1 Introduction

Land cover alteration, driven by increased anthropogenic activities, is a common phenomenon across the world but the intensity is high in developing countries like Bangladesh (Wurm and Taubenböck 2018; Panday 2020). Many studies have investigated drivers behind rapid land use and land cover (LULC) change, and have showed that population growth (Dewan and Corner 2014a; El-Zeiny and Effat 2017; Yohannes et al. 2021), economic affluence (Karakuş 2019), industrialization and rural to urban migration (Mberu et al. 2017) are major reasons behind LULC modification. This modification evidently impacts the environment in various ways, including deforestation (Behera et al. 2018), pollution (Hua 2017), flooding (Rahman et al. 2021), biodiversity loss (Sharma et al. 2018) and local warming (Ullah et al. 2019). Temperature is directly influenced by LULC change and has a substantial impact on global/regional climate (Cai et al. 2018; Peng et al. 2018; Wang et al. 2019; Karakuş 2019). It is also a critical variable for regulating surface energy balance (He et al. 2019). Though both air and land surface temperature (LST) are influenced by a number of factors, non-evaporative surfaces such as urban infrastructures, road, factories, buildings or any type of construction and dry surfaces significantly influence LST (Kayet et al. 2016; El-Zeiny and Effat 2017; Sannigrahi et al. 2018). Thus, LST is a major input to examine surface urban heat island (SUHI) (Chakraborty and Lee 2019), resulting from large-scale modification of land use/cover (Dewan et al. 2021). This kind of situation is pervasive in unplanned metropolis with insufficient amount of greenery and water bodies. Hence, it is well understood that LST estimation and identifying its relationship with LULC can be an effective tool for evaluating environmental conditions, inhabitability and sustainability of any region.

Both in-situ and remotely sensed data are used for LST and LULC mapping. However, in-situ data collection is costly, and unavailability of historical records can be a big obstacle. Moreover, mapping land cover type of a large region is difficult to carry out with field survey. Remotely sensed data are shown to overcome this obstacle. Recently, various satellite sensors like moderate resolution imaging spectroradiometer (MODIS), advanced very high-resolution radiometer (AVHRR), thematic mapper (TM), enhanced thematic mapper plus (ETM +), operational land imager/thermal infrared sensor (OLI/TIRS) are used in LST and LULC  mapping (Cristóbal et al. 2018; Shi and Zhang 2018; Xue et al. 2019). Among them, Landsat TM, ETM + and OLI/TIRS are particularly useful due to their high resolution (compared with MODIS or AVHRR) and long-term data availability (Shi and Zhang 2018; Soydan 2020).

The relationship between LULC and LST can be explored by two approaches (Zhou and Wang 2011). First, by associating LST with LULC (Voogt and Oke 2003). Second, by establishing the relationship with different biophysical variables such as normalized difference vegetation index (NDVI) (Rouse et al. 1974), normalized difference built-up index (NDBI) (Zha et al. 2003), modified normalized difference water index (MNDWI) (Xu 2006) and bare soil index (BSI) (Rikimaru et al. 2002) with LST. Some studies have used both techniques to provide evidence of increasing LST as a result of land use/cover change (Dewan and Corner 2012; Trotter et al. 2017; Das and Angadi 2020; Tariq et al. 2020).

Though many works have illustrated the relationship between LULC and LST on different cities of Bangladesh including Chittagong city corporation area (Chaudhuri and Mishra 2016; Trotter et al. 2017; Dewan et al. 2021; Raja et al. 2021; Roy et al. 2021; Gazi et al. 2021), study utilizing multi-seasonal data is lacking. For example, most of the existing works on LST in Bangladeshi cities have dealt with a single month temperature to represent entire year (Chaudhuri and Mishra 2016; Raja et al. 2021). However, a clear and accurate representation of the whole year requires LST computation for different seasons. Also, most of the previous works have used maximum likelihood method for classification (Dewan and Yamaguchi 2009; Chaudhuri and Mishra 2016). Few, however, have used support vector machine (SVM) method, but there is no work that has used advanced artificial neural network (ANN) algorithm for LULC classification. In addition, recent studies have showed that the intensity of SUHI has become a common feature for Chittagong area with marked seasonal variations, resulting in the narrowing of diurnal temperature range (DTR) (Dewan et al. 2021). On the other hand, attributes such as population density, elevation, slope, aspect, road networks, etc. have been used to forecast LST but studies examining the association between biophysical variables and LST is few and far between.

In this study, we have addressed a research gap by detailing the impact of LULC change on land surface temperature in Chittagong Metropolitan Area (CMA). We extended our work by modeling the impact of land cover change on LST for 2050.

2 Study Area

The CMA is located between latitude 22° 6′ to 22° 3′ and longitude 91° 41′ to 92° 3′, and is bounded by the Bay of Bengal on the west, the Karnafuli River on the southwest, the Halda River on the northeast, and Rangamati district on the east (Fig. 1). It is the second largest metropolitan area of Bangladesh with a population of around 8 million people within an area of 715.16 km2 (BBS 2011). Population, which is a major factor behind rapid urban expansion, is increasing in a substantial manner due to widespread employment opportunities. Apart from population increase, urbanization is driven by economic development, commercialization, and modernization (Roy et al. 2020).

Fig. 1
figure 1

Location of the study area

Full size image

Geographically, CMA is a coastal hilly region, and the hills on the southeastern portion are known as the Chittagong Hill Tracts (CHT). Climatologically, CMA has hot summer (April to June) and relatively cold winter (November to February) due to tropical monsoon climate (Adnan et al. 2019). The maximum temperature recorded by Bangladesh Meteorological Department (BMD) is 38.9 °C in April and the lowest is 5.2 °C in January, but the average temperate remains between 24 and 28 °C (BBS 2011). Annual precipitation ranges from 2400 to 3000 mm. Chittagong is also known as the economic capital of Bangladesh and the center of all business and economic activities. It had been an important seaport of the South Asian region since the ninth century. The area is also full of natural resources, especially forest resources. However, CMA is prone to a range of disasters, including cyclones, landslides, and earthquakes (Islam et al. 2017). Also, the area is certainly missing sustainable planning aspect regarding its geographical location. Therefore, continuous monitoring of its climate and land change pattern is important for future planning of the city, and to tackle any uncertain catastrophic activities in the economic hub of Bangladesh.

3 Materials and Methods

3.1 Data and Image Pre-processing

A total of 18 images from Landsat 5 TM (1990 and 2005) and Landsat 8 OLI/TIRS (2020) sensors were acquired for three different seasons (summer, winter and autumn) from USGS (Landsat 7 Data Users Handbook 2019; Landsat 8 Data Users Handbook 2019). They were from path 136/44 and 136/45, which were then mosaiced. The data (Level 1 Terrain Corrected) were geo-referenced to UTM zone 45 (Table 1).

Table 1 Attributes of Landsat data used in this study
Full size table

Using the fast line-of-sight atmospheric analysis of spectral hypercubes (FLAASH) algorithm, atmospheric correction was done.

3.2 Methods

3.2.1 LULC Classification

Among the three seasons, images of winter for all respective years were used for LULC analysis, as the sky remains relatively clear during winter. The images were classified using ANN algorithm for its high accuracy in ENVI (Environment for Visualizing Images) software. This is a non-parametric supervised classifier which uses back-propagation learning algorithm and consists of three layers; i.e., input, hidden and output layer. During the image classification, neurons in input layer denote image bands, whereas output layer neurons represent LULC features (Srivastava et al. 2012). The network error is minimized by back-propagation till the neural network can generate continuous input–output relations with a certain level of accuracy (Gopal and Woodcock 1996). The hidden layer's learning rate was set to 100, while output layer’s was set to 0.01 with a 0.001 stopping criteria. However, false color composition (FCC) was created via reflective bands of Landsat and collected 250 training samples for 5 different land features (vegetation, agriculture, water bodies, bare land and built-up) to train the model. The function of ANN can be expressed with the following equation:

$$o_{j} = {1 \mathord{\left/ {\vphantom {1 {\left( {1 + e^{{ - \lambda {\text{net}}_{j} }} } \right) }}} \right. \kern-\nulldelimiterspace} {\left( {1 + e^{{ - \lambda {\text{net}}_{j} }} } \right) }} $$
(1)

where \({O}_{j}\) is the output of external input \(j\) and \(\lambda\) is a gain factor. The term \({\mathrm{net}}_{j}\) can be computed by (Schalkoff 1997):

$${\text{net}}_{j} = \mathop \sum \limits_{i} w_{ji} o_{i} $$
(2)

where \({w}_{ji}\) is the weight of interconnection channel to unit \(j\), from unit \(i\) and \(o\) is the output of external unit \(i\).

3.2.2 Accuracy Assessment

The best method to collect sample data for accuracy assessment is field survey. But rapid land use changes often make field work challenging. Therefore, usage of satellite images as the source of reference is also accepted in the scientific community. The most frequently used validation tool for land cover classified maps is kappa statistics (Cohenx 1960), as it gives significantly precise results than other validation methods (Foody 1992). It ranges from 0 to 1, with 0 denoting poor agreement and 1 representing nearly perfect agreement. Also, accuracy of each LULC class can be quantified by applying user’s and producer’s accuracy (Story and Congalton 1986). However, in our study, the accuracy of a LULC classification was calculated by developing a confusion or error matrix with 500 sample sites which compared the classified LULC maps with reference data. The kappa coefficient (K), overall accuracy (OA), user’s accuracy (UA), and producer’s accuracy (PA) were then derived from this error matrix.

3.3 Extraction of LST

To understand the effect of LULC changes and urbanization on temperature, LST is used. It has also a direct relation with the biophysical components. There are a number of algorithms available to compute LST from Landsat data, however, mono window algorithms (MWA) (Qin et al. 2001) and single channel algorithms (SCA) (Jiménez‐Muñoz and Sobrino 2003) require water vapor data which was not available for our study area. Alternatively, we have used an image-based method developed by USGS to calculate LST from Landsat images (USGS 2016). Band 6 and band 10 of TM and TIRS data were used to retrieve LST using the following steps:

  • Step 1 Transformation of digital number (DN) to spectral radiance (\({L}_{\lambda }\))

Every object emits electromagnetic energy as the object temperature is greater than 0 (zero) Kelvin (K) or absolute 0 (zero). Thermal sensor collects signals and can transform to sensor radiance. For TIRS (Landsat 8), spectral radiance estimation (\({L}_{\lambda }\)) was performed using the following formula (Landsat 8 Data Users Handbook 2019):

$$L_{\lambda } = M_{{\text{L}}} *Q_{{{\text{CAL}}}} + A_{{\text{L}}} $$
(3)

where \({L}_{\lambda }\) is the sensor spectral radiance, \({M}_{\mathrm{L}}\) is the band’s radiance multiplicative scaling factor, \({A}_{\mathrm{L}}\) is the band’s radiance additive scaling factor, \({Q}_{\mathrm{CAL}}\) is Quantized and calibrated radiance value. For TM/ETM + (Landsat 4,5/7), the spectral radiance required the following formula (Landsat 7 Data Users Handbook 2019):

$$L_{\lambda } = L_{\min \lambda } \left[ {{{\left( {L_{\max \lambda } - L_{\min \lambda } } \right)} \mathord{\left/ {\vphantom {{\left( {L_{\max \lambda } - L_{\min \lambda } } \right)} {\left( {Q_{{{\text{CAL}}_{\max } }} - Q_{{{\text{CAL}}_{\min } }} } \right)}}} \right. \kern-\nulldelimiterspace} {\left( {Q_{{{\text{CAL}}_{\max } }} - Q_{{{\text{CAL}}_{\min } }} } \right)}}} \right] * Q_{{{\text{CAL}}}} $$
(4)

where \({L}_{\mathrm{max}\lambda }\) and \({L}_{\mathrm{min}\lambda }\) are the maximum and minimum spectral radiance (Wm−2sr−1 µm−1). \({Q}_{{\mathrm{CAL}}_{\mathrm{max}}}\) is the maximum DN value of the image and \({Q}_{{\mathrm{CAL}}_{\mathrm{min}}}\) is the minimum DN value of the image, here 255.

  • Step 2 Transformation of spectral radiance (\({L}_{\lambda }\)) to brightness temperature (\(\mathrm{BT}\))

It is important to convert spectral radiance (\({L}_{\lambda }\)) to brightness temperature (\(\mathrm{BT}\)). The following theorem measures \(\mathrm{BT}\) in °C (Landsat 7 Data Users Handbook 2019; Landsat 8 Data Users Handbook 2019):

$${\text{BT}} = \left[ {K2/(\ln \left( {K1/L_{\lambda } ) + 1)} \right)} \right] - 273.15$$
(5)

The value for \(K1\) and \(K2\) can be retrieved from respective metadata file.

  • Step 3 Estimating land surface emissivity

Land surface emissivity is considered as a proportionality factor of Plank’s Law. This can be calculated as (Sobrino et al. 2004):

$$\epsilon = 0.004Pv + 0.986 $$
(6)

where \(\epsilon\) is the land surface emissivity and \({P}_{v}\) is the proportion of vegetation.

$$P_{v} = \left[ {{{\left( {{\text{NDVI}} - {\text{NDVI}}_{\min } } \right)} \mathord{\left/ {\vphantom {{\left( {{\text{NDVI}} - {\text{NDVI}}_{\min } } \right)} {({\text{NDVI}}_{\max } - {\text{NDVI}}_{\min } )}}} \right. \kern-\nulldelimiterspace} {({\text{NDVI}}_{\max } - {\text{NDVI}}_{\min } )}}} \right]^{2} $$
(7)

where \({\mathrm{NDVI}}_{\mathrm{max}}\) is the maximum value of \(\mathrm{NDVI}\) and \({\mathrm{NDVI}}_{\mathrm{min}}\) is the minimum.

  • Step 4 Estimating LST

This is the final stage of calculating LST. This can be computed as:

$${\text{LST}} = {\text{BT}}/[1 + \left( {\left( {\lambda {\text{BT}}\sigma /hc)*\ln \epsilon } \right)} \right] $$
(8)

where \(\mathrm{BT}\) is the brightness temperature, \(\lambda\) is the spectral radiance, \(h\) is the Plank’s Constant (6.626 × 10 − 34 J K−1), \(c\) is the velocity of light (2.998 × 108 m s−1) and \(\sigma\) is the Boltzmann constant (1.38 × 10 − 23 J K−1). While comparing LST from different time periods, especially when comparing seasonal variation of LST, it is recommended to normalize or standardizee LST values (Carlson and Traci Arthur 2000; Trotter et al. 2017). Thus, our LST values were normalized.

3.4 Extraction of Biophysical Parameters

A few biophysical indices were calculated for a better perception of the change of a specific land cover. The relationship between these indices and LST were also studied to understand the impact on LST. Nevertheless, NDVI (Rouse et al. 1974), NDBI (Zha et al. 2003), MNDWI (Xu 2006) and BSI (Rikimaru et al. 2002), which respectively, indicate greenness or vegetation, built-up or impervious surfaces, open water bodies and soil without grass or any object of any area, were calculated using the following formulas.

$${\text{NDVI}} = {{\left( {B_{{{\text{NIR}}}} - B_{{\text{R}}} } \right)} \mathord{\left/ {\vphantom {{\left( {B_{{{\text{NIR}}}} - B_{{\text{R}}} } \right)} {(B_{{{\text{NIR}}}} + B_{{\text{R}}} )}}} \right. \kern-\nulldelimiterspace} {(B_{{{\text{NIR}}}} + B_{{\text{R}}} )}}$$
(9)
$${\text{NDBI}} = {{\left( {B_{{{\text{SWIR}}1}} - B_{{{\text{NIR}}}} } \right)} \mathord{\left/ {\vphantom {{\left( {B_{{{\text{SWIR}}1}} - B_{{{\text{NIR}}}} } \right)} {\left( {B_{{{\text{SWIR}}1}} + B_{{{\text{NIR}}}} } \right)}}} \right. \kern-\nulldelimiterspace} {\left( {B_{{{\text{SWIR}}1}} + B_{{{\text{NIR}}}} } \right)}}$$
(10)
$${\text{MNDWI}} = {{\left( {B_{{\text{G}}} - B_{{{\text{MIR}}}} } \right)} \mathord{\left/ {\vphantom {{\left( {B_{{\text{G}}} - B_{{{\text{MIR}}}} } \right)} {\left( {B_{{\text{G}}} + B_{{{\text{MIR}}}} } \right)}}} \right. \kern-\nulldelimiterspace} {\left( {B_{{\text{G}}} + B_{{{\text{MIR}}}} } \right)}}$$
(11)
$${\text{BSI}} = \left[ {{{\left( {\left( {B_{{{\text{SWIR}}1}} + B_{{\text{R}}} } \right) - \left( {B_{{{\text{NIR}}}} + B_{{\text{B}}} } \right)} \right)} \mathord{\left/ {\vphantom {{\left( {\left( {B_{{{\text{SWIR}}1}} + B_{{\text{R}}} } \right) - \left( {B_{{{\text{NIR}}}} + B_{{\text{B}}} } \right)} \right)} {\left( {\left( {B_{{{\text{SWIR1}}}} + B_{{\text{R}}} } \right) + \left( {B_{{{\text{NIR}}}} + B_{{\text{B}}} } \right)} \right)}}} \right. \kern-\nulldelimiterspace} {\left( {\left( {B_{{{\text{SWIR1}}}} + B_{{\text{R}}} } \right) + \left( {B_{{{\text{NIR}}}} + B_{{\text{B}}} } \right)} \right)}}} \right] + 1 $$
(12)

where NIR is near-infrared (TM band 4 and OLI band 5), R is red band (TM band 3 and OLI band 4), SWIR1 is shortwave infrared1 (TM band 5 and OLI band 6), G indicates green band (TM band 2 and OLI band 3) and B is blue band (TM band 1 and OLI band 2).

3.5 Predicting LST

To ascertain long-term impact of LULC change on LST, it is useful to predict future scenario of LST under current LULC change. There are several methods such as artificial neural network (ANN), regression model and hybrid neural models, which are applied in various studies (Corner et al. 2014; Ghosh et al. 2019; Nurwanda and Honjo 2020). In this work, we have used ANN method to predict the LST for the year 2050, which is regarded as an effective method in predicting with preceding data (Mas and Flores 2008; Ullah et al. 2019; Al Kafy et al. 2021). A Multi-layer Feed forward back Propagation ANN method in MATLAB software was adopted to forecast LST for the year 2050. The Multi-Layer Perceptron (MLP) neural network uses autonomous calculations about how to adjust provided parameters for a better output. However, when ANN algorithm reads the patterns, it generates a random output with poor accuracy and then computes the difference between low accurate output and the desired output. The iterative cycle repeats until the network output meets the target output with an acceptable error (Silva and Clarke 2002; Ullah et al. 2019).

During model training, biophysical variables (NDVI, NDBI, MNDWI, and BSI), classified images, latitude and longitude were employed as input variables and retrieved LST data as output. The initial learning rate (μ) was set to 0.1, and the decay rate (β) was used to regulate it. The pixel values for all images were also converted to continuous data to improve the performance. Apart from network creation, training, and prediction, our prediction model also incorporates performance evaluation. Mean Square Error (MSE) and correlation coefficient (R) were used to measure reliability of the network.

4 Results and Discussion

4.1 LULC Change Detection

The spatial pattern of LULC for the years 1990, 2005 and 2020 is shown in Fig. 2. In both 1990 and 2005, vegetation was prevalent followed by bare land, agriculture, waterbodies and built-up in 1990 and bare land, agriculture, built-up and waterbodies in 2005 (Table 2). But in 2020, agricultural land has surpassed vegetation and was seen to cover almost 39% of the study area.

Fig. 2
figure 2

LULC map of CMA in 1990, 2005 and 2020

Full size image
Table 2 Area and percent of LULC covers in CMA
Full size table

As a result of excessive destruction of vegetative areas for development purpose, it has lost 10,447.51 hectares (ha) of land in the last 30 years. Bare lands have also dramatically decreased from 20,342.3 ha to 2695.6 ha from 1990 to 2020. Built-up areas and waterbodies on the other hand have increased in this 3 decades. With a maximum change rate (4.6%), built-up area has reached 17.5% of CMA in 2020 from 3.1% in 1990, and waterbodies has almost doubled between 1990 and 2020.

From Fig. 2, it can be noticed that most of the vegetation patches that were available in 1990 started disappearing in 2005 and were transformed to other LULC in 2020. Only a noticeable patch can be identified in the northern side of the city, which is basically a hilly region. Eastern and north-eastern parts also have some small patches which may also disappear soon, if present trend of land cover change continues. Again, most of the bare lands near the river or city center have turned into built-up areas and some have turned into agricultural lands in both 2005 and 2020. The city has expanded from the riverbank mainly to northern part and a small portion to southern part of CMA. However, rapid population rise is mainly responsible for this unplanned and unsustainable land cover change. In fact, to meet the need for housing and food of the growing population, vegetation and bare land have converted to either built-up or agricultural lands, and this changing trend is common in developing countries around the world, including Bangladesh (Mishra and Rai 2016; Pawe and Saikia 2018; Hatab et al. 2019; Astuti et al. 2019; Roy et al. 2021; Gazi et al. 2021).

4.2 Accuracy Assessment

Accuracy of land cover classification has showed an overall accuracy of our classified data as 86.40%, 87.20% and 88.00%, respectively, for the years 1990, 2005 and 2020. The kappa coefficient was 0.83 in 1990, 0.84 in 2005 and 0.85 in 2020. Thus, it can be said that the accuracy of our classified images has satisfied accuracy metric. The detailed result of accuracy assessment is shown in Table 3.

Table 3 Accuracy assessment of LULC classification
Full size table

4.3 Changing Trend of LST

Surface temperature of 1990, 2005 and 2020 for the winter, summer and autumn seasons were extracted. In the winter season, the lowest temperature of 1990 was 14.05 °C where the highest was 23.26 °C in 2005, the highest temperatures increased to almost 2 °C where the lowest was almost the same as 1990. But in 2020, both the highest and the lowest temperatures increased in noticeable manner, which ranges between 15.75 and 29.46 °C (Table 4). In the summer season, a slight decrease in the lowest temperature was seen in 2005 (25.16 °C) from 1990 (26.59 °C). This might not be the same for the whole season as our study has analyzed single day temperature. However, the highest temperature was 32.81 °C in 1990, 33.21 °C in 2005 and 36.53 °C in 2020. In the autumn season, lowest temperatures were almost the same for all the three years, but the highest temperature has gradually increased and ended up at 29.61 °C in 2020, which was 26.13 °C in 1990. The average temperature has increased almost 2.5 °C between 1990 and 2020 (Table 4). Overall, average temperature has increased by almost 2 °C in the last 30 years. Other studies also reported a rise of temperature in the study area (Roy et al. 2020; Dewan et al. 2021; Raja et al. 2021). The spatially distributed average temperature for the years 1990 and 2020 is displayed in Fig. 3.

Table 4 Descriptive statistics of LST
Full size table
Fig. 3
figure 3

Spatiotemporal distribution of LST

Full size image

4.4 Relationship Between LST and LULC

Figure 4 shows the range and density of temperature distribution of each land cover type. The average temperature of each class is indicated by a red dot. It can be seen that, in every class in every year, density is always high around the point of average temperature, indicating that the temperature of each class covers the majority of CMA. On the other hand, high and low temperature values of each class have very low density; thus, comprising very few numbers of pixels and covering a small portion of the study area.

Fig. 4
figure 4

LST according to land cover categories: a 1990, b 2005 and c 2020

Full size image

LST varies from one type of land cover to another type based on emission property of land features (Voogt and Oke 2003). The non-evaporative surfaces like built-up and bare lands exhibit more temperature than evaporative surfaces like waterbodies and vegetations. Correspondingly, in this work, bare lands and built-up have showed higher temperature range, whereas waterbodies, vegetation and agriculture have showed lower or moderate temperature range (Fig. 4). The average temperature of built-up and bare land had increased noticeably between 1990 and 2020. From 26.8 °C in 1990, average temperature of built-up has increased to 28.2 °C in 2005 and 31.2 °C in 2020. The average temperature of bare land had also increased to 27.6 °C in 2020 from 25.6 °C in 1990. On the other hand, vegetation and agricultural lands have showed, respectively, an average temperature of 22.5 °C and 24.8 °C in 1990, 22 °C and 25.7 °C in 2005 and 22.9 °C and 25.1 °C in 2020. However, average temperature of water bodies was almost the same for all three years. This indicates not only that built-up and bare land encompass higher temperature range but also the consistent increase in intensity of temperature in these features; conversely, features such as vegetation, waterbodies and agricultural lands keep the temperature low and almost constant.

According to Fig. 3, most of the high-temperature zones in 1990 were in the western part of the city, with some in the center. Because of the availability of water, industries and factories developed along the riverbank, high-temperature zones along the river are mostly caused by the heat generated and emitted by those industries. The northern part of the city showed lower temperature as most of the parts in the north are hilly regions. The eastern part of the city is basically the transition zone between urban and rural areas. Vegetation was also abundant in that part of CMA; thus, temperature was low. But, in 2005 and 2020, high to mid-range temperature zone engulfed almost the whole city as a result of unplanned and uncontrolled urbanization. Destruction of the vegetative cover in the northern and eastern part, which are mostly replaced by built-up area and agricultural lands, also had a significant impact on the temperature rise.

4.5 Relationship Between LST and Biophysical Components

Each biophysical index was measured for all the three seasons and averaged for that year. To identify the relationship between LST and biophysical indices, individual regression analysis was performed in the statistical software R (Figs. 5, 6, 7, 8). Furthermore, correlation between LST and NDVI, NDBI, MNDWI and BSI was also calculated for each year Fig. 9.

Fig. 5
figure 5

Correlations between LST and NDVI: a 1990, b 2005 and c 2020

Full size image
Fig. 6
figure 6

Correlations between LST and MNDWI: a 1990, b 2005 and c 2020

Full size image
Fig. 7
figure 7

Correlations between LST and NDBI: a 1990, b 2005 and c 2020

Full size image
Fig. 8
figure 8

Correlations between LST and BSI: a 1990, b 2005 and c 2020

Full size image

Figure 5 shows the slope between NDVI and LST to be downwards for both the years, which indicates that NDVI and LST are negatively correlated (− 0.78 in 1990, − 0.27 in 2005 and − 0.35 in 2020). MNDWI also shows a negative correlation (− 0.25 in 1990, − 0.16 in 2005 and − 0.43 in 2020) with LST as in Fig. 6.

Conversely, the slope of NDBI and BSI has showed an upward trend. The correlation value between LST and NDBI was 0.65, 0.48 and 0.62 in 1990, 2005 and 2020, respectively (Fig. 7). This indicates that the intensity of built-up areas or impervious surfaces is directly related to the temperature rise. The scenario is similar for BSI where the correlation between LST and BSI was 0.44, 0.11 and 0.31 in 1990, 2005 and 2020, respectively (Fig. 8). This is because the non-evaporative surfaces cannot release their latent heat through evaporation or evapotranspiration, and surface absorbs or reflects the energy from the Sun; exhibiting high temperature as a result.

However, a clear picture of the relationship between LST and NDVI, NDBI, NDWI, MNDWI and BSI is depicted in Fig. 9. In 1990, the highest correlation was seen between LST and NDBI (0.65), and the lowest between LST and NDVI (− 0.78). A significant positive correlation was also witnessed between LST and BSI (0.44); and a negative correlation between NDVI and NDBI (− 0.59), NDVI and BSI (− 0.43). In 2005, NDBI and BSI also showed the highest correlation (0.56) and BSI and MNDWI showed the lowest correlation (− 0.66). A significant positive correlation was seen between LST and NDBI (0.48) where significant negative correlation was seen between NDVI and MNDWI (− 0.47), NDBI and MNDWI (− 0.53). Again, in 2020, the highest positive correlation was seen between LST and NDBI (0.62). Moreover, NDVI and BSI (0.36), NDBI and BSI (0.39) have also showed significant positive correlation. Conversely, the lowest correlation was seen between MNDWI and BSI (− 0.85), where LST and NDVI (− 0.35), NDVI and NDBI (− 0.58), LST and MNDWI (− 0.43), NDVI and MNDWI (− 0.48) also showed strong negative correlation.

Fig. 9
figure 9

Correlation between LST and NDVI, NDBI, MNDWI, and BSI: a 1990, b 2005 and c 2020

Full size image

4.6 Predicting LST for 2050

With the help of ANN, this work has predicted the LST for 2050 (Fig. 10). The MSE and R were 0.56 and 0.82, respectively, indicating a strong positive correlation between the predicted and estimated LST for CMA. It can be seen that the average temperature will rise to 28.04 °C in 2050 from 26.70 °C in 2020. Except for some areas in the eastern and northern part of CMA, where most of the lands are covered with vegetation, water bodies or agriculture, the whole city exhibited high temperature. The temperature is most intense not only over the previous built-up areas but also over its surroundings. This can be an indication that the built-up areas may encompass majority of the areas within 2050.

Fig. 10
figure 10

Predicted temperature (°C) for the year 2050

Full size image

5 Conclusions

In this study, we have estimated land cover and land surface temperature of CMA for the years 1990, 2005 and 2020 using Landsat TM and OLI/TIRS data and explored the relationship between LULC and LST. Result has revealed that the most of the infrastructural development took place near the Karnafuli river, and LST is high in this region. Conversely, area with vegetation, especially northern hilly region, and water bodies showed low temperature. Moreover, we have measured correlation between temperature and four biophysical indices, i.e., NDVI, NDBI, MNDWI and BSI, to obtain more clear understanding on the influence of land cover change on temperature, and it was found that NDVI and MNDWI are negatively correlated with LST, whereas NDBI and BSI are positively correlated. Finally, to perceive the long-term impact of LULC on LST, we have predicted the temperature of CMA with an ANN model for 2050; the outcome was alarming. Almost the whole area is predicted to exhibit higher temperatures except for some small patches on northern and eastern part of CMA.

The current urbanization pattern of CMA is unplanned and unsustainable, which is already affecting the environment in several ways including rise of LST. Although, the temperature is still in a tolerable limit, according to our prediction, it will rise tremendously by 2050. Therefore, immediate measures should be taken to obstruct the current trend of land cover change by establishing a strong and sustainable city plan. It is also important to promote public awareness so that inhabitants do not destroy natural resources like vegetation and refrain from filling up water bodies for agricultural or construction purposes. Finally, CDA (Chittagong Development Authority) and other relevant organizations should come forward with the aim of a sustainable model city and take proper measures accordingly.