (Meters)
				Summer	Winter
Landsat 8	Operational Land Imager (OLI)	Collection 2, Level 2	2020	6 May 2020 10:12:22.68	15 January 2020 10:13:8.38	Band 2—Blue	0.452–0.512	30
						Band 3—Green	0.533–0.590
						Band 4—Red	0.636–0.673
						Band 5—Near-Infrared (NIR)	0.851–0.879
						Band 6—Shortwave Infrared (SWIR) 1	1.566–1.651
						Band 7—Shortwave Infrared (SWIR) 2	2.107–2.294
Landsat 8	Thermal Infrared Sensor (TIRS)	Collection 2, Level 1	2020	6 May 2020 10:12:22.68	15 January 2020 10:13:8.38	Band 10—Thermal Infrared (TIR) 1	10.600–11.190	100 (resampled to 30)

* Time in Indian Standard Time (IST).

). Landsat Collection 2, Level 2 datasets were used for the classification of LULC and calculating indices for the analysis, which provides atmospherically corrected surface reflectance values suitable for calculating indices. For the analysis of LST, this study utilizes Landsat Thermal Infrared Sensor (TIRS) Collection 2, Level 1 datasets, which provide top-of-atmosphere (TOA) reflectance values. The data were acquired from the USGS Earth Explorer platform, with the selection criterion ensuring less than 5% cloud cover during both the summer and winter seasons. For the summer season, images corresponding to the highest recorded temperatures in the study area were chosen, while for the winter season, pixels representing the lowest temperatures were selected. All raster data processing was performed using ArcGIS Pro v3.1, and additional data analysis and processing were carried out using Microsoft Excel and R Studio v2024.04.2+764. The methodology flowchart and the datasets used are presented in Figure 2, with detailed technical information on the satellite bands used in the study summarized in Table 1.

3.1. Conversion of Digital Number (DN) Values to Surface Reflectance

The OLI satellite data were converted to surface reflectance values, and the TIRS images were converted to their respective TOA radiance value from their 16-bit Digital Number (DN) using Equation (1):

$$\rho =(DN\times GAIN)\pm BIAS$$

(1)

where $\rho$ is surface reflectance, $DN$ is the pixel value, and $GAIN$ and $BIAS$ are the multiplicative and additive parameters given in the metadata file of the satellite image.

3.2. Land Use–Land Cover

LULC is critical for understanding the characteristics of the land surface and the dynamics of urban and rural activities. This study area incorporates key LULC features developed from satellite imagery, classified into seven classes, i.e., water bodies, built-up areas, barren land, mining areas, grasslands, wetlands, and dense vegetation. The combination of the B, G, R, NIR, SWIR1, and SWIR2 surface reflectance bands of Landsat 8 OLI was used due to high spatial resolution for the classification of LULC. The classification was performed on the composite satellite image using the support vector machine (SVM) [48] classifier available in the ArcGIS Pro v3.1 software with a minimum of 200 samples per class spread throughout the map. Random stratified sampling was used to generate 361 reference points for the accuracy assessment of the LULC map generated.

3.3. Analysis Indices

This study used the following indices to make inferences.

3.3.1. Normalized Difference Vegetation Index (NDVI)

The NDVI is one of the most widely used and earliest remote sensing analytical tools for vegetation assessment [49]. It leverages the red (RED) and near-infrared (NIR) bands of satellite data to differentiate vegetation from water-based areas through the high reflectance of vegetation in the NIR spectrum and the low reflectance of water. The NDVI has proven effective in distinguishing forested and non-forested areas and is mainly used to estimate parameters such as the leaf area index (LAI), biomass, chlorophyll concentration in leaves, plant productivity, fractional vegetation cover, and plant stress [49].

Vegetation assessment is crucial since it directly impacts local temperature regulation. Increased vegetation cover enhances evapotranspiration rates, which helps to mitigate high temperatures and maintain thermal balance in the environment. The NDVI ranges from −1 to 1, with 1 indicating high vegetation density and −1 indicating water and barren areas with no vegetation. This makes the NDVI an invaluable tool for studying the vegetation phenomenon and its influence on land surface temperature. The NDVI can be calculated using the formula given below [50]:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\mathrm{D}}}$$

(2)

3.3.2. Modified Normalized Difference Water Index (MNDWI)

The MNDWI is widely used to distinguish water bodies from non-water regions. This method utilizes the green (GREEN) and short-wave infrared 1 (SWIR1) spectral bands to produce raster values ranging from −1 to 1 [51]. Typically, water bodies are represented by positive or near-zero values, while negative values characterize non-water regions. Interpreting the MNDWI is particularly valuable for identifying areas with low surface water availability, which can directly and significantly impact land surface temperature (LST). By providing spatially explicit insights into water distribution, the MNDWI is a crucial tool for understanding hydrological characteristics and their relationship with thermal patterns in various landscapes.

$$\mathrm{M}\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1}{\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1}}$$

(3)

3.3.3. Normalized Difference Moisture Index (NDMI)

The NDMI is commonly used to monitor crop water stress and differentiate between waterlogged and water-stressed regions. This index is calculated using the ratio of the difference to the sum of the near-infrared (NIR) and short-wave infrared 1 (SWIR1) bands [52]. NDMI values range between −1 and 1, where lower values typically correspond to built-up, barren and mining areas, while higher values are associated with vegetation and water bodies. This characteristic makes the NDMI particularly useful for distinguishing between various land cover types, particularly in regions affected by different levels of moisture availability.

$$\mathrm{N}\mathrm{D}\mathrm{M}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1}}$$

(4)

3.3.4. Normalized Difference Barren Index (NDBI)

The NDBI is derived from the near-infrared (NIR) and short-wave infrared 2 (SWIR2) spectral bands and is designed to identify barren regions in contrast to non-barren areas. This index is particularly important for distinguishing barren landscapes, such as bare land and mining areas, which play a critical role in understanding the LST phenomenon.

The NDBI values range from −1 to 1, with −1 representing non-barren regions and +1 corresponding to barren regions. Compared to the traditional barren index that utilizes the SWIR1 and NIR bands, the inclusion of SWIR2 instead of the SWIR1 band offers improved accuracy in distinguishing barren from non-barren areas, as analyzed in this study. This enhancement is especially valuable for more precise land cover classification in regions influenced by mining and other human activities to distinguish the barren and non-barren regions better, as discussed in the Results section.

$$\mathrm{N}\mathrm{D}\mathrm{B}\mathrm{I}={\displaystyle \frac{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2-\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2+\mathrm{N}\mathrm{I}\mathrm{R}}}$$

(5)

3.4. Land Surface Temperature (LST)

LST was retrieved using the radiative transfer equation (RTE), given below by [53]

$$L_{\lambda }^s=\left[\epsilon B_{\lambda }\left(T_s\right)+\left(1-\epsilon \right)L_{atm\lambda }^{\downarrow }\right]\tau +L_{atm\lambda }^{\uparrow }$$

(6)

where $L_{\lambda }^s$ (W.m^−2.sr⁻¹.µm⁻¹) is the at-sensor registered radiance of the TIR band, $B_{\lambda }$ (W.m^−2.sr⁻¹.µm⁻¹) is the blackbody radiance, $\epsilon$ is the emissivity of the target, $L_{atm\lambda }^{\downarrow }$ is the downwelling radiance, and $L_{atm\lambda }^{\uparrow }$ is the upwelling radiance. The blackbody radiance ($B_{\lambda }$) at the land surface temperature ($T_s$) was obtained by inverting the RTE [27]:

$$B_{\lambda }\left(T_s\right)={\displaystyle \frac{L_{\lambda }^s-L_{atm\lambda }^{\uparrow }-\tau \left(1-\epsilon \right)L_{atm\lambda }^{\downarrow }}{\tau \epsilon }}$$

(7)

The emissivity ($\epsilon$) was calculated using the Sobrino model from the satellite data [53,54]:

$$\epsilon =\left\{\begin{array}{c}{a}_i{\rho }_R+b_i; \mathrm{NDVI}<0.2\hfill \\ {\epsilon }_v+{\epsilon }_s\left(1-{\rho }_v\right)+d_{ϵ}, where d_{ϵ}=\left(1-{\epsilon }_s\right)\left(1-{\rho }_v\right)F_{ev}; 0.2\le \mathrm{NDVI}\le 0.5\hfill \\ {\epsilon }_v+d_{ϵ}; \mathrm{NDVI}>0.5\hfill \end{array}\right.$$

(8)

where $a_i$ and $b_i$ are empirical relationships between red-band reflectance and the MODIS emissivity library, $b_i$ is the reflectance of the red band, ${\epsilon }_v$ and ${\epsilon }_s$ are the emissivity of vegetation and soil, respectively, and ${\rho }_v$ is the fractional vegetation cover calculated using the following equation [55]:

$${\rho }_v={\left[{\displaystyle \frac{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}-{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}{{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}\right]}^2$$

(9)

$T_s$ is then calculated by inverting Plank’s formula [27]:

$$T_s={\displaystyle \frac{K_2}{ln\left(\frac{K_1}{\frac{L_{\lambda }^s-L_{atm\lambda }^{\uparrow }-\tau \left(1-\epsilon \right)L_{atm\lambda }^{\downarrow }}{\tau \epsilon }}\right)+1}}$$

(10)

where $K_1$ and $K_2$ are calibration constants provided in the metadata file of the satellite image. The thermal infrared (TIR) 1 band was use for the retrieval of LST since USGS has cautioned the use of the TIR 2 band due to calibration uncertainties [53].

3.5. Identification of Coal Fires

The threshold value for distinguishing fire-affected and non-fire-affected pixels was determined based on the statistical method developed by Biswal and Gorai [56]. The seasonal variations, climatic conditions, and the intensity and depth of coal fires can affect the LST of surrounding areas. For the winter season, the threshold was determined using the mean temperature ($m_{LST}$) of the map along with its standard deviation (${\sigma }_{LST}$). Pixels with values beyond the calculated threshold were categorized as fire-affected pixels. This method enables better identification of areas impacted by coal fires, considering the influence of environmental factors affecting the LST. Field verification was conducted using a Sonel KT-200 Thermal Imager (Make: Sonel S.A., Świdnica, Poland) to verify the presence of surface and subsurface coal-fire zones demarcated by the threshold, calculated by the following equation:

$${\gamma }_{LST}=m_{LST}+{3\sigma }_{LST}$$

(11)

3.6. Geographically Weighted Regression (GWR)

GWR was utilized to investigate spatial variations and the relationship between the LST and various affecting variables. GWR is a multivariate linear regression model that can incorporate many variables to predict the dependent variable [57]. Unlike traditional global regression models, GWR fits local regression models for each data point, using weights based on proximity or distance to neighboring points. This regional exploratory analysis technique measures a set of local parameters provided to the model to map and analyze the effects of the local and global correlation of variables.

The general equation for GWR is expressed as follows [57]:

$$y_{i}u={\beta }_{0i}u+{\beta }_{1i}u{x}_{1i}+{\beta }_{2i}u{x}_{2i}+\dots +{\beta }_{mi}u{x}_{mi}$$

(12)

where $y$ is the dependent variable, $x$ is the independent variable, and $\beta$ is the varying weight on the given point $i$, with a number of independent variables ranging from $0$ to $m$. The estimation of $\beta$ is based on the unbiased estimation of assets of observations. The value of $\beta$ differs at different locations based on the values of independent variables [32]. The Geographically Weighted Regression (GWR) model employs two types of local weighting schemes: Bisquare and Gaussian. The Bisquare weighting scheme assumes that features beyond the assigned distance have a weight of zero, meaning they do not influence the local weight calculation for a given point. In contrast, the Gaussian weighting scheme assigns weights to all features, with weights decreasing exponentially as the distance increases. The assignment of weights can follow two neighborhood types: the number of neighbors or the distance band. In the number of neighbors approach, a fixed number of neighboring features are selected for each feature, with the spatial extent of the band determined by the density of these neighbors. In the distance band approach, a fixed distance is applied to each feature based on the spatial distribution of features across the study area. The selection of neighborhood, whether based on a number of neighbors or distance band, can be carried using either a golden search algorithm, manual intervals or user-defined interval. The golden search algorithm determines the optimal distance or number of neighbors based on the golden section search method, while the manual interval and user-defined methods are both specified by the user. The neighborhood type with the lowest Akaike information criterion (AICc) is best-suited for running the GWR model [58].

The area was populated with 20,000 random points generated using a stratified sampling approach based on LULC classification. GWR analysis was then conducted for the winter season, taking LST as the dependent variable and NDVI, MNDWI, NDMI, and NDBI as independent variables. Assuming that the effect of the variables does not become zero after a fixed distance, a Gaussian kernel function with a distance band was applied, and the bandwidth was determined using the golden search algorithm in ArcGIS Pro since the points were generated at random [32]. The GWR analysis gives an output in the form of standardized residuals by using Equation (12) to predict the value of the dependent variable. The difference in the original and predicted outputs is used to compute the standardized residuals. The standardized residuals have a mean value of 0 and standard deviation of 1. These standardized residuals are divided into 7 different categories based on an interval size of 1 standard deviation. The seven distinct classes were categorized as −C, −B, −A, A, +A, +B, and +C based on the range of standardized residuals given in

Table 2. Categorization based on the range of standardized residuals.

Range of Standardized Residuals	Category Name
<−2.5	−C
−2.5 to −1.5	−B
−1.5 to −0.5	−A
−0.5 to 0.5	A
0.5 to 1.5	+A
1.5 to 2.5	+B
>2.5	+C

. An interval size of −0.5 to 0.5 (A) represents the regions with strongest local associations and a good model fit. Interval sizes of −1.5 to −0.5 (−A) and −2.5 to −1.5 (−B) represent the regions where the predicted variables are greater than the observed variables. Interval sizes of 0.5 to 1.5 (+A) and 1.5 to 2.5 (+B) represent the regions where the predicted variables are smaller than the observed variables. Areas with standardized residuals greater than 2.5 or less than −2.5 are identified as regions exhibiting anomalies. If the extreme standardized residuals are spread across the map randomly, then the output is completely random, and there is no association between the dependent and explanatory variables, but the presence of clustering in these high- or low-residual areas suggests that these anomalies cannot be explained by the explanatory variables included in the model. Global Moran’s I of the standardized residual was used to test the randomness and null hypothesis [32]. To better understand these anomalies, the LST values in the anomalous regions were compared with features such as LULC, coal fire areas, and other noted attributes within and around the affected clusters. Since categorical data like these cannot be directly incorporated into the GWR model, this comparative analysis was utilized to help identify the underlying causes of these anomalies. The local R² value from the GWR analysis was compared with the standardized residual and the LST to assess the relationship among these three parameters.

4. Results and Discussions

4.1. Land Use–Land Cover Map

Using 361 sampling points, the accuracy of the study area was found to be 93.63%, with a Kappa coefficient of 0.91. The LULC map of the study is shown in Figure 3. In the urban area, the LULC distribution is 0.91% water bodies, 23.00% built-up areas, 1.19% barren land, 9.20% mining areas, 43.18% grasslands, 10.05% wetlands, and 12.47% dense vegetation. The region is located near the mining belt, with scattered grasslands and wetlands interspersed among the mining and built-up areas. The rural area LULC distribution is 0.54% water bodies, 2.4% built-up areas, 2.37% barren land, 0.80% mining areas, 57.26% grasslands, 10.94% wetlands, and 26.41% dense vegetation. The built-up areas are sparsely distributed, with the northern section dominated by a significant patch of dense vegetation. Wetlands are evenly spread across the grassland areas, enhancing the ecological diversity of the rural landscape. This comparison underscores the differences in LULC between urban and rural areas, specifically the prevalence of built-up environments and mining activities in urban regions compared to the dominance of grasslands and dense vegetation in rural regions. These distinctions can offer insights into the spatial dynamics of land use and their potential impact on the LST.

4.2. NDVI Data

The NDVI indicates vegetation quantity and health, with healthy vegetation reflecting a more significant amount of near-infrared (NIR) radiation compared to unhealthy or sparse vegetation. The NDVI values range from 0.84 to −0.31 (Figure 4a,

Table 3. NDVI, MNDWI, NDMI, and NDBI values for the rural study area.

Rural LULC	NDVI				MNDWI				NDMI				NDBI
	Min	Max	Mean	Std	Min	Max	Mean	Std	Min	Max	Mean	Std	Min	Max	Mean	Std
Waterbody	−0.31	0.56	0.17	0.17	−0.34	0.67	0.13	0.22	0.01	0.48	0.25	0.09	−0.64	−0.20	−0.44	0.10
Built-up	0.23	0.54	0.38	0.07	−0.57	−0.36	−0.45	0.08	−0.14	0.11	−0.01	0.05	−0.31	0.07	−0.13	0.08
Barren	0.24	0.5	0.36	0.06	−0.60	−0.47	−0.53	0.04	−0.15	−0.02	−0.09	0.03	−0.18	0.05	−0.06	0.06
Mining	0.20	0.42	0.29	0.09	−0.47	−0.13	−0.28	0.13	−0.09	0.17	0.05	0.08	−0.41	−0.07	−0.24	0.10
Grassland	0.33	0.68	0.51	0.06	−0.60	−0.43	−0.51	0.03	−0.11	0.20	0.03	0.05	−0.47	−0.02	−0.25	0.08
Wetland	0.39	0.64	0.51	0.05	−0.51	−0.32	−0.42	0.05	0.03	0.25	0.14	0.04	−0.51	−0.24	−0.37	0.05
Dense Vegetation	0.55	0.84	0.73	0.06	−0.57	−0.43	−0.5	0.03	0.05	0.43	0.25	0.06	−0.72	−0.28	−0.53	0.08

and

Table 4. NDVI, MNDWI, NDMI, and NDBI values for the urban study area.

Urban LULC	NDVI				MNDWI				NDMI				NDBI
	Min	Max	Mean	Std	Min	Max	Mean	Std	Min	Max	Mean	Std	Min	Max	Mean	Std
Waterbody	−0.14	0.57	0.19	0.15	−0.33	0.56	0.06	0.19	−0.06	0.52	0.19	0.11	−0.70	−0.08	−0.35	0.15
Built-up	0.13	0.60	0.37	0.09	−0.51	−0.23	−0.37	0.05	−0.16	0.18	0.01	0.06	−0.40	0.12	−0.14	0.10
Barren	0.11	0.48	0.30	0.07	−0.59	−0.34	−0.46	0.05	−0.19	−0.01	−0.11	0.05	−0.19	0.16	0.00	0.10
Mining	0.04	0.51	0.22	0.10	−0.52	−0.15	−0.34	0.07	−0.26	0.06	−0.10	0.06	−0.26	0.31	0.02	0.11
Grassland	0.34	0.69	0.52	0.06	−0.57	−0.38	−0.48	0.04	−0.1	0.21	0.05	0.06	−0.48	−0.06	−0.27	0.08
Wetland	0.39	0.66	0.53	0.05	−0.49	−0.29	−0.4	0.04	0.04	0.26	0.15	0.04	−0.52	−0.23	−0.38	0.06
Dense Vegetation	0.51	0.79	0.65	0.06	−0.56	−0.33	−0.44	0.05	0.06	0.33	0.19	0.05	−0.64	−0.27	−0.46	0.07

). In the rural region, the mean NDVI for dense vegetation was 0.73 ± 0.06, while in the urban region, it was 0.65 ± 0.06. This indicates that vegetation health is better in rural areas than in urban areas, with a higher mean NDVI. Figure 4a further indicates that the rural region exhibits higher vegetation density than the urban region. Water bodies in both urban and rural areas show the lowest NDVI values due to a lack of vegetation. This study also highlights that the northern patch of the Tundi forest has a higher overall NDVI, suggesting relatively healthier and denser vegetation in that area. The NDVI can distinguish vegetation from other land cover classes, as shown in Figure 4a. Kawashima [59], Guha et al. [60], and Guha and Govil [61] have observed a negative relationship between the NDVI and LST in urban and suburban regions, where the degree of effect of vegetation on LST depends on the percentage area of other LULC classes to that of vegetated areas. It was confirmed in the study that a higher vegetation density with a higher percentage area in the rural region results in lower minimum and maximum LSTs by 2.82 °C and 1.24 °C, respectively.

4.3. MNDWI Data

The MNDWI assesses the distribution of water bodies within an area. The MNDWI values ranged from 0.67 to −0.60 (Figure 4b, Table 3 and Table 4). Within the study area, the mean MNDWI values for water bodies in the rural and urban regions were 0.13 ± 0.22 and 0.06 ± 0.19, indicating a higher concentration of water bodies in urban areas. Larger water bodies were primarily located within the open-cast mining areas, as shown in Figure 4b. This presence of water bodies significantly affects the LST of surrounding areas by keeping the area cooler [22]. Larger water bodies, as seen in Figure 4b in the mining areas, can have an even cooling effect around the area by absorbing the shortwave radiation from the sun to a greater extent [62] as well as absorbing heat from surrounding coal fires. The higher surface porosity in open-cast mines allows mine sumps to significantly contribute to the cooling of surrounding areas. The presence of water within these sumps enhances evaporative cooling, which reduces the overall land surface temperature (LST) in the surrounding regions, particularly during the daytime. [63]. The negative values of the MNDWI in other land classes reflects the absence of significant water bodies.

4.4. NDMI Data

The values of the NDMI ranged from 0.52 to −0.26 (Figure 5a, Table 3 and Table 4). This index highlights water bodies, wetlands, and dense vegetation or any areas with higher surface moisture content. The average moisture content in rural areas, particularly in the water body and dense vegetation classes, is significantly higher compared to water bodies and dense vegetation classes in urban areas, which can lead to better cooling effects in rural regions, resulting in lower LST. In contrast, barren regions in both urban and rural areas exhibit notably lower moisture content, which can directly influence the LST in these regions. Land cover classes with higher moisture content should demonstrate significantly lower LSTs due to the cooling effect of evapotranspiration [21]. This underscores the importance of vegetation and water bodies in moderating the LST.

4.5. NDBI Data

The values of the NDBI ranged from 0.31 to −0.62 (Figure 5b, Table 3 and Table 4). The mining and barren classes exhibited the highest average values in the urban region, with values of 0.00 ± 0.10 and 0.02 ± 0.11. In the rural region, the barren class had slightly lower NDBI values, with a mean of −0.06 ± 0.06. The LST is positively associated with the NDBI but can be influenced by factors such as rock surfaces, dry or wet soil, and heterogeneous man-made materials [33]. Notably, the NDBI values calculated using the NIR and SWIR2 bands provide better separation between mining, built-up, and barren areas and other land cover classes, as demonstrated in Figure 5b, compared to the conventional NDBI calculated using NIR and SWIR1 bands.

4.6. LST in Urban and Rural Areas

The average LST values for urban and rural areas are similar across most LULC classes, except for the mining class in the urban region (

Table 5. Urban and rural areas’ LST (°C) values during summer and winter seasons.

		Urban LST (°C)				Rural LST (°C)
Season	LULC	Min	Max	Mean	Std	Min	Max	Mean	Std
Summer	Water	28.03	39.32	31.74	1.87	28.77	34.66	31.8	1.05
	Built-up	28.8	40.55	32.5	1.23	27.84	35.24	32.71	0.81
	Barren	29.62	45.3	32.95	1.35	30.09	35.77	33.58	0.7
	Mining	28.52	47.68	35.37	2.84	30.24	35.06	32.91	0.82
	Grassland	28.37	39.67	32.21	0.98	25.23	35.71	32.95	0.81
	Wetland	28.42	35.85	31.63	0.88	29.35	35.22	32.71	0.76
	Dense Vegetation	27.11	36.31	30.89	1.17	21.11	35.21	30.46	1.69
Winter	Water	19.87	33.33	22.72	1.7	19.49	25.22	22.22	0.97
	Built-up	20.31	33.64	23.08	1.18	18.56	26.39	23.02	0.86
	Barren	20.83	41.88	24.05	1.51	20.63	27.04	24.4	0.87
	Mining	20.02	41.01	26.47	3.33	20.55	25.58	23.44	0.9
	Grassland	19.72	29.57	22.99	0.81	19.36	27.02	23.2	0.84
	Wetland	20.1	27.11	22.19	0.58	20.03	25.41	22.45	0.62
	Dense Vegetation	19.81	27.1	21.99	0.66	16.99	25.86	21.09	1.18

, Figure 6a,b and Figure 7). The high-temperature peak is mainly due to coal mine fires in the open-cast mining areas, a finding that was confirmed through field observations. The elevated LST values in the mining regions are also supported by the maximum LST values recorded during summer and winter. The higher LST values in the barren class within the mining regions are largely caused by the self-heating of waste overburden dumps and areas impacted by exposed underground coal fires. Studies also suggest that environmental factors like relative humidity, oxygen concentration, moisture, and sun radiation can trigger the spontaneous combustion of soil heaps, waste overburden dumps, and even coal seams [64]. The standard deviation of the LST in the mining class of the urban region was the highest, with values of ±2.84 °C and ±3.33 °C in the summer and winter seasons, respectively. This variability arises from the presence of both burning and non-burning areas around the coal mines. While the coal fire-affected areas have significantly contributed to higher LST, studies also show that the influence of oxygen on coal at low temperatures is exothermic in nature and can also increase the temperature in mining areas [10,65], resulting in the self-heating of coal. It is also important to note that the barren, grassland, and wetland classes in the rural region had slightly higher mean LSTs in both the summer and winter seasons (Table 5, Figure 7a). The higher mean LST of grassland and wetland can be traced back to the lower mean values of the NDVI and NDMI in rural regions than in urban regions (Table 3 and Table 4). Both classes have relatively lower vegetation and moisture content in rural regions spread across a larger area, resulting in a higher LST. Zheng et al. [65] and Taripanah and Ranjbar [66] in their studies also stated a negative correlation of the NDMI with the LST, relating lower vegetation to lower moisture content and therefore a higher LST. The waterbody class of the urban region followed with standard deviations of ±1.87 °C and ±1.7 °C for the summer and winter seasons, respectively. The maximum LST values of the water body class in urban areas in both the summer and winter seasons (Table 5, Figure 7b,c) are close to those of the built-up class, which can be attributed to the proximity of coal fires to the water bodies based on the analysis presented in Section 4.6. The dense vegetation class in both urban and rural regions consistently displayed significantly cooler LSTs in both seasons compared to other land cover classes, which supports the findings of [67], who attributed the reduced LST within green areas to low heat storage capacity and heat losses through the evapotranspiration of dense vegetation areas. Furthermore, the maximum LST values of wetlands and dense vegetation in urban and rural regions showed minimal seasonal differences. However, the difference in maximum temperatures between water body and built-up areas in both urban and rural regions was notably greater, while the difference in the LST between barren and mining areas in these regions was the most pronounced. Studies suggest that the LSTs of urban areas are affected by LCZs based on the built-up and land cover types [17] and urban surfaces [68]. Other reasons for the higher maximum LST of built-up areas was the close proximity of coal mine fires. The minimum LST values across most classes in urban and rural areas were generally comparable, except for the grassland and dense vegetation classes, where the minimum LST in rural areas was lower than that in urban regions.

4.7. Identification of Coal Fire Areas

The threshold value for identifying areas affected by coal fires using Equation (11) was 27.33 °C, and this covered a 691.11 ha area within the study area (Figure 8a). The LST values for fire-affected areas ranged from 27.33 °C to 41.88 °C. While most of the affected area is centered around mining regions, some sections of nearby urban areas, water bodies within mines, barren land, and grasslands also display elevated LST values, likely due to underlying coal fires or the close proximity of mine fires. Numerous accidents and substantial loss of life due to coal fires and subsidence have also been reported in these regions [44,45]. Field verifications carried out using a thermal camera confirmed the presence of underlying coal fires in the vicinity of the coal mine region in Figure 8b. The three sample images used for field verification (Figure 8b) are from (i) 23°45′27.0” N, 86°24′00.8″ E; (ii) 23°47′43.1″ N, 86°20′36.0″ E; and (iii) 23°46′55.2″ N, 86°23′27.3″ E. Figure 8b (i) is of a coal seam on fire with a nearby water sump inside an open-cast mine at a temperature of 38.2 °C. Figure 8b (ii) is a burning overburden dump situated near a state highway with a mine fire on both sides of the highway. Figure 8b (iii) shows the edge of an open-cast coal mine, which was once a commutable road, and an old, abandoned building structure affected by a coal fire near the surface.

4.8. Relationship Between LST and Different Affecting Variables

This study analyzed correlations between the LST and different variables within the urban and rural regions using whole pixel values from the satellite image of the study area (265299 points each for urban and rural regions) (Figure 9). The results showed an inverse relationship between the LST and the NDVI (Figure 9a,b), with a stronger correlation in the rural region (R² = 0.50) compared to the urban region (R² = 0.24). This suggests that the influence of the NDVI on the LST is more pronounced in rural areas than in urban areas. According to Ge et al. [69], the higher the NDVI, the higher the heat loss through evaporation and transpiration. Denser vegetation regions with higher area coverage result in much lower LSTs than denser vegetation with fragmented areas [70]. This fragmentation of denser vegetation could be the reason why, even with higher NDVI values in urban areas, the temperature is greater than 20 °C, while in rural areas, the LST can go as low as 17 °C. The presence of water bodies results in lower overall correlation values between the LST and NDVI, as the relationship between the NDVI of water bodies and the LST is moderately positive [33]. Since water absorbs the NIR band, resulting in negative NDVI values, but due to evaporation, it has a cooling effect, the correlation between the NDVI and the LST is more negative for green spaces and built-up areas than for water bodies.

The correlation between the LST and the MNDWI (Figure 9c,d) in both urban and rural regions was very weak, with R² values of 0 and 0.03, respectively, indicating no relationship between them in the mining region. However, the MNDWI can still distinguish water bodies from other land cover types, which can be helpful to study the local effect. In other studies, the MNDWI has had a negative correlation with LST, as water bodies have distinct positive values and other classes have negative values [71,72]. According to Peng et al. [73], while other indices, such as the Normalized Difference Water Index (NDWI), used to distinguish water are always integrated with construction land features, the MNDWI can better characterize open water bodies, though they also found a very low correlation between the MNDWI and the LST with a 0.001 R². The reason for this lower correlation was that the area of open water bodies was very small compared to other land classes, which can also be seen in this study, where water bodies account for only 0.54% of the study area.

A strong negative correlation was found between the LST and the NDMI (Figure 9e,f) in the rural region (R² = 0.60) compared to urban regions (R² = 0.28). This suggests that moisture content plays a significant role in influencing LST in rural areas, while the impact is somewhat diminished in urban regions due to other anthropogenic factors, such as mining and built-up areas. Studies suggest that the NDMI and LST have a negative correlation [65,66]. Soil moisture can depend on the amount of rainfall, slope, soil type, and vegetation, as well as the presence of water bodies and wetlands, which can induce a cooling effect [66]. Higher vegetation also contributes to evapotranspiration, which can inversely affect the LST. The higher negative correlation between the NDMI and the LST is a result of the combined cooling effects from water bodies, grasslands, wetlands, and dense vegetation through evapotranspiration. The lower moisture levels present in built-up and mining areas lead to the slow transpiration of land surfaces, which promotes a rise in the LST [1,74], as seen in the urban region of the study area, though due to coal fires, the LST rises to a much greater extent, thus reducing the overall correlation between the NDMI and the LST in the urban region.

The LST and NDBI exhibited a stronger positive correlation (Figure 9g,h) in rural areas (R² = 0.61) than in urban areas (R² = 0.36). In rural areas, higher NDBI values correspond to a higher LST. However, in urban areas affected by coal mine fires, even lower NDBI values, such as for water bodies, can exhibit significantly elevated LSTs. This anomaly in the urban region can influence the correlation between the LST and the NDBI. Since the NDBI reflects barren information, which includes barren land, mining areas, and built-up regions, it correlates positively with the LST. The higher the density of these features, the higher the correlation between the LST and the NDBI will be [74]. This study also found a similar correlation with higher LSTs in regions with a higher NDBI.

Overall, the NDMI and NDBI are more effective for simple linear correlation analysis than the NDVI in mining regions, as the NDVI primarily accounts for vegetation only. At the same time, the NDMI encompasses vegetation, wetlands, and water bodies, offering a broader perspective on moisture content that directly impacts the LST due to evapotranspiration. Due to a low overall correlation, this linear relationship fails to address the local parameters affecting the LST and lower R² values. However, it can give an overall general idea of the relationship between the LST and these variables.

4.9. GWR Analysis

The coefficient of determination (R²) for the GWR analyses for the NDVI, MNDWI, NDMI, and NDBI variables was 0.7831. Thus, these variables can account for 78.31% of the variability in the LST. This is more significant than the individual relationships of the variables with the LST, as GWR uses the combined effect of these variables to explain the variations in the study area [33]. The GWR analysis divides the research area into seven groups based on standardized residuals. Category A of the standardized residual is the region with the best GWR model fit, where all four indices explain the LST trend with the least error. Categories −A and +A follow the 1 std deviation range away from A, and −B and +B follow the 1 std deviation away from −A and +A, respectively. The areas with the weakest local associations with the independent variables are indicated by standardized residual values greater than 2.5 (+C) or less than −2.5 (−C) (Figure 10a). The Global Moran’s Index and the z-score for the standardized residual are 0.42 and 112.86, respectively (Figure 10c), which suggests high clustering in the residuals and rejection of the null hypothesis. This result is important as it supports the alternate hypothesis with a strong indication of the external local factors affecting the LST. Upon examination, it has been observed that while categories A, +A, +B, −A, and −B are spread across the study area randomly, categories +C and −C are the only categories present in urban mining areas in higher percentages. While category A has the least standardized residuals, there are various factors that can affect the LST in categories −A and −B, resulting in the observed LST being lower than the predicted LST. Features such as the density of vegetation and wetland, leaf area index, humidity, type of soil, wind speed, elevation, and aspect (shadow region) are some of the environmental factors that can affect the LST [17,32,75]. There are other anthropogenic factors, such as population density, distance from roads, distance from mines, building height, building coverage ratio, sky view factor, and distance from mine fire, that can also affect the LST [30,32,75]. These factors can also affect categories +A and +B, resulting in a lower predicted LST than the observed LST. In the rural region, the shadow zones of the Tundi forest are the areas with the most data points with category −C, while data points with category +C are mostly found in barren areas. The dense vegetation class in the rural region and the mining class in the urban region exhibit higher local R² values, demonstrating a higher correlation among the affecting variables (Figure 10b). These values are significantly lower in the grassland region, indicating that areas with mixed features have weaker correlations than those with a single predominant feature. In the urban area, the percentage of anomalies is higher, particularly within the mining zones, while there are no significant clusters of category +C or −C in the rural region. Urban mining areas show a higher concentration of these anomalies, with both the +C and −C categories in higher clusters. While some studies consider residuals larger than +3 or lower than −3 as outliers and advise that they be examined [58], in this study, based on a range of one standard deviation, residuals with a value lower than −2.5 and greater than +2.5 are considered outliers. Since both categories −C and +C are clustered in a specific land class in the urban region, there is a greater influence of mining activity in these two categories. The LST map shows a significant difference between the minimum and maximum urban LST values in mining and barren regions, with differences exceeding 20 °C, as shown in Figure 7c. The areas with category +C are predominantly those affected by coal fires (Figure 8a and Figure 10a), leading to LST values that are considerably higher than those in the surrounding regions. Studies show that the self-heating of coal due to oxidization, spontaneous burning, and the absorption of NIR-SWIR bands by the C–H, O–H, and N–H functional groups in coal can result in higher temperatures in the mining area [10,76,77]. The surface moisture of mining was very low, resulting in the least to no effect of evaporation on the LST. In contrast, the areas with category −C found within coal mine pits show that these areas are not affected by coal fires, with the accumulation of water in the sumps, increasing the surface moisture (Table 4). This resulted in the LST being significantly lower due to evaporation [33,62,63]. The scatter plot depicting standardized residuals versus local R² for urban regions (Figure 10d) also validates that clusters of extreme residuals are present in higher numbers in urban areas. The category +C datasets correlate with high LSTs (Figure 10f), while category −C datasets are related to lower LSTs. Categories A, −A, −B, and +B in the rural region are randomly distributed across the area, with clusters of categories −C corresponding to the lower LST-C category, primarily located in the northern part of the rural region (Figure 10a,e,g). It was also observed that the foothills of the Tundi forest region in the northern part of the rural region have higher clusters of category +A. Clusters of the +A category can be seen near the foothills of Tundi forest. These increased temperatures in close vicinity to foothills could be due to the direction of the slope towards the southeast [78]. From the GWR analysis, the high and low standardized residuals in the urban region can be interpreted in relation to mining operations, coal fire-affected areas, and the self-heating of coal, which raise the LST values, as well as the presence of accumulated water in sumps, which considerably reduces the LST in coal mine areas due to evaporation.

4.10. Limitations of the Study and Future Plans

The variations in the LST of coal mining areas can be explained effectively using GWR modeling with the help of affecting variables, such as mining and coal fires, along with indices such as the NDVI, MNDWI, NDMI, and NDBI. However, there are other parameters that can also affect the LST which are not explored in this research. The lack of explanation for the lower local R² values in the grassland region of the study area requires further analysis of additional parameters affecting the LST in these LULC classes. The areas with categories −B, −A, +A, and +B will only be of significance if there is clustering in the residual; otherwise, it can be difficult to explain the impact of underlying independent variables using the GWR model. The GWR model predicts the dependent variables based on the statistics of the input independent variables, which can be continuous or categorical but should not have multicollinearity. Another shortcoming of the GWR model is its kernel size. The moving kernel assumes that the response to the predictor relationship operates over the same spatial scale across the study area, which is not true in the real world and affects the accuracy of the output data. Future analysis can focus on how much each affecting variable is contributing to the LST in the mining region to assess the local effects in different LULC classes along with the effect on urban heat islands in these regions using GWR as well as other recent advanced methods.

5. Conclusions

This study uses Landsat datasets to analyze the spatial relationships for urban and rural values of the NDVI, MNDWI, NDMI, NDBI, and LST across various land classes in a coal mining region impacted by coal mine fires. Some summarized results are as follows: (1) Different land classes exhibit varying LST values depending on their location. The same land class in urban and rural areas can display different spectral properties and LSTs due to the surrounding land uses. (2) Areas near mining activities, particularly those affected by coal fires, show LST values that are significantly higher than average. The correlation between the LST, NDVI, NDMI, and NDBI is weaker in urban regions than in rural regions. Furthermore, no significant correlation was found between the LST and the MNDWI. (3) The NDMI, being highly correlated with the LST, can help mitigate higher temperatures near coal mining areas by focusing on the land cover classes that are directly related to the LST, such as water bodies, wetlands, and dense vegetation. (4) The NDVI and MNDWI indicators are complementary indices used to study the relationship of vegetation and water with LST. The NDVI only focuses on vegetation health, and the MNDWI focuses on surface water. (5) The GWR analysis suggests that, in addition to the NDVI, MNDWI, NDMI, and NDBI, other local variables can also influence the LST in the study area, which can be attributed to mining activities, the occurrence of coal fires, accumulated water in the mine sump, and the land surface. (6) The urban region with mining activities shows many more anomalies than the rural region.

This study offers valuable insights into the relationship between the LST and various indices in both urban and rural areas in coal mining regions. Additionally, it enhances our understanding of urban heat islands near these coal mining sites, primarily driven by coal mine fires. The indices applied in this research can aid in identifying and delineating hot and cool zones around mining and urban areas, providing essential information for urban planning aimed at mitigating high LSTs. Reducing coal fires in barren and mining areas is necessary to lower the LST. The GWR analysis results highlight the substantial impact of open-cast mining and coal fires in Dhanbad City, especially when compared to rural locations. This emphasizes the need for targeted strategies to manage temperature variations in mining-affected areas.