Assessing variations in land cover-land use and surface temperature dynamics for Dehradun, India, using multi-time and multi-sensor landsat data

Abstract

Rapid urbanisation and industrialisation coupled with overpopulation have altered land cover/land use (LCLU) and surface temperature (ST) patterns in Dehradun. Monitoring these changes through satellite-based remote sensing is required to ensure the sustained development of this ecologically fragile region. Here, LU and ST dynamics of the Dehradun municipal area have been estimated using Landsat-5 datasets for 1991, 1998, and 2008 and Landsat-8 dataset for 2018. LU maps have been extracted using a Gaussian Maximum Likelihood classifier with an overall accuracy of over 88% and Kappa coefficients above 0.85. Results reveal that the urban region expanded by 80.6% in the 27 years while dense vegetation and dry river bed classes have declined sharply. Sparse vegetation has risen by 3 km², whereas bare ground has decreased by about 4.3 km². Mean ST has increased above 9 °C from 1991 to 2018 in every season. A seasonal influence is evident on the mean ST per LU class’s trend, which rose between 8 °C and 12 °C for every LU class, indicating significant warming across each LU class. ST probably has non-linear relationships with its causal factors represented by spectral indices, elevation, and population density. Urban heat island (UHI) formation is thus evinced, promulgating the administration’s urgent action to save the environment and redrawing policies for ambitious projects like smart cities.

Introduction

Urbanisation has peaked in the world today, with more people settling in towns than ever. Much of this urban development is happening in third-world countries, and the global urban population is expected to rise to 5 billion by 2030 (United Nations Population Fund, 2016). Consequently, natural land cover (LC), like vegetated and agricultural lands and water bodies, has been transformed into impervious surfaces (Ding & Shi, 2013), which cover about 40% of the land on the planet (Sterling & Ducharne, 2008). The surface energy and radiation budgets have also been altered due to an increase in the absorption of shortwave radiation and a reduction in energy loss because of the emission of longwave radiation (Oke, 1976). This alteration profoundly influences ecosystem functioning and climatic conditions, evident in the land, water, and air pollution and enhanced urban heat islands (UHIs) (Chan & Yao, 2008; Shao et al., 2006). Urban areas experiencing UHI are significantly warmer than the surrounding peri-urban or rural areas owing to a significant presence of imperviousness as LC. According to the United States Environmental Protection Agency (2008), increased energy use, increased greenhouse gas releases, and reduced levels of well-being are the major effects of UHI.

Conventionally, UHI studies were performed at isolated locations, and the difference between in-situ air temperature measurements was determined. With the arrival of thermal sensors, it has become possible to study this phenomenon continuously, locally and globally. Extraction of land surface temperature (LST), defined as the skin temperature of the bare Earth surface or the temperature at the interface between the Earth’s surface and the atmosphere, forms a central theme in such research works (Gallo & Owen, 1998; Kidder & Wu, 1987; Liu & Zhang, 2011; Ramachandra et al., 2012; Streutker, 2002; Yuan & Bauer, 2007). Initially performed using the National Oceanic and Atmospheric Administration Advanced Very High-Resolution Radiometer (NOAA AVHRR) data at a spatial resolution of 1.1 km, urban temperature mapping and UHI detection have improved considerably from regional to the city level with the availability of 120 m and 60 m spatial resolution Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper + (ETM +) products. LST relates more closely to the urban canopy layer (UCL) heat islands (Weng et al., 2004). However, the canopy layer may have a discontinuity in an urban area, necessitating the requirement of a high resolution of satellite imagery for accurate UHI representation (Nichol, 1994).

Further research on LST reveals that surface radiant temperature response expressed as a function of latent and sensible heat fluxes varies according to topsoil, water content, and vegetation cover (Owen et al., 1998). In contrast to the impervious surface, green spaces prevent direct surface heating from solar radiation, which decreases LST and helps to generate cool island effects (Li et al., 2012). The urban fraction exhibited the maximum correlation with LST in the National Capital Territory of Delhi (Mallick et al., 2013). It was concluded that the dense built-up at Lucknow’s centre recorded the largest LST in contrast with the open spaces in the immediate vicinity, which were characterised by vegetation and water bodies (Shukla & Jain, 2021; Singh et al., 2017; Verma & Kundapura, 2020). Normalised difference vegetation index (NDVI) (Rouse et al., 1974) is the most commonly used quantitative representation of vegetation density and health. NDVI’s relationship with LST is majorly negative, as Imhoff et al. (2010) and Molnar (2016) reported.

Similar to NDVI is the normalised difference water index (NDWI) (McFeeters, 1996) which monitors the changes in water bodies using green and near-infrared (NIR) channels. Spectral indices have also been developed to monitor built-up areas (Zha et al., 2003) and barren lands (Zhao & Chen, 2005). Representing the LC types by these indices and interpreting their relationship with LST helps understand the UHI intensity over a temporal scale, as Chen et al. (2006) performed for the Pearl River delta and Nimish et al. (2020) for the Kolkata metropolitan area. Quantifying the land use (LU) changes in urban areas using supervised classification and relating them with LST also helps understand the impact of anthropogenic activity on the UHI intensity. Hence, recent research combines these analyses, i.e. LCLU changes, while investigating the UHI effect in a particular study area (Choudhury et al., 2019; Nimish et al., 2020; Ramachandra et al., 2012). Population growth, measured through population density, i.e. the number of persons per unit area, is also related to LST to understand the local population’s adaptability to the existing climatic conditions and study the modifications in the climatic patterns brought in by increasing population and their economic activities (Dushi & Berila, 2022; Mallick & Rahman, 2012).

Dehradun has remained a thoroughly investigated study area since it became the capital city of Uttarakhand in 2000 (Singh et al., 2013). Recently published works mostly cater to analysing the LCLU changes in the Doon valley between 2000 and 2009 (Tiwari & Khanduri, 2011). Jana et al. (2020) used two Landsat images from 2000 and 2019 to investigate the impact of spatio-temporal urban expansion on green spaces and thermal behaviour in the Doon valley. Piyoosh and Ghosh (2018) developed new indices: modified normalised difference soil index (MNDSI) and normalised ratio urban index (NRUI). MNDSI is the output of the division of the difference between the digital number (DN) of the shortwave infrared—2 (SWIR—2) and panchromatic (PAN) bands and the sum of the DN of SWIR—2 and PAN bands. NRUI is the result of the difference between the ratio urban index (RUI) and MNDSI divided by the sum of RUI and MNDSI, wherein RUI is the quotient of the biophysical composition index (BCI) (Deng & Wu, 2012) and MNDSI. They showed the efficacy of MNDSI and NRUI over other spectral indices in distinguishing soil and urban areas, respectively, while performing semi-automatic extraction of areas of anthropogenic activity in the Landsat-8 image of the Dehradun planning area (Piyoosh & Ghosh, 2017). Saini and Tiwari (2017) studied the spatio-temporal variation of LST in Dehradun city and its surrounding areas from 1998 to 2017 and assessed LST’s relationship with NDVI during the same period. LCLU maps were extracted using new and existing spectral indices (Piyoosh & Ghosh, 2020) and artificial neural network (ANN) (Nautiyal et al., 2021) for the Dehradun planning area to study LCLU changes and correlate with LST. Nautiyal et al. (2021) also analysed LST hotspots and studied the variation in human thermal comfort levels. Maithani et al. (2020) studied the effect of the national lockdown due to the Covid-19 pandemic on the spatio-temporal patterns of LST in the Dehradun municipal area by taking Landsat-8 images of April–May 2017, 2018, 2019, and 2020.

The present research considers only the Dehradun municipal area compared to the larger study area definitions adopted in the above-cited works. An attempt is made to extract LU and spectral index maps using four October–November images of 1991, 1998, 2008, and 2018 and study the changes in LU patterns during these 27 years. LST maps for 1991, 1998, 2008, and 2018 are extracted from 31 Landsat-5 and 8 images to understand LST's behaviour season-wise and for each LU class during the considered time frame. The cross-validation approach has also been introduced to examine Landsat-derived LST’s reliability in the study area instead of field measurements using the classic Moderate Resolution Imaging Spectroradiometer (MODIS) MOD11A1 version 6 (Wan et al., 2015) LST product and the newly launched MOD21A1D version 6.1 (Hulley & Hook, 2017) LST product. MOD21A1D LST product has not been utilised in any research, as per the authors’ knowledge.

Moreover, qualitative changes in LC have also been studied using spectral indices, and their relationships have been analysed with LST. Another novelty of the present work is incorporating elevation as a causal factor (Mathew et al., 2016) and assessing its influence on the LST and spectral indices during these 27 years. Lastly, population density maps derived from gridded population datasets: WorldPop (Lloyd et al., 2019) and Global Human Settlement Layer (GHSL; Freire et al., 2016; Schiavina et al., 2022) instead of the traditional ward-wise population density maps have been used to understand the effect of population growth on LST using correlation analysis.

Study area

Dehradun, located in the Uttarakhand state of India, is the investigation region (Fig. 1). The Census of India website (https://www.censusindia.gov.in/2011census/dchb/DCHB.html) states that the city had a population of 578,420 in 2011, and 714,223 citizens resided in the larger urban agglomeration area. Nestled in an undulating valley between the Shivalik hills and the lesser Himalaya range, the latitudinal and longitudinal extents are 30.266362 N to 30.403137 N and 77.985389 E to 78.102823 E, respectively. With a humid subtropical climate, the average yearly highest and lowest temperatures are 28 °C and 15 °C, respectively (Indian Meteorological Department, 2015). The region experiences an average annual rainfall of above 2000 mm. Once famed for its litchi orchards, tea gardens, basmati rice, and canal system designed during the British Raj, Dehradun witnessed rapid urbanisation since its designation as Uttarakhand’s interim capital in 2000. Dehradun’s civic agency was upgraded to a municipal corporation in 1998 by the then Uttar Pradesh government (Kazmi, 2013). Abolished by the Uttarakhand government in 2000 and reconstituted in 2003, the municipal limits have expanded more than five times to accommodate 100 wards in 2018 from 34 in 1991 (Pant, 2017; Rajeshwari, 2006). The district lost about 213.03 km² of forest area to large-scale construction activities and industries in the first 2 decades of the twenty-first century (Forest Survey of India, 2019).

Fig. 1.

Location of the study area

Materials

The datasets utilised for this research work belong to three sensors of the Landsat series of satellites, viz. TM of Landsat-5 for 1991, 1998, and 2008; and Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS) of Landsat-8 for the year 2018. TM has six bands (bands 1–5 and band 7) in the wavelength range of 0.45–2. 35 µm and one band (band 6) in the wavelength range of 10.41–12.5 µm (https://www.landsat.gsfc.nasa.gov/landsat-4-5/tm). OLI has eight bands (bands 1–7 and band 9) in the wavelength range of 0.433–2.3 µm and a panchromatic band (band 8) spanning the 0.5–0.68 µm region of the electromagnetic spectrum (https://www.landsat.gsfc.nasa.gov/landsat-8/oli-requirements). TIRS has two bands in the wavelength range of 10.6–12.5 µm (Irons et al., 2012). TM data products have a radiometric resolution of 8-bit, while OLI and TIRS data products have a 12-bit radiometric resolution scaled to 16-bit integers. Both Landsat-5 and Landsat-8 have a temporal resolution of 16 days. Table 1 presents the details of the scenes: 21 from TM and ten from OLI and TIRS. Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model Version 3 (GDEM 003) possesses improved horizontal and vertical accuracy compared to the previous GDEM versions and is used for obtaining the study area’s elevation values.

Table 1.

Description of multi-sensor multi-temporal satellite data

Satellite	Sensor	Path	Row	Date of acquisition	Spatial resolution	Projection parameters	Application
Landsat—5	TM	146	039	1991 (January 9, March 30, May 17, November 10)	30 m (Bands 1–5 and 7); 120 m resampled to 30 m (Band 6)	Projection System: Universal Transverse Mercator (UTM) Zone: 44 North (N) Datum: WGS 84	LCLU classification and calculation of spectral indices from images corresponding to November 10, 1991, November 12, 1998, October 22, 2008, and October 2, 2018; derivation of LST from all scenes of TM and OLI and TIRS
				1998 (March 1, April 2, April 18, May 4, May 20, October 11, October 27, November 12, December 14, December 30)
				2008 (April 13, April 29, May 31, October 22, November 23, December 9, December 25)
Landsat—8	OLI and TIRS			2018 (January 19, February 4, February 20, March 8, March 24, April 25, May 11, June 12, October 2, December 21)	30 m (Bands 1 – 7 and 9); 15 m (Band 8); 100 m resampled to 30 m (Bands 10 and 11)

The Nagar Nigam Dehradun website (https://www.nagarnigamdehradun.com/) provides the individual ward boundaries in pdf. Georeferencing the municipal ward map uses the scanned and co-registered Survey of India toposheets (sheet numbers: 53F/15, 53 J/3). The Census of India website (https://www.censusindia.gov.in/) supplies the demographics and socio-economic profile through the district census handbook for the census years 1991, 2001, and 2011. The Google Earth portal gives high-resolution reference data for validating the classification results.

The ever-changing ward boundaries due to frequent delimitation exercises render any ward-wise analysis unsuccessful, thereby promulgating the need for standardised data products to facilitate time-series socio-economic analyses in urban areas. WorldPop and GHSL initiatives provide high-resolution, open-source gridded data on various demographic indicators and spatial extent of urban areas at global and country levels. Gridded surface area and population count datasets at 100 m spatial resolution for India have been utilised to generate population density maps for 2000, 2008, and 2018 (WorldPop & Center for International Earth Science Information Network, Columbia University, 2018). A methodology for deriving grid-cell surface area proposed by Santini et al. (2010) is adopted and modified by Lloyd et al. (2019) to derive WorldPop gridded-cell surface areas, which remain unaffected by time. Population count datasets have been generated using an unconstrained top-down approach (Lloyd et al., 2019; Stevens et al., 2015). The WorldPop population counts are available from 2000 to 2020 only. Therefore, the latest release of GHSL-POP, i.e. R2022A (Schiavina et al., 2022), has been used to estimate population density for 1990 at 100 m spatial resolution. Freire et al. (2016) discuss the methodology for arriving at the gridded residential population counts in GHSL datasets. The WorldPop datasets are available in Geographic Coordinate System (GCS) World Geodetic System -1984 (WGS 84), whereas the GHSL-POP dataset has World Mollweide WGS 84 coordinate system.

Two MODIS LST products have been utilised for validating the LST derived from Landsat datasets. One is MOD11A1 version 6, in which LST is generated using the standard split-window algorithm. This product is customised from the Application for Extracting and Exploring Analysis Ready Samples (AρρEEARS) (AppEEARS Team, 2021) portal. MOD21A1D uses the physics-based temperature-emissivity separation algorithm to retrieve LST from the three thermal channels 29, 30, and 31 of MODIS. It is available for download from the Earthdata (https://earthdata.nasa.gov/) portal. Processing for MOD21A1D version 6.1 scenes from 2011 to 2018 is anticipated to be complete by the end of 2021. Hence, this product was not available for 2018. Table 2 lists the specifications of the MODIS LST products selected for validation.

Table 2.

Specification of MODIS LST product chosen for validation of Landsat-derived LST

Satellite	Sensor	LST product	Date of acquisition	Spatial resolution (m)	Projection parameters	File format
Terra	MODIS	MOD11A1	October 22, 2008	1000	Projection System: GCS Datum: WGS 84	GeoTiff
			October 2, 2018
		MOD21A1D	October 22, 2008	926.6	Projection System: Sinusoidal Datum: WGS 84	Hierarchical Data Format (hdf)

Methodology

The overall methodology for this research work involves significant steps: study area boundary preparation; data pre-processing; LST estimation and validation; generation of LU maps; calculation of spectral indices; LU change analysis; determination of season-wise LST pattern and LST statistics per LU class; generation of population density maps; and assessment of the relationships between LST and spectral indices, LST and elevation, spectral indices and elevation, and LST and population density. The following sub-sections describe these steps.

Preparation of study area boundary

Individual ward boundaries are converted from pdf to jpg format. Then these images are georeferenced using third-degree polynomial affine transformation and 56 ground control points (GCPs) of important landmarks obtained from the co-registered toposheets and cross-validated from a high-resolution Google Earth image.

The municipal ward map is prepared by digitising and correcting topological errors. Attributes from the district census handbook like ward number, ward name, the total population of the ward, ward-wise male and female population, ward-wise population density per km² and hectare, ward area in a hectare, km² and acres, ward perimeter in km, and ward-wise road length in km were added. The ward boundary map was exported in shapefile format, and the ward boundaries were dissolved to create a single boundary shapefile of the entire Dehradun municipal area.

The latest studies (Jaiswal et al., 2021; Nautiyal et al., 2021; Sai Krishna et al., 2017) consider the municipal extent and ward definitions given in the district census handbook of 2011. The extent representative of 100 wards in 2018 corresponds to more than 50% of the Dehradun planning area, already considered in multiple studies cited in the “Introduction” section. No study has attempted to assess the LCLU and LST dynamics, considering the municipal extent at the start of Dehradun’s rapid and haphazard urbanisation, which now forms the urban core of the newly expanded city limits in 2018. Hence, the single boundary shapefile created herein represents the area limit to which the Uttarakhand government reconstituted the municipal corporation in 2003, as mentioned in the “Study Area” section.

Data pre-processing

Landsat-5 TM data

Pre-processing includes generating radiance images from the band images containing DN values, converting radiance images to the top of atmosphere (TOA) reflectance images and cropping the radiance, reflectance, and DN value band images to the study area boundary. Production of radiance images follows Eq. 1 (Chander et al., 2009):

$$L^{\ast }=\frac{(L_{\mathit{max}}-L_{\mathit{min}})DN+L_{\mathit{min}}}{{\mathit{DN}}_{\mathit{max}}}$$ 1

where $L^{\ast }$ $=$ at-sensor spectral radiance, $L_{\mathit{min}}$ $=$ lowest spectral radiance, $L_{\mathit{max}}$ $=$ highest spectral radiance, and ${\mathit{DN}}_{\mathit{max}}$ $=$ highest DN. The metadata file associated with the TM dataset gives $L_{\mathit{min}}$, $L_{\mathit{max}}$ and ${\mathit{DN}}_{\mathit{max}}$ values for each band. Radiance images were generated for bands 2, 3, 4, and 6 of the TM datasets.

Generation of TOA reflectance is possible from Eq. 2 (Chander et al., 2009).

$${\rho }_{\lambda }=\frac{\pi L^{\ast }{d}^2}{\cos {\theta }_s{\mathit{ESUN}}_{\lambda }}$$ 2

where ${\rho }_{\lambda }$ $=$ TOA reflectance, $\pi$ $=$ 3.14159, $d$ $=$ distance of Earth from the Sun in astronomical units, ${\theta }_s$ $=$ solar zenith angle, and ${\mathit{ESUN}}_{\lambda }$ $=$ average exoatmospheric solar irradiance. Readers must refer to Chander et al. (2009) for the values of $d$ and ${\mathit{ESUN}}_{\lambda }$ whereas the metadata file provides the value of solar elevation angle, which is the complement of ${\theta }_s$. The radiance images of bands 2, 3, and 4 serve as input for producing TOA reflectance images.

Radiance and reflectance images were clipped using the prepared boundary shapefile to extract the data of the study area. The unprocessed band 5 image was also cropped according to the study area boundary.

Visual interpretation of the Landsat scene dated May 17, 1991, revealed clouds and shadows over some parts of the study area. Therefore, the “Landsat Quality Assurance (QA) ArcGIS Toolbox” (Department of the Interior United States Geological Survey, 2017) was used to obtain pixel-wise information about clouds, cloud shadow, closeness to clouds, and snow. Such pixels were converted into three polygons: cloud, cloud shadow, and closeness to cloud and snow. The “erase” tool removed these areas from the study area boundary shapefile. Then, band images for this scene were clipped using the new study area shapefile, and after that, the pre-processing steps were employed.

Landsat-8 OLI and TIRS data

According to the Landsat-8 Data Users Handbook (Department of the Interior United States Geological Survey, 2019), it is possible to obtain radiance images by Eq. 3:

$$L^{\ast }=M_L\times DN+A_L$$ 3

where $M_L$ $=$ band-wise multiplication constant and $A_L$ $=$ band-wise addition constant. The metadata file supplies these values. Radiance images were generated for bands 3, 4, 5, and 10 of the OLI and TIRS dataset.

TOA reflectance generation follows Eqs. 4 and 5 (Department of the Interior United States Geological Survey, 2019).

$${\rho }_{{\lambda }^{{}^{\prime }}}=M_{\rho }\times DN+A_{\rho }$$ 4

where ${\rho }_{{\lambda }^{{}^{\prime }}}$ $=$ apparent TOA reflectance without solar angle correction, $M_{\rho }$ $=$ band-wise multiplication constant, and $A_{\rho }$ $=$ band-wise addition constant. Again, the metadata file gives $M_{\rho }$ and $A_{\rho }$.

$${\rho }_{\lambda }=\frac{{\rho }_{{\lambda }^{{}^{\prime }}}}{\sin \theta }$$ 5

where ${\rho }_{\lambda }$ $=$ actual TOA reflectance, and $\theta$ $=$ solar elevation angle obtained from the metadata.

Again, the study area was clipped from the radiance and reflectance images according to the shapefile. The unprocessed band 6 image was also cropped.

DEM

Two scenes of the ASTER GDEM 003 product were mosaicked. According to the boundary shapefile, the study area’s elevation values were cropped from the mosaicked scene.

MODIS

The study area boundary shapefile was converted from.shp format to.geoJSON format using the R programming language and uploaded on the AρρEEARS portal while submitting an area sample request for the Level 3 MOD11A1 Version 6 daily LST product in GeoTiff format. LST values in Kelvin (K) were obtained by multiplying the product by 0.02. Subtracting 273.15 K gave the LST in degree Celsius (°C) units.

The MOD21A1D Version 6.1 daily LST product for the study area was available in hdf from the Earthdata portal (https://search.earthdata.nasa.gov/). Again, LST values in K were retrieved by multiplying 0.02 and subtracting 273.15 K gave the LST in °C. The LST dataset was resampled from 926.6 to 1000 m and reprojected to the GCS WGS 84 coordinate system to facilitate spatial resolution and projection system uniformity during the validation step.

WorldPop and GHSL-POP data

The study area boundary shapefile was reprojected to the coordinate system of the WorldPop surface area and population count datasets, i.e. GCS WGS 84. The population count for the considered years and surface area corresponding to the study area were clipped according to the reprojected shapefile.

The GHSL-POP dataset was reprojected to GCS WGS 84, and the population count corresponding to the study area was clipped according to the study area shapefile.

Estimation of LST

LST has been estimated from each Landsat scene acquired on the dates specified in Table 1 for the considered years. LST retrieval from the Landsat thermal data involves estimating two inputs: at-sensor brightness temperature and surface emissivity. Band 6 and band 10 radiance images serve as input for retrieving at-sensor radiant or brightness temperature in the case of TM and TIRS datasets using Eq. 6 (Chander et al., 2009; Department of the Interior United States Geological Survey, 2019):

$$T=\frac{K_2}{ln(1+\frac{K_1}{L^{\ast }})}$$ 6

where $T$ $=$ at-sensor brightness temperature in K, $K_1$ and $K_2$ $=$ thermal conversion constants for the respective band available from the metadata file. One of TIRS’s channels, i.e. band 11, shows significant calibration uncertainty (Landsat Missions, 2019). Hence, band 10 is used instead.

Retrieving surface emissivity follows Eq. 7 (Sobrino et al., 2004), which considers pixels containing both soil and vegetation proportions.

$$\epsilon =0.004P_v+0.986$$ 7

where $\epsilon$ $=$ surface emissivity and $P_v$ $=$ vegetation fraction determined according to Eq. 8 below (Carlson & Ripley, 1997).

$$P_v={[\frac{NDVI-{\mathit{NDVI}}_{\mathit{min}}}{{\mathit{NDVI}}_{\mathit{max}}-{\mathit{NDVI}}_{\mathit{min}}}]}^2$$ 8

where ${\mathit{NDVI}}_{\mathit{max}}=0.5$ and ${\mathit{NDVI}}_{\mathit{min}}=0.2$, as Sobrino et al. (2004) recommended for global conditions. Using $\epsilon$ and $T$, it is possible to obtain LST according to the equation below (Artis & Carnahan, 1982; Stathopoulou & Cartalis, 2007; Weng et al., 2004)

$$LST=\frac{T}{1+ln\epsilon (69.50346T\lambda )}$$ 9

where $LST=$ LST in K, $\lambda =$ emitted radiance wavelength (central wavelength of bands 6 and 10, respectively) (Markham & Barker, 1985).

Estimation of LST in °C uses the following equation:

$$LST{(}^{\circ }\mathrm{C})=\mathrm{LST}-273.15$$ 10

Validation of LST estimated from Landsat data

Point-based ground measurements using thermal infrared (TIR) guns help validate LST only in a small homogenous area due to the high variations observed in LST measurements with space and time (Nikam et al., 2016). The cross-validation approach (Qian et al., 2013) overcomes this limitation. Two MODIS-derived daily daytime LST products: MOD11A1 and MOD21A1D, acquired over the study area on the date and time of acquiring Landsat datasets, serve as the ground truth. The choice of MODIS LST products over other standard LST products is due to Terra’s equatorial passing time of 10:30 am, which nears Landsat’s equatorial passing time, i.e. 10 am to 10:25 am. MOD11A1 and MOD21A1D products were utilised for October 22, 2008, and only MOD11A1 for October 2, 2018, due to the non-availability of MOD21A1D. Since the launch of MODIS onboard Terra happened in 1999 and the subsequent data availability began in February 2000, historical standard LST products are unavailable. Hence, cross-validation was not performed for 1991 and 1998.

For cross-validation, the projection system and spatial resolution of the derived and standard LST products must be identical. Therefore, the Landsat-derived LST was resampled to 1000 m and reprojected to the GCS WGS 84 coordinate system. LST values were extracted from Landsat and MODIS products at 3337 randomly sampled points within the study area. Correlation analysis was performed in the R programming language, considering Landsat-derived LST as the dependent variable and MODIS-derived LST as the independent variable.

Creation of LU maps

A false colour composite (FCC) was prepared for TM datasets of November 10, 1991, November 12, 1998, and October 22, 2008, by stacking the atmospherically corrected bands 4, 3, and 2. Atmospherically corrected bands 5, 4, and 3 were stacked to obtain FCC for the OLI dataset of October 2, 2018. These dates’ images were chosen as the scene, and land cloud covers were the lowest amongst all TM and OLI datasets according to metadata, and the study area had zero cloud cover upon visual examination in these dates’ images. Using visual interpretation skills and overlaying the FCCs on the high-resolution Google Earth imagery of the respective date and time of satellite data acquisition, randomly distributed regions of interest (ROIs) were assigned for each LU class. Each ROI is a digitised polygon representing true pixels of that LU class on the FCC image and serves as a signature for training the Gaussian Maximum Likelihood Classifier (GMLC). GMLC utilises a probability density function instead of a distance measure to designate a pixel to a particular class, thereby giving superior results to other supervised classification algorithms (Duda et al., 2012). The FCC images were classified into five classes: dense vegetation, sparse vegetation, built-up, dry river bed, and bare ground/soil to extract LU maps of 1991, 1998, 2008, and 2018. Validation of the classified outputs was possible by evaluating Cohen’s kappa statistic and overall and class-wise producer’s and user’s accuracies with the help of a confusion matrix. ROIs digitised on the FCC image around randomly sampled GCPs representative of each LU class on the high-resolution Google Earth imagery served as ground truth for accuracy assessment.

Calculation of spectral indices

As specified in Table 1, four spectral indices were calculated from TM datasets of November 10, 1991, November 12, 1998, and October 22, 2008, and the OLI dataset of October 2, 890–2018, to represent the LC types found in the study area. The reason for choosing these dates’ images is mentioned in the previous sub-section. They are the normalised difference built-up index (NDBI), normalised difference bareness index (NDBaI), NDVI, and NDWI. NDVI divides the difference between NIR and red bands by the sum of NIR and red bands to indicate vegetated areas. NDWI is the output of the division of the difference between green and NIR channels by the sum of the green and NIR channels. This index helps to detect water bodies. The detection of built-up areas is possible using NDBI, the value obtained after dividing the difference of shortwave infrared (SWIR) and NIR bands by its sum. NDBaI identifies barren lands such as open spaces and dry river beds by considering the division of the difference by the sum of SWIR and TIR channels. All four indices fall between −1 and 1, with plus figures indicating the presence of the LC class detected by the respective index and minus figures signifying the occurrence of other LC classes. Moreover, NDVI and NDWI utilise band images scaled to reflectance units as input, whereas NDBI and NDBaI use the band images expressed in DN as input. Table 3 shows the band number representing green, red, NIR, SWIR, and TIR channels in Landsat-5 and 8 datasets, which serve as input for estimating spectral indices.

Table 3.

Band number representing different channels in Landsat datasets

Sensor	Channel	Band number
TM	Green	2
	Red	3
	NIR	4
	SWIR	5
	TIR	6
OLI	Green	3
	Red	4
	NIR	5
	SWIR	6
TIRS	TIR	10

LU change analysis

Four LU change analyses: 1991–1998, 1998–2008, 2008–2018, and 1991–2018 were performed using a post-classification thematic change assessment. This assessment compares “before” and “after” classified images reflecting the amount and direction of change from one LU class to another during the time under consideration. The matrix union tool generated an output file consisting of a matrix of possible LU class conversions between the two classified images. The area of such conversions was calculated in km². A pivot table was produced to interpret the LU change patterns and trends.

Estimation of season-wise LST patterns and LST statistics per LU class

According to Sharma et al. (2012), the entire year was divided into four seasons for studying season-wise LST patterns: winter (December–February), (summer March–May), monsoon (June–September), and post-monsoon (October–November). Scene dates given in Table 1 were arranged for each year according to this division. A vector file consisting of 10,000 randomly spaced points was generated, out of which only 3337 random points were in the study area, and values corresponding to these points were extracted from each LU map raster. Repetitive and null values were discarded before determining minimum, maximum, range, standard deviation, and mean LST. Season-wise column plots were generated for minimum, mean, and maximum LST to study the seasonal patterns.

An attribute table containing the LU class grid code was generated for every LU map raster. LST maps of a particular year were overlaid over the LU map raster of the corresponding year. Then, LST and grid code values corresponding to 3337 points are extracted. Again, repetitive null values are discarded before sorting out the data according to grid code in increasing order, following which minimum, maximum, range, standard deviation, and mean LST are determined for each LU class seasonally.

Generation of population density maps

The cell-wise population density for the study area was obtained by dividing population counts by surface area. To generate population density as the “number of persons per square kilometre”, the obtained population density is multiplied by 1,000,000.

As the year-wise population count product is unavailable for 1998 in the WorldPop database and the year 2000 is the nearest to 1998, the population count product for 2000 was chosen to arrive at the population density variable to be correlated with LST values of 1998. Similarly, 1990 is the nearest to 1991, and the modelled population count for 1990 is available in the GHSL-POP data. Therefore, dividing the GHSL-POP data by the time-invariant WorldPop surface area dataset and multiplying by 1,000,000 generates population density as the “number of persons per square kilometre” for 1990 at 100 m spatial resolution.

Correlation analysis for assessing relationships

Four types of relationships are assessed in this study using three correlation tests in the R programming language. The tests are Pearson’s, Kendall, and Spearman’s rank correlation. Pearson’s correlation assumes a linear relationship between the dependent and independent variables, provided both follow a normal distribution. Both Kendall and Spearman correlation tests measure the strength and direction of association between two ranked variables, which do not follow a normal distribution. A t-test assesses the significance of dependence between the two variables considering a significance level α = 0.05 and 95% confidence.

The relationship types are LST and spectral indices, LST and elevation, spectral indices and elevation, and LST and population density. In the first, second, and fourth types, LST is the dependent variable, whereas NDVI, NDWI, NDBI, NDBaI, elevation, and population density are the independent variables. Each spectral index is the dependent variable in the third relationship type, whereas elevation is the independent variable.

For the relationship types involving elevation, we followed Mathew et al. (2016) and extracted elevation values corresponding to 3337 points. Their histogram was plotted, and only those values were considered whose frequency counts were eight and above. Therefore, points with elevations 552 m to 748 m were kept, and the remaining were discarded. Then, corresponding values from the dependent variable image were extracted for correlation tests.

In each case and relationship type, repetitive and null values were discarded before executing the correlation test.

Results and discussion

Temporal LU maps

Figure 2 below shows the LU maps generated by supervised GMLC for 1991, 1998, 2008, and 2018. A visual inspection of the classified outputs reveals an increment in the built-up class over the last 27 years in every direction. The densification of the existing built-up area over the years is also apparent from these LU maps. There is a considerable decrease in the bare ground class with the steady conversion of such lands into built-up or sparse vegetation classes. The dry river bed class evident in the 1991 and 1998 LU maps is only traceable in the 2008 LU map and is not visible in the 2018 LU map, indicating possibilities of encroachment of such lands. The dense vegetation class experienced a steady decline from 1991 to 2018, converting into sparse vegetation or bare ground and subsequently into a built-up class. The remaining dense vegetation patches in the 2018 LU map appear heavily fragmented.

Fig. 2.

LU maps of the Dehradun municipal area for 1991, 1998, 2008, and 2018

Tables 4 and 5 show the accuracy assessment results of individual LU maps. Every LU map reports an overall accuracy (OA) above the acceptable norm of 85%. The high values of the Kappa coefficient indicate a strong agreement of the classified output with the ground truth data. The individual class producer’s accuracy (PA) and user’s accuracy (UA) also echo this observation for 1991. In the 1998 LU map, the built-up class recorded a lower PA indicating a higher error of commission, while the dry river bed class recorded a lower UA indicating a higher error of omission. Hence, a possible assignment of dry river bed pixels to the built-up class cannot be ruled out. The very low PAs of the dry river bed class in the 2008 and 2018 LU maps indicate that this class occupies only a tiny proportion of a pixel containing more than one LU class, i.e. a mixed pixel. Such mixed pixels are assigned to the dominant LU class in that pixel by the hard classifier, here GMLC, thereby promulgating the need for soft classification. These dominant LU classes could be the built-up class and bare ground/soil class for the 2008 LU map and the built-up class only for the 2018 LU map. The low UA of the bare ground/soil class in the 2018 LU map indicates a misassignment of bare ground/soil class pixels to the built-up class. Similarly, though smaller, such misassignments are also possible between dense and sparse vegetation classes in the 2008 and 2018 LU maps.

Table 4.

OA and Kappa coefficient of LU maps obtained by supervised GMLC

Year	OA (%)	Kappa
1991	96.15	0.9503
1998	88.37	0.8532
2008	90.29	0.8572
2018	91.08	0.8667

Table 5.

Individual class PA and UA in LU maps of 1991, 1998, 2008, and 2018

Class/year	1991		1998		2008		2018
	PA (%)	UA (%)	PA (%)	UA (%)	PA (%)	UA (%)	PA (%)	UA (%)
Dense vegetation	96.85	98.62	93.24	97.18	93.58	87.06	87.50	92.02
Sparse vegetation	95.45	94.59	94.27	92.50	79.61	87.68	90.43	87.18
Built-up	99.04	96.88	75.61	88.07	98.11	94.06	98.10	96.50
Dry river bed	88.81	95.20	82.96	70.00	35.29	63.16	37.50	80.00
Bare ground/soil	96.43	94.41	100.00	92.00	89.83	91.38	81.25	56.52

Spectral index maps

Figure 3 above shows the NDBaI maps for 1991, 1998, 2008, and 2018. The maximum value of NDBaI increased from 0.19 in 1991 to 0.32 in 2008 and then declined to 0 in 2018. The minimum value decreased from −0.77 in 1991 to −0.87 in 1998 and then increased to −0.59 in 2018. Barren lands, such as open spaces and sandy river beds, are delineable in red, while the remaining LC classes are in other colours.

Fig. 3.

NDBaI maps for 1991, 1998, 2008, and 2018

Figure 4 shows the NDBI maps for 1991, 1998, 2008, and 2018. The maximum value of NDBI decreased from 0.60 in 1991 to 0.21 in 2018, while the minimum value rises from −0.46 in 1991 to −0.30 in 2008. A slight decline is visible in the minimum value to −0.39 in 2018. Unlike NDBaI maps, NDBI maps mix built-up, water, and barren lands leading to confusion in their extraction. Only vegetated areas are in grey.

Fig. 4.

NDBI maps for 1991, 1998, 2008, and 2018

Figure 5 shows the NDVI maps for 1991, 1998, 2008, and 2018. The greenest areas are easily differentiable in red, while other LC classes are visible in other colours. The red patches decreased from 1991 to 2018, indicating a steady conversion of vegetated areas into other LC classes.

Fig. 5.

NDVI maps for 1991, 1998, 2008, and 2018

Figure 6 shows the NDWI maps for 1991, 1998, 2008, and 2018. The decline of the maximum value from 0.43 in 1991 to 0 in 2018 indicates the incapability of the moderate resolution Landsat data to detect water in the narrow river beds over time. It also indicates the absence of water in these 27 years and possible encroachment of the river beds. Barren lands and vegetated areas are visible in grey and yellow. Built-up areas occupy red and blue shades, indicating confusion and intermixing with the water channel areas.

Fig. 6.

NDWI maps for 1991, 1998, 2008, and 2018

LU change

Figure 7 shows the LU change maps for 1991–1998, 1998–2008, 2008–2018, and 1991–2018. A look at these maps suggests the conversion of every LU class into the built-up class and subsequent manifold increase in the built-up area, especially in the low-lying southern and eastern parts. The densely vegetated northern parts of the city towards Mussoorie have become fragmented with conversion into sparse vegetation and built-up classes. The seasonal rivulets forming part of the dry river bed class have disappeared into the built-up or bare ground/soil classes. Conversion of bare ground/soil class into sparse vegetation is also apparent in the city’s south-eastern and southwestern portions.

Fig. 7.

LU change maps (1991–1998, 1998–2008, 2008–2018, and 1991–2018)

Table 6 shows the change in individual LU classes’ area from 1991 to 1998. It also shows the amount of area converted from one class to another during this period. The bare ground/soil class rose from 12.47 km² in 1991 to 12.85 km² in 1998. Out of 12.47 km², only 53% remained bare ground/soil in 1998, and the remaining 47% changed into other LU classes with maximum conversion into the sparse vegetation class and minimal conversion into the dry river bed class. The built-up class also increased from 18.22 km² in 1991 to 20.52 km² in 1998. The dry river bed class contributed the maximum to this increment. The dense vegetation class shrank by about 2.43 km² in 1998. About 2.64 km², 1. 62 km² and 1.25 km² of dense vegetation changed into sparse vegetation, bare ground/soil, and built-up classes, respectively. The dry river bed class also declined by about 0.78 km². In contrast, the sparse vegetation class increased from 7.49 km² in 1991 to 8.02 km² in 1998.

Table 6.

Change in area of individual LU class (1991–1998)

1991/1998	Bare ground/soil	Built-up	Dense vegetation	Dry river bed	Sparse vegetation	Total area (km2)
Bare ground/soil	6.65	1.50	1.26	0.88	2.18	12.47
built-up	1.16	14.06	0.34	2.01	0.65	18.22
Dense vegetation	1.62	1.25	7.38	0.16	2.64	13.05
dry river bed	1.46	2.32	0.11	1.78	0.22	5.89
Sparse vegetation	1.96	1.39	1.53	0.28	2.33	7.49
Total area (km2)	12.85	20.52	10.62	5.11	8.02	57.12

Table 7 shows the change in individual LU class areas from 1998 to 2008. The bare ground/soil class declined by about 8.97 km². At the same time, the built-up class increased by about 8.09 km². During these ten years, the dry river bed and dense vegetation classes shrunk by 2.88 km² and 1.11 km², respectively. The sparse vegetation class also increased by 4.87 km², with maximum contributions from converting bare ground/soil and dense vegetation classes into the sparse vegetation class. About 4.58 km² of bare ground/soil, 4.20 km² of the dry river bed and 2.24 km² of sparse vegetation changed into the built-up class. The least area conversion occurred from the dry river bed class to the dense vegetation class, i.e. 0.06 km².

Table 7.

Change in area of individual LU class (1998–2008)

1998/2008	Bare ground/soil	Built-up	Dense vegetation	Dry river bed	Sparse vegetation	Total area (km2)
Bare ground/soil	2.13	4.58	1.05	0.08	5.01	12.85
built-up	0.28	16.54	0.51	1.66	1.53	20.52
Dense vegetation	0.40	1.05	6.66	0.21	2.30	10.62
dry river bed	0.19	4.20	0.06	0.21	0.45	5.11
Sparse vegetation	0.88	2.24	1.23	0.07	3.6	8.02
Total area (km2)	3.88	28.61	9.51	2.23	12.89	57.12

Table 14 in the Appendix section shows the change in individual LU classes’ area from 2008 to 2018. The built-up class has shown a further increase of 4.26 km² to 32.87 km² in 2018. About 4.15 km² of sparse vegetation changed into built-up in these 10 years, contributing to the maximum increase in the built-up class area. Bare ground/soil increased from 3.88 km² in 2008 to 8.18 km² in 2018. At the same time, the dense vegetation class declines by about 4.89 km². The dry river bed and sparse vegetation classes decreased by 1.62 km² and 2.05 km², respectively, during this period. The maximum conversion of the dry river bed and bare ground/soil pixels occurs into built-up pixels in these 10 years. Similarly, about 3.24 km² of dense vegetation changes into sparse vegetation, contributing to the maximum increase in the sparse vegetation area.

Table 14.

Change in area of individual LU class (2008—2018)

2008/2018	Bare ground/soil	Built-up	Dense vegetation	Dry river bed	Sparse vegetation	Total area (km 2 )
Bare ground/soil	1.16	1.57	0.04	0.00	1.11	3.88
built-up	2.62	24.58	0.06	0.39	0.96	28.61
Dense vegetation	1.36	0.85	4.04	0.02	3.24	9.51
dry river bed	0.23	1.72	0.05	0.16	0.07	2.23
Sparse vegetation	2.81	4.15	0.43	0.04	5.46	12.89
Total area (km2)	8.18	32.87	4.62	0.61	10.84	57.12

Table 15 in the Appendix section shows the change in individual LU classes’ area from 1991 to 2018. The urban area rose from 18.22 km² in 1991 to 32.87 km² in 2018, while the rest of the LU classes experienced a decline, with the maximum being in the dense vegetation class. During these 27 years, about 6.29 km² and 3.46 km² of bare ground/soil were converted into built-up and sparse vegetation classes. On the contrary, only 1.71 km² of built-up was converted into bare ground/soil. About 4.23 km² of dry river red changed into built-up class, affirming the observation of encroachment of river beds by human settlements arrived upon by visual inspection of the LU maps. 2.21 km² and 3.58 km² of dense vegetation transformed into bare ground/soil and built-up classes during the same period.

Table 15.

Change in area of individual LU class (1991–2018)

1991/2018	Bare ground/soil	Built-up	Dense vegetation	Dry river bed	Sparse vegetation	Total area (km 2 )
Bare ground/soil	2.30	6.29	0.39	0.03	3.46	12.47
built-up	1.71	15.07	0.19	0.40	0.85	18.22
Dense vegetation	2.21	3.58	3.50	0.06	3.70	13.05
dry river bed	1.48	4.23	0.11	0.07	0.00	5.89
Sparse vegetation	0.48	3.70	0.43	0.05	2.83	7.49
Total area (km2)	8.18	32.87	4.62	0.61	10.84	57.12

On the contrary, only about 0.43 km² of sparse vegetation turned into dense vegetation. An area of 3.70 km² belonging to the sparse vegetation class also transformed into the built-up class. This continuous transition of other LU classes into the built-up class reflects the fast-paced infrastructure development and construction activity happening to house an ever-increasing population in the city and cater to their needs.

LST patterns

Figure 8 shows the representative LST maps generated from Landsat data for November 10, 1991, November 12, 1998, October 22, 2008, and October 2, 2018. Table 8 reveals the season-wise ranges of LST and its standard deviation for 1991, 1998, 2008, and 2018. The reader must note that the season-wise LST ranges in Table 8 reflect the minimum and maximum LST values of all the dates considered for the respective season in the considered years. Moreover, the column’ standard deviation range’ in Table 8 specifies only a single value if a single image is available for the respective season in the considered years. Figure 9 shows the minimum, maximum, and mean LSTs for the seasons under consideration. In summer, there is an increment of about 13.66 °C in the minimum temperature from 1991 to 2008, rising from 11.26 to 24.92 °C followed by a decrease of about 2.69 °C to 22.23 °C in 2018. The maximum summer temperature records an increase of 12.55 °C during the same time, rising from 32.65 to 45.2 °C. The mean LST values for the summer months indicate a gradual rise in temperature from March onwards to late April and May beginning in the years considered. The LST pattern for the monsoon months could not be studied properly as no cloud-free images were available for 1991, 1998, and 2008, and only a single cloud-free scene was available for 2018. LST figures for June 12, 2018, indicate a higher minimum temperature than the rest of the seasons. The maximum and mean temperatures are slightly less than summer’s maximum and mean temperatures in 2018, indicating the continuation of the hot and humid summer weather. In the post-monsoon months, the minimum and maximum LSTs rose by 1.32 °C and 5.15 °C in 1998, followed by a drop of 2.63 °C and 1.7 °C respectively, in 2008. Again, there is an increase of 8.13 °C and 5.41 °C in the minimum and maximum LST in 2018. Overall, there has been an increment of about 10 °C in mean LST during the post-monsoon period from 1991 to 2018. The lowest increment in minimum LST among the four seasons is in winter during these 27 years, which is 4.46 °C. The maximum LST has risen by 12.03 °C at the same time.

Fig. 8.

Representative Landsat-derived LST maps

Table 8.

Season-wise range of LST and standard deviation for 1991, 1998, 2008, and 2018

Year	Season (date)	LST range (°C)	Standard deviation range
1991	Summer (March 30 and May 17)	11.26–32.65	1.96–2.44
	Monsoon	-	-
	Post-monsoon (November 10)	15.94–25.46	1.03
	Winter (January 9)	9.52–15.87	0.67
1998	Summer (March 1, April 2, April 18, May 4, and May 20)	14.74–41.02	0.94–2.28
	Monsoon	-	-
	Post-monsoon (October 11, October 27, and November 12)	17.26–30.61	1.08–1.37
	Winter (December 14 and December 30)	11.7–22.45	0.97–0.98
2008	Summer (April 13, April 29, and May 31)	24.92–45.2	1.58–1.66
	Monsoon	-	-
	Post-monsoon (October 22 and November 23)	14.63–28.91	0.77–1.22
	Winter (December 9 and December 25)	12.25–25.08	0.78–0.79
2018	Summer (March 8, March 24, April 25, and May 11)	22.23–45.2	1.21–1.71
	Monsoon (June 12)	29.23–44.91	2.29
	Post-monsoon (October 2)	22.76–34.32	1.39
	Winter (January 19, February 4, February 20, December 21)	13.98–27.9	0.82 –0.87

Fig. 9.

a Column charts showing minimum, maximum, and mean LST for a summer (1991, 1998, 2008, 2018), b monsoon (2018), c post-monsoon (1991, 1998, 2008, 2018), d winter (1991, 1998, 2008, 2018)

The mean LST increment in winter has a similar pattern as the post-monsoon months, rising about 10 °C from 1991 to 2018. It is found upon examining the standard deviation ranges that the highest variation in LST values could be in summer and monsoon months and the least in winter. In the summer of 1991, the difference between the minimum and maximum LST is the highest, falling between 19.05 °C to 21 °C. It decreases subsequently to stabilise between 10 °C and 16 °C in the summer of 2018. Such a high difference could be possible because of the high standard deviation in summer. A temperature difference of 6 °C to 9.5 °C is evident in November, December, January, and February, irrespective of the year, and this difference has narrowed down from 1991 to 2018 instead of widening. If a uniform increment rate is assumed every year without considering any other influencing factor or data, the minimum and maximum summer temperatures rise by 0.44 °C and 1.05 °C per year from 1991 to 1998. This increment rate increased to 0.93 °C per year for minimum summer temperature from 1998 to 2008, whereas this rate decreased to 0.38 °C per year for maximum summer temperature. From 2008 to 2018, the minimum summer temperature decreased by 0.24 °C per year, while the maximum summer temperature remained constant. For winter, the increment rate stands at 0.27 °C and 0.82 °C per year for minimum and maximum LST, respectively, from 1991 to 1998. The increment rate declined to 0.05 °C and 0.24 °C per year for minimum and maximum winter LST from 1998 to 2008. From 2008 to 2018, the increment rate rose slightly to 0.16 °C and 0.26 °C per year for minimum and maximum winter LST. During October–November, the minimum and maximum LST rise by 0.17 °C and 0.64 °C per year, respectively, from 1991 to 1998. In the next 10 years, both minimum and maximum LST declined by 0.24 °C and 0.15 °C per year, respectively. From 2008 to 2018, the minimum temperature rose sharply by 0.73 °C per year compared to the increment in maximum LST, which was only 0.49 °C per year.

Fig. 10 shows the LST derived from MODIS products MOD11A1 and MOD21A1D for October 22, 2008, and the resampled Landsat-derived LST. The nearest neighbour resampling of the Landsat-derived LST to the spatial resolution of the MODIS product increases the minimum temperature from 19.59 to 21.85 °C and decreases the maximum temperature from 28.91 to 28.07 °C. Due to the difference in algorithms adopted for retrieving LST from MODIS data, both the MODIS products record a difference in their LST ranges, i.e. 22.27–26.33 °C (MOD11A1) and 23.65–27.43 °C (MOD21A1D).

Fig. 10.

a MODIS LST (MOD11A1.006), b MODIS LST (MOD21A1D.061), c Resampled Landsat LST for 2008

Figure 11 shows the LST derived from MODIS product MOD11A1 for October 2, 2018, and the resampled Landsat-derived LST. The nearest neighbour resampling of the Landsat-derived LST to the spatial resolution of the MODIS product increases the minimum temperature from 22.76 to 24.55 °C and decreases the maximum temperature from 34.32 to 31.95 °C. The MODIS LST range is 24.39–28.41 °C.

Fig. 11.

a MODIS LST (MOD11A1.006), b Resampled Landsat LST for 2018

For 2008, the correlation coefficient of 0.54 shows that both Landsat and MOD11A1 LST are moderately correlated, indicating moderate reliability of the LST derived from Landsat. The correlation coefficient between Landsat and MOD21A1D is only slightly lower at 0.48. On the other hand, 0.70 is the value reported for 2018, indicating the high reliability of the LST derived from Landsat. For both years, the Landsat LST values have cleared the t-test with MODIS LST values using α = 0.05. There are three reasons for the moderate correlation between Landsat LST and MODIS LST for 2008. First, the difference in approaches of retrieval of LST.

The split-window algorithm utilised for LST retrieval in the MOD11A1 product reduces the impact of atmospheric absorption in the TIR on the derived LST.

The temperature-emissivity separation algorithm utilised for LST retrieval in the MOD21A1D product uses a physics-based scheme and an improved water vapour scaling atmospheric correction model to minimise the effect of atmospheric parameters on the derived LST. Moreover, the methodology for deriving LST from Landsat considers global atmospheric conditions due to the absence of historical localised meteorological parameters. It utilises a single-channel NDVI-thresholding scheme for calculating emissivity under global conditions.

Second, there is a time difference of about 25 to 30 min between the equatorial crossing times of Landsat and Terra satellites which could lead to different LST calculations in the case of MODIS. Third, the resampling of moderate resolution Landsat-derived LST to the coarser spatial resolution of MODIS LST might have introduced uncertainty in the data.

Nevertheless, to further ascertain the reliability of Landsat-derived LST, the ranges of both the Landsat and MODIS products were examined, and the difference was computed between them. MOD11A1 LST product reports a difference of only 0.62 °C and −1.74 °C in the minimum and maximum LST values from the respective Landsat LST values for 2008. On the other hand, the MOD21A1D LST product records a difference of 1.8 °C and only −0.64 °C in the minimum and maximum LST values from the respective Landsat LST values for 2008. For 2018, the MOD11A1 LST product reports a difference of only −0.16 °C and −3.54 °C in the minimum and maximum LST values from the respective Landsat LST values. The mean temperature values further reflect the closeness of the LST values between both the datasets: 25.31 °C (MOD11A1), 26.54 °C (MOD21A1D), and 25.011 °C (Landsat) for 2008; 27.39 °C (MOD11A1), and 29.07 °C (Landsat) for 2018. Hence, it is possible to use Landsat-derived LST for further analysis.

LST patterns for individual LU class

Table 9 shows the season-wise LST statistics for each LU class. Every LU class has experienced an increment in minimum and maximum temperatures during these 27 years, irrespective of the season. The dense vegetation class records an 11.32 °C and 10.33 °C increase in the lowest and highest summer LST between 1991 and 2008. There is a decline of 2.26 °C and 1.57 °C in the minimum and maximum summer LST from 2008 to 2018 for the dense vegetation class. The sparse vegetation class records an 11.4 °C and 12.25 °C increase in the minimum and maximum summer temperature from 1991 to 2008. In contrast, there is a decrease of 2.59 °C and 0.47 °C in the summer minima and maxima for the next 10 years.

Table 9.

LST (°C) per LU class

LU Class	Statistic	Summer				Post-monsoon				Winter				Monsoon
		‘91	‘98	‘08	‘18	‘91	‘98	‘08	‘18	‘91	‘98	‘08	‘18	‘18
Dense vegetation	Mean	24.49	26.24	33.06	30.93	19.31	22.02	20.63	26.67	12.50	15.74	17.11	18.94	33.54
	Range	13.73–31.84	15.67–38.29	25.05–42.17	22.79–40.60	16.37–21.90	17.79–29.59	15.36–27.97	23.57–29.76	10.34–15.51	12.13–20.40	13.17–22.22	15.08–24.81	29.86–41.97
	Standard deviation	3.00	4.90	2.73	4.17	0.90	1.75	2.94	1.33	0.81	1.63	2.03	2.03	2.15
Sparse vegetation	Mean	25.61	27.49	34.55	32.73	19.55	23.13	21.68	28.07	12.56	15.84	17.74	19.64	35.72
	Range	14.68–31.82	16.51–39.46	26.08–44.07	23.49–43.60	17.79–21.86	19.12–30.18	16.75–27.60	24.14–31.00	11.21–14.65	12.54–20.94	13.69–23.34	15.74–27.35	30.43–44.06
	Standard deviation	2.16	5.64	2.81	4.60	0.77	1.88	3.07	1.09	0.60	1.55	2.03	2.24	2.28
Built-up	Mean	26.81	28.73	35.23	33.28	20.03	24.11	22.17	29.68	12.48	15.79	17.50	19.44	37.83
	Range	14.19–32.65	16.72–37.86	28.49–41.75	23.61–41.83	17.84–23.72	18.78–30.17	16.53–28.09	25.85–33.54	11.18–14.67	12.28–21.10	14.17–23.33	14.78–25.12	31.61–41.64
	Standard deviation	2.39	5.97	2.38	4.51	0.78	2.09	3.45	0.93	0.52	1.59	2.05	2.30	1.66
Dry river bed	Mean	26.80	28.43	35.09	33.38	20.68	24.03	22.15	29.76	12.69	16.03	17.46	19.52	37.99
	Range	15.15–32.24	16.41–37.50	30.18–41.02	27.01–40.60	17.87–25.46	20.18–28.94	17.00–27.64	27.75–31.88	11.20–15.60	12.74–20.64	14.19–21.12	15.98–24.77	34.89–41.29
	Standard deviation	2.35	5.91	2.32	4.44	1.04	1.89	3.43	0.90	0.62	1.58	2.03	2.34	1.47
Bare ground/soil	Mean	25.96	27.93	35.17	32.80	20.60	23.68	22.25	28.81	12.85	16.59	18.21	19.57	36.54
	Range	14.67–31.83	15.62–39.46	26.61–44.82	22.59–44.56	18.27–25.00	20.07–29.36	16.95–28.51	24.69–33.92	11.13–15.15	13.01–21.57	14.20–25.07	15.57–26.77	30.77–41.88
	Standard deviation	2.35	5.98	3.10	4.52	1.00	1.73	3.26	1.21	0.74	1.79	2.19	2.24	2.30

In the built-up class, the minimum and maximum summer temperature increases are 14.3 °C and 9.1 °C between 1991 and 2008. In contrast, the minimum temperature declined by 4.88 °C and the maximum temperature increased by 0.08 °C from 2008 to 2018. Among all the LU classes, the maximum rise in minimum summer LST has been for the dry river bed class from 1991 to 2008, i.e. about 15.03 °C. The maximum summer LST rose by only 8.78 °C during the same period. However, minimum and maximum LST experienced a decline of 3.17 °C and 0.42 °C, respectively, in the next 10 years. The bare ground/soil class reports an 11.94 °C and 12.99 °C increase in the lowest and highest summer LST between 1991 and 2008. A decrement of 4.02 °C and 0.26 °C was observed in the minimum and maximum LST from 2008 to 2018. The dry river bed class had the highest minimum LST, while the built-up class had the highest maximum LST in the 1991 summer. In the 1998 summer, the built-up class had the largest minimum LST among all the LU classes, whereas sparse vegetation and bare ground classes jointly had the highest maximum LST. In the 2008 and 2018 summers, the dry river bed and bare ground/soil classes recorded the highest minimum and maximum LST among all the LU classes.

Although LST results from June 12, 2018, cannot reflect the monsoon pattern, they represent a version of extended summer. Again, the dry river bed and bare ground/soil classes have the highest minimum and maximum LST among all the LU classes. The sparse vegetation class has the largest difference between the maximum and minimum LST at 13.63 °C.

In October–November, the minimum LST for the dense vegetation class increased by 1.42 °C in 1998, followed by a decline of 2.43 °C in 2008 and a sharp increase of 8.21 °C in 2018. The maximum LST had a sharp increase of 7.69 °C in 1998, followed by a decrease of 1.62 °C in 2008 and an increment of 1.79 °C in 2018. For the sparse vegetation class, the minimum LST increased by 1.33 °C in 1998, followed by a decline of 2.37 °C in 2008 and a sharp increment of 7.39 °C in 2018. The maximum LST rose sharply by 8.32 °C in 1998, followed by a decrement of 2.58 °C in 2008 and an increase of 3.4 °C in 2018. The built-up class’s minimum LST increased by 0.94 °C in 1998, followed by a decrease of 2.25 °C in 2008 and a sharp rise of 9.32 °C in 2018. The maximum LST increased sharply by 6.45 °C in 1998, then a 2.08 °C decrease in 2008 and a 5.45 °C increase in 2018. For the dry river bed class, the minimum LST increased by 2.31 °C in 1998, then a 3.18 °C decline in 2008 and after that, a 10.75 °C increase in 2018, which is the highest among all classes. The maximum LST rose by 3.48 °C in 1998, followed by a decrement of 1.3 °C in 2008 and a 4.24 °C increase in 2018. The bare ground/soil class’s minimum LST has a similar trend as other classes’ minimum LST, experiencing a 1.8 °C rise in 1998, followed by a 3.12 °C decline in 2008 and afterwards a 7.74 °C sharp increase in 2018. The maximum LST experienced a 4.36 °C increase in 1998, followed by a 0.85 °C decrease in 2008 and a 5.41 °C increase in 2018.

In the post-monsoon of 1991, the bare ground/soil and dry river bed classes had the highest minimum and maximum LST among all the LU classes. In 1998, 2008, and 2018 post-monsoon months, the dry river bed had the highest minimum LST. The sparse vegetation class had the highest maximum LST in 1998 post-monsoon, whereas the bare ground/soil class had the highest maximum LST in both 2008 and 2018 post-monsoon. The built-up class is a close second in 1998, 2008, and 2018.

Irrespective of the year, the winter months see a consistent rise in minimum and maximum LST for every LU class. The minimum and maximum LST for dense vegetation rose by 4.74 °C and 9.3 °C between 1991 and 2018. Sparse vegetation’s minimum and maximum LST increased by 4.53 °C and 12.7 °C simultaneously. The built-up class recorded an increment of 3.6 °C and 10.45 °C in the minimum and maximum LST during these 27 years. An increase of 4.78 °C and 9.17 °C is observed in the dry river bed’s minimum and maximum LST from 1991 to 2018. For the bare ground/soil class, the minimum and maximum LST reported 4.44 °C and 11.62 °C increases between 1991 and 2018.

In the 1991 winter, sparse vegetation and dry river bed classes had the highest minimum and maximum LST among the LU classes. In contrast, bare ground/soil had the highest minimum and maximum LST in the 1998 and 2008 winters. In the 2018 winter, the dry river bed and sparse vegetation classes had the highest minimum and maximum LST, respectively.

The mean LST per LU class has also risen during these 27 years, and its trend depends on the season, being identical to the trends of the minimum and maximum LST per LU class. The highest mean LST for each LU class is the results of June 12, 2018, or the “monsoon '18” column in Table 9, followed by the 2008 and 2018 summers, respectively. The bare ground/soil class experienced the maximum rise in mean LST between 1991 and 2008 summer. It also experienced the sharpest decline in the ten summers following the 2008 summer. Between the post-monsoons of 1991 and 1998, the built-up class experienced the highest rise in mean LST and even the highest decline between the post-monsoons of 1998 and 2008. However, the dry river bed class has the highest warming from 2008 to 2018 post-monsoon. The sparse vegetation class has the highest increment in the mean LST between the winters of 1991 and 2018.

The standard deviation varies between 0.52 and 5.98 and depends on the season. The maximum deviation is maximum in the summer months, with the highest in the 1998 summer, followed by the 2018 summer. Each LU class has the lowest standard deviation in the 1991 winter and the highest standard deviation in the 1998 summer. The winter season reports a constant increase in every LU class’s standard deviation during these 27 years. The trend of a standard deviation per LU class in the post-monsoon season reflects the trend of minimum and maximum LST per LU class in the summer season between 1991 and 2018. Likewise, the standard deviation per LU class’s trend in the summer season resembles the trend of maximum and minimum LST per LU class in the post-monsoon season from 1991 to 2018. In 1991 summer, the sparse vegetation and dense vegetation classes had the lowest and highest standard deviation among all the LU classes. The dense vegetation and dry river bed classes had the lowest standard deviation in the 1998 and 2008 summers.

In contrast, the bare ground/soil class had the highest standard deviation in the 1998 and 2008 summers. In the 2018 summer, dense and sparse vegetation classes had the lowest and highest standard deviations. In the post-monsoon season, sparse vegetation and dry river bed classes had the lowest and highest standard deviation in 1991. In 1998 and 2008, bare ground/soil and dense vegetation classes had the minimum standard deviation, respectively, whereas the built-up class had the maximum standard deviation in both 1998 and 2008.

In contrast, the dry river bed and dense vegetation classes had the lowest and highest standard deviation in 2018 post-monsoon. For winter, built-up and dense vegetation classes had the minimum and maximum standard deviation in 1991, while in 1998, sparse vegetation had the lowest standard deviation. In 2008, sparse vegetation had the lowest standard deviation, together with dense vegetation and dry river bed classes. In both the 1998 and 2008 winters, bare ground/soil had the maximum standard deviation. The dry river bed and dense vegetation classes had the highest and lowest standard deviation in the 2018 winter.

Relationships between LST and its causal factors

Figure 12 shows the population density maps of Dehradun for 1990, 2000, 2008, and 2018. Figure 13 shows the elevation of the Dehradun municipal area. The positions of the randomly distributed points selected for validation of LST and the examination of the relationship between LST and its causal factors are also shown in Fig. 13.

Fig. 12.

Population density of Dehradun for 1990, 2000, 2008, and 2018

Fig. 13.

a Elevation of Dehradun municipal area above mean sea level, b Location of randomly sampled points chosen for correlation analyses

The population density has increased significantly during these 27 years, with more and more areas falling in the “high” and “very high” categories. Till 2000, only the city centre at Ghantaghar (30.324426 N, 78.041833 E) and its surrounding areas up to a radius of 4–5 km had “medium” to “very high” population density. From 2000 to 2018, the low-lying southern areas adjoining the Shivalik range and the Rajaji tiger reserve saw a drastic concentration of population with the rapid construction of houses, government offices, the new inter-state bus terminus, and road widening activities, thereby transcending from “low” and “medium”, respectively to “medium” and “high” categories. Even the highly elevated northern areas near Rajpur village and the Malsi reserve forest have changed from the “very low” to “low” category, as shown in the 2018 population density map.

The creation of upscale neighbourhoods such as Doon Vihar at 800 m and above in the Jakhan area within 10 years has changed the population density category of this area from “low” in 2008 to “medium” and “high” in 2018. Figure 13 shows that the altitude ranges from 539 m in the southern to 1007 m in the northern areas. Tables 10, 11, 12, and 13 show the correlation coefficients between elevation and spectral indices from 1991 to 2018. As per convention, irrespective of sign, correlation coefficients below 0.4 indicate a low and insignificant correlation, above 0.4 and below 0.7 indicate a moderate correlation, and above 0.7 indicate a strong correlation. A low positive correlation is evident between NDVI and elevation. A low negative correlation is evident between NDWI and elevation. Similarly, low negative correlations are possible for NDBI and NDBaI with elevation.

Table 10.

Correlation results for NDVI vs elevation (1991, 1998, 2008, 2018)

NDVI vs elevation
Correlation coefficient/year	1991	1998	2008	2018
Pearson	0.17	0.25	0.29	0.37
Kendall	0.07	0.12	0.10	0.15
Spearman	0.11	0.20	0.16	0.23

Table 11.

Correlation results for NDWI vs elevation (1991, 1998, 2008, 2018)

NDWI vs elevation
Correlation coefficient/year	1991	1998	2008	2018
Pearson	−0.16	−0.25	−0.27	−0.36
Kendall	−0.06	−0.10	−0.09	−0.16
Spearman	−0.08	−0.16	−0.14	−0.23

Table 12.

Correlation results for NDBI vs elevation (1991, 1998, 2008, 2018)

NDBI vs elevation
Correlation coefficient/year	1991	1998	2008	2018
Pearson	−0.20	−0.26	−0.29	−0.28
Kendall	−0.19	−0.19	−0.15	−0.11
Spearman	−0.28	−0.29	−0.23	−0.17

Table 13.

Correlation results for NDBaI vs elevation (1991, 1998, 2008, 2018)

NDBaI vs elevation
Correlation coefficient/year	1991	1998	2008	2018
Pearson	−0.17	−0.13	−0.19	−0.21
Kendall	−0.24	−0.22	−0.24	−0.18
Spearman	−0.35	−0.33	−0.34	−0.27

Tables 16, 17, 18, and 19 in the Appendix section show the results of the correlation analyses between LST and spectral indices for 1991, 1998, 2008, and 2018, respectively. The reader must note that the correlation analysis between LST and the respective spectral index has been performed only for a particular day in the considered year, i.e. November 10, 1991; November 12, 1998; October 22, 2008; and October 2, 2018. These results must not be generalised to any particular season or year. The correlation coefficients in these tables clearly show that the independent variables do not follow a normal frequency distribution. A non-linear equation of more than one independent variable can model the dependent variable’s relationship with the independent variables. Among the spectral indices, NDBI and NDBaI are moderately positively correlated with LST, indicating an increase in LST with an increase in built-up and barren lands. On the other hand, NDVI is moderately negatively correlated with LST, indicating a decreasing LST with an increment in vegetation. The correlation coefficient between NDWI and LST is low and not significant.

Table 16.

Correlation results for LST vs spectral indices (1991)

LST vs spectral indices (1991)
Correlation coefficient/index or elevation	NDVI	NDWI	NDBI	NDBaI
Pearson	−0.22	0.13	0.46	0.23
Kendall	−0.35	0.26	0.48	0.37
Spearman	−0.48	0.37	0.66	0.52

Table 17.

Correlation results for LST vs spectral indices (1998)

LST vs spectral indices (1998)
Correlation coefficient/index or elevation	NDVI	NDWI	NDBI	NDBaI
Pearson	−0.27	0.24	0.42	0.09
Kendall	−0.39	0.36	0.45	0.26
Spearman	−0.52	0.48	0.62	0.37

Table 18.

Correlation results for LST vs spectral indices (2008)

LST vs spectral indices (2008)
Correlation coefficient/index or elevation	NDVI	NDWI	NDBI	NDBaI
Pearson	−0.41	0.39	0.49	−0.11
Kendall	−0.5	0.49	0.49	0.01
Spearman	−0.66	0.66	0.66	0.03

Table 19.

Correlation results for LST vs spectral indices (2018)

LST vs spectral indices (2018)
Correlation coefficient/index or elevation	NDVI	NDWI	NDBI	NDBaI
Pearson	−0.48	0.47	0.45	−0.14
Kendall	−0.53	0.53	0.53	0.20
Spearman	−0.72	0.72	0.71	0.31

For 1998, NDBI is moderately positively correlated with LST, indicating an increase in LST with increased built-up. Surprisingly, NDWI is also moderately positively correlated with LST, although it fails to separate rivulets from the built-up areas. NDVI is moderately negatively correlated with LST, indicating a decreasing LST with increasing vegetation. NDBaI and LST do not have significant agreeability.

Again, NDBI and NDWI are moderately positively correlated with LST, reporting identical correlation coefficients. NDVI is moderately negatively correlated with LST. The relationship between NDBaI and LST is insignificant for 2008.

For 2018, NDWI and NDBI reveal a highly positive correlation with LST, whereas NDVI reveals a highly negative correlation. A low positive correlation is also present between NDBaI and LST.

Table 20 in the Appendix section shows the results of the correlation analysis between LST and elevation. In 1991, the correlation between LST and elevation was moderate throughout, except significant drop in March and the sign change from positive to negative in November. In 1998, March and April’s beginnings showed a weak correlation. By mid-April, the correlation sign became negative, and the coefficient magnitude increased drastically, reaching 0.90 by mid-May. October showed a moderate negative correlation which became a high negative correlation by November beginning. The correlation was insignificant for December. In 2008, the correlation was negative throughout, irrespective of the month. A weak correlation in mid-April turned into a moderate correlation by May end. A moderate correlation in October transitioned to a high one as November ended. A weak correlation was observed as December began, which changed into a moderate correlation around Christmas eve. In 2018, only January showed a positive correlation, which was moderate. The rest of the year reported a negative correlation. February 4 reported a weak correlation, whereas February 20 showed a moderate correlation.

Table 20.

Correlation results for LST vs elevation

LST vs elevation
LST date/correlation coefficient	Pearson	Kendall	Spearman
January 9, 1991	0.55	0.12	0.20
March 30, 1991	0.19	0.17	0.30
May 17, 1991	0.46	0.36	0.50
November 10, 1991	−0.41	−0.40	−0.54
March 1, 1998	0.24	0.17	0.25
April 2, 1998	0.14	0.12	0.20
April 18, 1998	−0.39	−0.15	−0.22
May 4, 1998	−0.66	−0.45	−0.60
May 20, 1998	−0.90	−0.72	−0.86
October 11, 1998	−0.49	−0.17	−0.26
October 27, 1998	−0.49	−0.20	−0.30
November 12, 1998	−0.71	−0.62	−0.81
December 14, 1998	−0.09	−0.15	−0.19
December 30, 1998	−0.19	−0.33	−0.42
April 13, 2008	−0.18	−0.13	−0.18
April 29, 2008	−0.57	−0.47	−0.63
May 31, 2008	−0.68	−0.50	−0.68
October 22, 2008	−0.44	−0.20	−0.36
November 23, 2008	−0.65	−0.57	−0.75
December 9, 2008	−0.22	−0.28	−0.37
December 25, 2008	−0.47	−0.54	−0.61
January 19, 2018	0.49	0.40	0.54
February 4, 2018	−0.09	−0.07	−0.11
February 20, 2018	−0.37	−0.39	−0.54
March 8, 2018	−0.02	−0.03	−0.03
March 24, 2018	−0.58	−0.43	−0.60
April 25, 2018	−0.71	−0.59	−0.74
May 11, 2018	−0.85	−0.76	−0.90
June 12, 2018	−0.83	−0.54	−0.72
October 2, 2018	−0.60	−0.30	−0.47
December 21, 2018	−0.48	−0.46	−0.61

March’s beginning showed an insignificant correlation. From March 24 to June 12, the correlation magnitude increased, becoming moderate to high. The dates in both October and December reported a correlation coefficient of 0.60, indicating a moderately negative relationship between LST and elevation. Table 20 shows that wherever Pearson’s coefficient is low, Spearman’s coefficient is high, indicating a non-linear relationship between the dependent, i.e. LST, and independent, i.e. elevation, variable. Kendall’s coefficient always has the lowest value among the three coefficients. The month-wise/season-wise change from weak to moderate to high correlation can be attributed to the growth cycle of vegetation and its type in this region. The region is predominantly characterised by tropical moist and dry deciduous forest types, which begin shedding their leaves in October–November only to begin growing back from August onwards (Chandrashekhar et al., 2005). Hence, bare land is visible in satellite datasets of this region in the first 4–5 months of the year, while datasets acquired in the later half of the year show a dense forest canopy. So, the thermal sensor effectively measures the warming of bare land at the beginning of the year, while in the latter half, the warming of vegetation is measured, hence the change of sign in correlation coefficient from positive to negative and vice-versa.

Tables 21, 22, 23, and 24 in the Appendix section show the results of correlation analyses between LST and population density for the respective years. The correlation was insignificant to weak in 1991. In 1998, only December reported a weak to moderate negative correlation, while the rest of the year was weak to moderate positive. Above 0.5 values were reported for both April and October. Again in 2008, December recorded a negative correlation being weak and insignificant. For the remainder of the year, the correlation was weakly positive at the end of April, May, and November. The datasets at the beginning of April and October reported a moderate correlation of above 0.5. In 2018, both December and January reported a weak negative correlation.

Table 21.

Correlation results for LST (1991) vs population density (1990)

LST (1991) vs population density (1990)
LST date/correlation coefficient	Pearson	Kendall	Spearman
January 9, 1991	−0.07	−0.002	0.09
March 30, 1991	0.33	0.23	0.35
May 17, 1991	0.19	0.12	0.17
November 10, 1991	0.07	0.09	0.13

Table 22.

Correlation results for LST (1998) vs population density (2000)

LST (1998) vs population density (2000)
LST date/correlation coefficient	Pearson	Kendall	Spearman
March 1, 1998	0.29	0.19	0.27
April 2, 1998	0.57	0.38	0.54
April 18, 1998	0.55	0.39	0.55
May 4, 1998	0.40	0.26	0.37
May 20, 1998	0.37	0.24	0.34
October 11, 1998	0.56	0.41	0.58
October 27, 1998	0.50	0.35	0.49
November 12, 1998	0.15	0.13	0.18
December 14, 1998	−0.24	−0.15	-0.24
December 30, 1998	−0.38	−0.27	-0.40

Table 23.

Correlation results for LST (2008) vs Population Density (2008)

LST (2008) vs population density (2008)
LST Date/correlation coefficient	Pearson	Kendall	Spearman
April 13, 2008	0.50	0.35	0.51
April 29, 2008	0.28	0.18	0.26
May 31, 2008	0.29	0.22	0.30
October 22, 2008	0.54	0.42	0.58
November 23, 2008	0.16	0.09	0.13
December 9, 2008	−0.08	−0.02	−0.03
December 25, 2008	−0.23	−0.14	−0.21

Table 24.

Correlation results for LST (2018) vs population density (2018)

LST (2018) vs population density (2018)
LST date/correlation coefficient	Pearson	Kendall	Spearman
January 19, 2018	−0.24	−0.10	−0.16
February 4, 2018	0.21	0.15	0.21
February 20, 2018	0.25	0.17	0.25
March 8, 2018	0.34	0.24	0.35
March 24, 2018	0.46	0.32	0.47
April 25, 2018	0.42	0.29	0.42
May 11, 2018	0.37	0.25	0.36
June 12, 2018	0.63	0.44	0.61
October 2, 2018	0.70	0.53	0.72
December 21, 2018	−0.08	−0.01	−0.03

From February onwards, the correlation coefficient became positive, and its magnitude rose to a peak at the end of March. Again, it declined in April and May, only to increase in June to above 0.6. A further increment is observed in October, leading to a high positive correlation between LST and population density. A reason for the weak correlation in 1991 could be the data product, i.e. GHSL-POP. GHSL-POP is lightly modelled data, considering only the population estimates and projections of the United Nations Population Division (UNPD) and the building footprint information (Leyk et al., 2019). The 1990 population density map shown in Fig. 12 has been synthesized using GHSL-POP. The map has white spaces, indicating missing data. In contrast, WorldPop is heavily modelled data, considering many ancillary geospatial datasets: actual country-wise census data, roads, land cover, urban extent, night-time lights, infrastructure, climate, topography, elevation, water bodies, and protected areas (Leyk et al., 2019), thereby giving a more accurate picture of the population density. Hence, the floating population is also considered in the WorldPop data product, especially tourists, which are abundant in April–June and October–December and may be a reason for the moderate to high correlation in these months in 2008 and 2018.

Conclusion

This study successfully extracted LU maps of the Dehradun municipal area for 1991, 1998, 2008, and 2018, with OAs and Kappa coefficients above 88% and 0.85, respectively. Individual class PAs and UAs indicated misassignment of the dry river bed pixels to the bare ground/soil or the built-up pixels. A slight misassignment of bare ground/soil pixels to the built-up pixels was also prevalent. Similarly, sparse vegetation pixels were assigned to dense vegetation pixels and vice-versa, prompting the need to use soft classifiers.

LU change patterns revealed an increase of built-up from 18.22 km² in 1991 to 32.87 km² in 2018, while the remaining LU classes declined in their areas. About 6.29 km² of bare ground/soil, 4.23 km² of dry river red, 3.70 km² of sparse vegetation, and 3.58 km² of dense vegetation converted into the built-up area during these 27 years.

The maximum value of NDWI and NDBaI decreased over the years indicating such maps highlight the non-target LC classes, especially the built-up, in a better manner. It is difficult to differentiate built-up from water and barren areas in NDBI maps. Such observations highlight the need to incorporate alternate spectral indices which can easily detect the target LC classes and facilitate their extraction.

The LST derived from Landsat was successfully validated with MODIS LST for both the MODIS LST products. Season-wise LST figures indicate the largest increments in summer followed by post-monsoon and winter during these 27 years. In the absence of cloud-free data, no pattern for LST could be determined for monsoon months. The highest and lowest variations are in the summer and winter months. The mean LST has risen above 9 °C every season, whereas the minimum and maximum LST have increased to 8–12 °C in summer and post-monsoon. Even in winter, the minimum LST has risen above 4 °C. These figures indicate the formation of UHI in the study area. The rise and fall pattern for the LST of each LU class is dependent on the season and mimics the rise and fall pattern of season-wise minimum, maximum, and mean LST from 1991 to 2018.

A moderate positive correlation exists between NDBI, NDWI, and LST in most cases, whereas a negative correlation exists between NDVI, elevation and LST. No significant correlation exists between the spectral indices and elevation, prompting the need to ascertain the nature of frequency distribution of the spectral indices and elevation variables for modelling the relationship between them. A possible influence of the vegetation growth cycle is evident in the LST vs elevation correlation coefficient’s sign and magnitude. Choosing gridded population datasets to understand population density due to constantly changing municipal ward boundaries proved successful. The WorldPop data product has shown its effectiveness in retrieving accurate population densities for 2000, 2008, and 2018 and having a moderate to high correlation with LST, thereby implying its importance as a predictor variable along with spectral indices and elevation. An influence of season type is possible on the population density, as evinced from the correlation coefficient’s sign and magnitude in certain months. On the contrary, a weak correlation was reported between 1991 LST and 1990 population density due to the use of the GHSL-POP data product.

Further work pertains to understanding the influence of meteorological variables on spectral indices, elevation, population density, and LU change during the considered period. Analyzing the LU change with respect to other demographic data such as age, gender, workforce participation, the proportion of disadvantaged sections of society, literacy, infant and maternal mortality rates, and life expectancy will be considered in the future study.

Author contribution

Kavach Mishra: conceptualisation, data processing, analysis and interpretation, writing—original draft preparation and editing. Rahul Dev Garg: supervision, reviewing, editing.

Data availability

All data generated or analysed during this study are included in this published article.

Declarations

Competing interests

The authors declare no competing interests.