Introduction

Due to natural and human activities, the Earth's surface undergoes constant change (Grimm et al., 2008; John et al., 2020). Land surface changes generally fall into two categories: land use and land cover (Barnsley et al., 2001). According to Lambin and Ehrlich (1997), three main factors influence land use and land cover (LULC) change: institutional-political factors, biophysical factors, and technical and economic concerns. Nowadays, satellite images are commonly used to monitor spatial–temporal changes in LULC. Consequently, the causes and effects of changes related to human activity can be assessed (El-Zeiny & Effat, 2017). Remote sensing has numerous applications (Karimi et al., 2017a; b; Rangzan et al., 2022), including the discovery of processes altered between different periods, which forms the core of change detection studies, and the detection of changes not discernible by ground observations (Lakra and Sharma, 2019; Imran et al., 2021). Remote sensing is particularly useful for monitoring and evaluating LULC changes, especially in areas significantly impacted by human activities. Satellite images with moderate spatial resolution, such as Landsat data (Williams et al., 2006), are commonly employed for observing and assessing LULC changes. Numerous studies, such as Lewinski (2006), Al Fugara et al. (2009), and Pal and Ziaul (2017), have investigated LULC changes using Landsat images.

The modification of Land Surface Temperature (LST) in an urban setting is one of the major effects of Land Use and Land Cover (LULC) change (Dhar et al., 2019; Gohain et al., 2021). Over the past decade, global environmental processes have been studied to understand climate change, utilizing LST as an excellent indicator to monitor the physical properties of surface processes and climate change. For this reason, LST is an important input for estimating the energy balance equation (Filgueiras et al., 2019). It has been concluded that the main sources of human heat emission in urban centers correspond to certain types of land use-land cover. Therefore, the ratio of different types of land use-land cover may significantly affect LST, especially the ratio of industrial and commercial areas. Thus, urban and industrial development emerges as one of the key contributors to the rise in temperatures. Consequently, the following discussion will focus on the research conducted on LST and land use change.

Yoo et al. (2019) utilized Convolutional Neural Network (CNN) and Random Forest (RF) methods with Landsat images to classify urban areas. Their study revealed that the CNN classifier achieved higher accuracy in class separation compared to RF (Yoo et al., 2019). In 2020, Soleimani et al. investigated the effects of land use changes on the temporal and spatial patterns of Land Surface Temperature (LST) and thermal islands in Saqqez city. In this research, they employed Landsat satellite images from 1989 to 2018. For classification, they utilized the maximum likelihood method, and subsequently, to extract temperature data, they applied the split window algorithm. Their results indicated an increase in thermal islands in the northeast of the city from 2008 to 2018. Moreover, the highest temperatures were observed in vegetation, residential, and wasteland areas. Overall, their findings demonstrated the direct effects of land use on temperature rise (Soleimani et al., 2018).

Das et al. (2021) investigated the effect of Land Use and Land Cover (LULC) on Land Surface Temperature (LST) in the Sansol region of India. In their study, they considered data from 1993 and 2018 to prepare the LULC and LST maps. The Kappa coefficient was employed to evaluate the accuracy of the LULC maps. The LST maps revealed an increase in temperature by 0.15 °C and 0.19 °C per year, respectively, during summer and winter. The temperature rise was primarily attributed to urbanization, commercial activities, and coal mining regions. According to changes in the LULC pattern, urban areas expanded by 60%, while coal mining regions increased by 15%. The relationship between LST and various spatial indices such as the Normalized Difference Built-up Index (NDBI), Normalized Difference Vegetation Index (NDVI), and Normalized Difference Water Index (NDWI) was demonstrated through several correlations. Their findings indicated a negative association between LST and NDVI as well as LST and NDWI, whereas LST and NDBI showed a positive correlation. Finally, the simulation of temperature for the year 2041 suggested a potential rise of 0.21 °C per year in the forthcoming years (Das et al., 2021).

Although various classification algorithms have been applied in remote sensing studies, achieving high accuracy in detecting Land Use-Land Cover (LULC) changes remains a challenge. Previous studies have used various classification methods, such as Convolutional Neural Networks (CNNs) and Random Forest (RF), to improve the accuracy of LULC classification, revealing the superior performance of CNNs. The fully convolutional neural network approach and the Res-UNet model are used in our study to significantly improve the classification accuracy. This highlights the potential of deep learning to enhance change detection accuracy, addressing a critical research gap in the field. The purpose of this study is to evaluate and monitor Land Use and Land Cover (LULC) changes and their relation to Land Surface Temperature (LST) changes by integrating remote sensing, GIS, and deep learning techniques over the past 20 years (2001–2021) in Izeh city.

Study Area

Izeh city, covering an area of 4035 km2, is situated in the northeastern part of Khuzestan province in southwestern Iran. The average annual rainfall is 670 mm, and the average daily temperature is 21 °C over a 20-year period (2001–2021). The study area extends spatially from 31° 26′ 0″ N to 32° 22′ 0″ N and 49° 32′ 0″ E to 50° 28′ 0″ E (Fig. 1). Over this twenty-year period, the highest recorded air temperature in July was 48 °C, while the lowest temperature in February reached 9 °C. The highest point in the region stands at 3589 m, and the lowest is 348 m above sea level. The climate of the Izeh plain is characterized as semi-humid and moderate. The plain is predominantly surrounded by limestone highlands (Asmari formation) and is geologically classified as an open karst plain, with two lakes named Miangaran and Bandan (Moradi et al., 2020) present. The border of the Izeh zone corresponds to the fronts of the Balaroud and Kazeroun mountains, situated across a distinct topographic gap in the southwest of the Zagros fault (Asadi Mehmandosti et al., 2013).

Fig. 1
figure 1

Geographical location of the study area: Iran (a), Khouzestan (b), and the Digital Elevation Model (DEM) of Izeh (c)

Full size image

Materials and Methods

Materials/Datasets

Satellite Images

The present study utilized four Landsat images (path/row: 038/164 and 038/165) obtained for two distinct years from the Earth Explorer website of the USGS. Two images were selected from Landsat-7 ETM+ (for 2001) and two from Landsat-8 OLI/TIRS (for 2021), specifically for the months of August and September, considering minimal cloud cover (Amran et al., 2018). The details of the employed Landsat-8 images are provided in Table 1.

Table 1 Characteristics of the applied Landsat images
Full size table

Methodology

Image Pre-processing

Assuming that the spectral properties of non-changed areas remain stable, preprocessing is crucial in change detection studies. Inadequate preprocessing can lead to false change detection in the spectral space, increasing the risk of error (Wulder et al., 2006; Cooley et al., 2002). Prior to image processing, the preprocessing steps—radiometric control and image enhancement—were conducted (Aslami & Ghorbani, 2018).

An atmospheric correction tool called Fast Line-of-sight Atmospheric Analysis of Spectral Hypercube (FLAASH) is utilized to adjust remote sensor data in the 400 ± 3000 nm range (Jensen & Lulla, 1987). FLAASH employs MODTRAN simulations to generate spectral radiance data under different atmospheric, water vapor, and viewing conditions (solar angles) across various surface reflectances. These data are then utilized to create lookup tables for atmospheric parameters such as column water vapor, aerosol type, and visibility, which can be referenced for future analyses (Kruse, 2004; Adler-Golden et al., 2005; Pordel et al., 2019).

Land Use-Land Cover Map (LULC) Classification

In this study, two methods were employed for pixel-base classification: (a) The Maximum Likelihood Classifier (ML) and (b) the object-based Fully Convolutional Network (FCN), to classify the land use and land cover of the study area.

(a) Max Likelihood Classifier

Image classification categorizes pixels into different classes automatically (Lillesand et al., 2003). These classes can include urban, vegetation, water, and wasteland. Pixels are identified based on their spectral signatures, which reflect the relative reflectance of the area in various bands (Sabins, 1997). Among supervised classifiers, the Maximum Likelihood (ML) classifier is one of the most significant and accurate methods.

In this study, the LULC map was classified using the ML method in the ENVI software, which yielded the highest accuracy among the supervised classifiers (Lillesand et al., 2003). Training samples are required for each of these user-defined classes. The probability distribution of each class across the image is computed using the class means and covariances. Subsequently, each pixel is assigned to one of the classes based on its probability (Ayanlade & Howard, 2019).

According to the USGS definition, Land use-Land cover classes in the region include:

  • Water streams, canals, lakes, reservoirs, bays, or oceans.

  • Wetland mosaics of water, bare soil, and herbaceous or wooded vegetated cover.

  • Grassland shrubs and perennial or annual natural and domesticated grasses (e.g., pasture), forbs, or other forms of herbaceous vegetation at least 10% of the area and tree cover is less than 10% of the area Pasture.

  • Forest land spanning more than 0.5 hectares with trees higher than 5 m and a canopy cover of more than 10 percent.

  • Wasteland natural occurrences of soils, sand, or rocks where less than 10% of the area is vegetated.

  • Urban high-density residential, commercial, industrial, mining, or transportation.

We employed a random sampling method. The number of samples collected or selected for accuracy evaluation was calculated according to Professor Jensen's formula from the following Eq. (1):

$$\text{N}=\frac{{Z}^{2}Pq}{{E}^{2}} =\frac{{2}^{2}\left(85\right)\left(15\right)}{{5}^{2}}=204$$
(1)

where Z = 2, P is the required accuracy, q = 100-P and E is the acceptable error percentage. Therefore, the number of samples needed to determine the correct accuracy of the desired map was 204.

(b) Fully Convolutional Network (FCN) Structure

Numerous studies in machine learning, deep learning, and artificial intelligence have directed their attention toward diverse subjects, encompassing intelligent cities, weather forecasts, and change detection analysis (Elik and Gaziolu, 2020). Notably, Deep Learning has emerged as the predominant trend in image analysis, surpassing individual performance in challenging tasks (Torres et al., 2021).

CNN, a specific type of artificial neural network, is composed of convolutional layers, pooling layers, and fully connected layers (Yoo et al., 2019). The CNN approach has been widely successful in image classification. By incorporating fully connected layers, CNN can ascertain subsequent classification probability information. However, its application is limited to entire image classification, lacking pixel-level classification (Dai et al., 2016; Liu et al., 2020). Consequently, the Fully Convolutional Network (FCN) classification method was developed. This method transforms fully connected layers into convolutional layers, facilitating the creation of a classification network (Liu et al., 2020; Wu et al., 2021).

The FCN mainly consists of convolutional layers, an integration layer, and deconvolutional layers at its core. By semantically segmenting the data of the entire image, the FCN can classify pixels on a pixel-by-pixel level, significantly enhancing the algorithm's computing efficiency and accuracy (Wu et al., 2021). The FCN is utilized in the novel design of CNN models, which are configured as complete convolutional networks, optimizing the generation of proportionally sized outputs. This technique finds applications in tasks like edge detection (Xie and Tu, 2015; Ozturk et al., 2020; Wu et al., 2021), image classification (Yoo et al., 2019; Torres et al., 2021; Ghorbanzadeh et al., 2022), and more. An FCN can work with inputs of varying sizes, producing outputs with dimensions that match (potentially resampled) spatial dimensions. Furthermore, feedforward computation and backpropagation are considerably more effective when applied independently patch-by-patch across the entire image, particularly when receptive fields significantly overlap (Ozturk et al., 2020).

In this study, the Keras open-source library was employed to implement the FCN. Various methods can be explored to construct the FCN architecture, necessitating the identification of an optimal model that suits the data's characteristics. To evaluate the image classes, an object-based image analysis (OBIA) approach was adopted, utilizing the probabilities derived from the ResU-Net model with a 50-backbone and Adam optimizer set at a learning rate of 0.001 (Ghorbanzadeh et al., 2022).

Utilizing open-source ALOS DEM data and Landsat 8 images from 2001 and 2021 for both training and testing the FCN model, we trained the FCN model using these images. The network architecture of the ResU-Net that we used consists of a total of 15 convolutional layers. We employed the binary cross entropy loss function to determine the difference between each of the highest probability 1. The applied patch size is 64*64 and the bath size is 128. For data augmentation of the training sample patches, we utilized horizontal and vertical data flips. The number of image patches became 6000, of which 4000 were selected as training data and 2000 as test data. Additionally, a rule-based OBIA approach was devised during the object-based classification stage, incorporating land use-land cover classes.

Evaluation Method of LULC

(a) Assessing the Accuracy of the ML Method.

In this study, accuracy is evaluated using the error matrix which contains information about actual and anticipated pixel identifications (Jupp, 1989; Pal & Ziaul, 2017). Equation 2 is used to calculate the overall accuracy.

$$\text{T}=\frac{\sum {\text{D}}_{\text{ii}}}{\text{N}}$$
(2)

In this formula, T represents the overall accuracy, ∑Dii signifies the number of pixels that are correctly classified, and N is the total number of pixels in the error matrix. Equation 3 is used to calculate the producer's accuracy, and the user's accuracy is calculated using Eq. 4.

$$\text{PA}=\frac{\sum {\text{D}}_{\text{ij}}}{{\text{R}}_{\text{i}}}$$
(3)

where ∑Dij represents the number of pixels in a row I that are correctly classified, Ri stands for the total number of pixels in a row i, and PA represents the producer's accuracy.

$$\text{UA}=\frac{\sum {\text{D}}_{\text{ij}}}{{\text{C}}_{\text{j}}}$$
(4)

where ∑Dij signifies the number of pixels in column j that are correctly classified, Cj represents the total number of pixels in column j, and UA represents the user's accuracy.

Another accuracy coefficient, known as the kappa coefficient (Foody, 1992), was used in this research. The Kappa coefficient's value ranges between 0 and 1, where 0 indicates weak agreement, and 1 indicates almost complete agreement (Landis & Koch, 1977) (Fig. 2).

Fig. 2
figure 2

Flowchart of the current study

Full size image

(b) Assessing the Accuracy of the FCN Method

Common metrics for image classification tasks include accuracy, precision, recall, F1 score, confusion matrix, and area under the receiver operating characteristic curve (AUC-ROC). In this study accuracy, precision, and F1 score criteria were used to evaluate the accuracy of the FCN method and the theRes-UNet model.

Accuracy measures the proportion of correctly classified instances compared to the total instances. It's computed by dividing the number of correct predictions by the total number of predictions. While accuracy offers a comprehensive performance assessment, it may not suffice for datasets with imbalanced classes (Useya & Chen, 2018). Precision evaluates the ratio of true positive predictions among all positive predictions, calculated by dividing true positives by the sum of true positives and false positives. Precision is valuable when false positives are costly (Theres & Selvakunar, 2022). The F1 Score, a harmonic mean of precision and recall, balances both metrics and proves beneficial when considering false positives and false negatives. It's calculated as 2 * ((precision * recall)/(precision + recall)) (Chakhar et al., 2020).

Change Detection Method

The LULC classification for 2001 and 2021 was compared utilizing the change matrix, following the methodology proposed by Weng et al. (2004), within the ArcGIS 10.7 software. To detect changes, a combination of qualitative and quantitative techniques was employed. After identifying the change matrix, a map of changes was generated in the ArcGIS environment using spatial analysis tools.

LST Estimation Methods from the Thermal Band

Temperature data is collected by Landsat sensors (ETM + and OLI) and stored as digital numbers (DN) within a range of 0–255. For meaningful comparisons, all included images were captured during the same season and nearly simultaneously (Coll et al., 2010). Land Surface Temperature (LST) is derived using the thermal bands of Landsat 7 ETM + (band 6) and Landsat 8 OLI (band 10). However, the LST extraction process from Landsat ETM + and Landsat OLI calculates spectral radiance (Lλ) in distinct ways (Asgarian et al., 2014; Nguemhe Fils et al., 2017). The steps for extracting LST from Landsat images are outlined below (Govind & Ramesh, 2019).

Step 1: Convert DN to Spectral Radiance (Lλ).

The spectral radiance of the upper atmosphere (Lλ) is calculated using Eq. 5, which involves ETM + band six for Landsat 7.

$$L\lambda =LMIN\lambda +\left[\frac{LMAX\lambda -LMIN\lambda }{QCALMAX\lambda -QCALMIN\lambda }\times QCAL\right]$$
(5)

For each pixel, QCALMIN is set to 1, while QCALMAX is set to 255. The spectral radiance for band six is denoted by LMAXλ, with a value of 15.303, and LMINλ, with a value of 1.238. QCAL represents the numerical value of each pixel.

On Landsat 8, the thermal band is designated as OLI band 10, and the extraction of spectral radiance (Lλ) is accomplished using Eq. (6).

$${\text{L}}{{\uplambda }} = {\text{ML}} \times {\text{QCAL}} + {\text{AL}}$$
(6)

Here, Lλ represents the upper spectral radiance of the atmosphere, ML stands for the band-specific multiplicative scaling factor (0.0003342), AL is the band-specific incremental scaling factor (0.1), and QCAL signifies the pixel value of the standard quantized and calibrated product (Das et al., 2021).

Step 2: Converting Spectral Radiance to Brightness Temperature (BT)

Brightness temperature, also known as apparent temperature, corresponds to the temperature of the blackbody that produces the radiation captured by the sensor. It is also the temperature received by satellites. It's important to note that brightness temperature doesn't directly represent the actual temperature of the Earth; instead, it reflects the temperature of the satellite itself. Consequently, the data from Landsat's thermal bands can be transformed from spectral radiance to brightness temperature (Eq. 7). This conversion is achieved using the thermal constants provided within the metadata file (Sherafati et al., 2018; Ibrahim and Mallouh, 2018).

$$\text{BT}=\frac{{\text{K}}_{2}}{\text{ln}\left[\left(\frac{{\text{K}}_{1}}{{\text{L}}_{\uplambda }}+1\right)\right]}$$
(7)

In this context, BT denotes the brightness temperature of the satellite in Kelvin, while Lλ represents the spectral radiance recorded by the sensor (W/m2·sr·µm). K1 stands for the constant coefficient of the first calibration (W/m2·sr·µm), and K2 represents the constant coefficient of the second calibration (W/m2·sr·µm).

Step 3: Emissivity Index (ε)

The method proposed by James and Sobrino was employed to calculate emissivity. Emissivity is derived through thresholding the NDVI index (Eq. 8).

$$\text{PV}={\left[\frac{\text{NDVI}-{\text{NDVI}}_{\text{min}}}{{\text{NDVI}}_{\text{max}}-{\text{NDVI}}_{\text{min}}}\right]}^{2}$$
(8)

The emissivity index (ε) is obtained using the following Eq. (9) (Das et al., 2021).

$$\text{Land Surface Emissivity }(\upvarepsilon )=0.004\times \text{PV}+0.986$$
(9)

Step 4: Calculate the Land Surface Temperature

In this step, we determine the Atmospheric Water Vapor Index (AWVI) and the temperature of the Earth's surface using the single-channel algorithm.

For estimating the Land Surface Temperature (LST) using the single-channel method, it's crucial to ascertain the atmospheric water vapor content during the satellite's passage. The atmospheric water vapor was determined using the following Eq. (10), which incorporates meteorological information to derive the relative humidity (Nasseri, 2019).

$${\upomega {\text{i}}}=0.0981\times \left\{10\times 0.6801\times \text{exp}\left[\frac{17.27\times \left({\text{T}}_{0 }-273.15\right)}{273.15+\left({\text{T}}_{0}-273.15\right)}\right]\times \text{RH}\right\}+0.167$$
(10)

In this context, T0 represents the temperature of the air near the Earth's surface, RH stands for the relative humidity of the air, and ωi signifies the atmospheric water vapor content.

The single-channel algorithm employs thermal infrared bands to extract the Land Surface Temperature (LST). This approach is applicable to sensors equipped with a thermal band (bands 10 and 11) (Chatterjee et al., 2017).

$${T}_{s}=\upgamma \left[{\varepsilon }^{-1}\left({\varphi }_{1}{L}_{sensor}+{{\varphi }}_{2}\right)+{{\varphi }}_{3}\right]+\updelta$$
(11)
$$\upgamma =\left\{\frac{{C}_{2}{\text{L}}_{\text{sensor}}}{{{T}_{sensor}}^{2}}\left[\frac{{\uplambda }^{4}}{{\text{C}}_{1}} {\text{L}}_{\text{sensor}}+{\uplambda }^{-1}\right]\right\}^{-1}$$
(12)

Here, TS signifies the Land Surface Temperature (LST), Tsensor stands for the sensor's brightness temperature in Kelvin, ε represents surface emissivity, γ denotes the effective wavelength of a thermal infrared band, while γ and δ are parameters tied to the Planck function. Additionally, φ1, φ2, and φ3 are atmospheric correction parameters (Eqs. 11 and 12). By applying these parameters, the atmospheric influence is significantly mitigated or adjusted, a calculation achieved through the utilization of the following Eqs. (13, 14, 15, 16, and 17). It's important to note that all parameters are wavelength-dependent (Munoz et al., 2009). The value of C1 is 1.19104 × 10^8 W·m^-2.sr^-1.μm^4, and the value of C2 is 14,387.7 μm.K.

$$\left[ {\begin{array}{*{20}c} {{{\varphi }}_{1} } \\ {{{\varphi }}_{2} } \\ {{{\varphi }}_{3} } \\ \end{array} } \right] = \left[ {\begin{array}{*{20}c} {{\text{C}}11} & {{\text{C}}12} & {{\text{C}}13} \\ {{\text{C}}21} & {{\text{C}}22} & {{\text{C}}23} \\ {{\text{C}}31} & {{\text{C}}32} & {{\text{C}}33} \\ \end{array} } \right]\left[ {\begin{array}{*{20}c} {{\upomega }^{2} } \\ \omega \\ 1 \\ \end{array} } \right]$$
(13)

In the equation mentioned above, the C coefficients are acquired through simulation.

$${\varphi }=-\upgamma {\text{L}}_{\text{sensor}}+{\text{T}}_{\text{sensor}}$$
(14)
$${{\varphi }}_{1}=0.14714{\upomega }^{2}-0.15583\upomega +1.1234$$
(15)
$${{\varphi }}_{2}=-1.1836{\omega }^{2}-0.3760\omega -0.52894$$
(16)
$${\varphi }_{3}=-0.04554{\omega }^{2}+1.8719\omega -0.39071$$
(17)

Calculating Spatial Indices

(a) Normalized Differential Vegetation Index (NDVI)

The Normalized Difference Vegetation Index (NDVI) is a numerical indicator that utilizes the visible and near-infrared bands of the electromagnetic spectrum. It evaluates whether an observed target contains living vegetation. NDVI values range from − 1 to + 1. As values approach + 1, they indicate a higher presence of vegetation, while values associated with water and clouds are generally below zero. This is calculated using the following Eqs. (18) (Kayet et al., 2016).

$$\text{NDVI}=\frac{\text{NIR}-\text{Red}}{\text{NIR}+\text{Red}}$$
(18)

NIR indicates the extent of reflection in the infrared band, while Red indicates the degree of reflection in the red band.

(b) Normalized Differential Water Index (NDWI)

NDWI is an additional index utilized to demarcate open water boundaries and identify them using remote sensing data based on near-infrared and visible radiation. The following Eq. (19) can be employed to calculate NDWI (McFeeters, 1996):

$$\text{NDWI}=\frac{\text{Green}-\text{NIR}}{\text{Green}+\text{NIR}}$$
(19)

Green refers to band 2 for Landsat 7 (ETM+) images and band 3 for Landsat 8 (OLI) images. Near-infrared (NIR) corresponds to band 4 for Landsat ETM+ and band 5 for Landsat OLI images. Smaller values, including negative values, indicate the presence of vegetation, whereas NDWI values exceeding 0.5 signify water bodies. Thus, water bodies can be readily distinguished from vegetation. Values ranging from 0 to 0.2 are typically associated with human-made areas (Das et al., 2021).

(c) Normalized Differential Build-up Index (NDBI)

By utilizing the mid-infrared (MIR) and near-infrared (NIR) bands, we can calculate this index using remote sensing data. The equation provided below (20) was employed to compute the Normalized Difference Built-Up Index (NDBI) (Zha et al., 2003).

$$\text{NDBI}=\frac{\text{MIR}-\text{NIR}}{\text{MIR}+\text{NIR}}$$
(20)

Here, MIR refers to the mid-infrared band (band 5 for Landsat ETM+ and band 6 for Landsat OLI), while NIR corresponds to the near-infrared band (band 4 for Landsat ETM+ and band 5 for Landsat OLI). NDBI values span from − 1 to + 1, where the range of 0 to 1 is associated with urban areas. A value around 1 indicates a dense concentration of built-up regions (Choudhury et al., 2019).

Evaluation of the Relationship Between the LST and Spatial Indicators

To understand the influence of different spatial characteristics on Land Surface Temperature (LST), Pearson's correlation function within the SPSS statistical package was employed. In this study, raster correlations, such as those between LST and NDVI, LST and NDWI, etc., were visualized using Saga software (as detailed in the results and discussion section).

Results and Discussion

Land Use-Land Cover Results

Results of the ML Classifier

The LULC maps of the study area were classified into six classes: water, urban, wasteland, grassland, wetland, and forest. Table 2 presents the area of each LULC class for both 2001 and 2021. To assess the accuracy of the LULC maps, ground observation points were collected with the assistance of Google Earth. The accuracy of the LULC maps was also analyzed using the Kappa coefficient. The Kappa coefficient values were 0.77 for 2001 and 0.88 for 2021 (Table 3), indicating both LULC classifications were achieved with acceptable accuracy.

Table 2 The area of each LULC (Km2)
Full size table
Table 3 Assessment of the accuracy of LULC in 2001 and 2021
Full size table

In 2001, the area of the water class was 33,201.64 km2 (Fig. 3a and b). However, by 2021, it had expanded to 35,851.13 km2 due to the construction of the Karun 3 dam and the formation of its associated lake. Observing the distribution of settlements, it's evident that in 2001, they were primarily concentrated in the central core of Izeh city. However, by 2021, settlements had expanded, leading to urban and rural area development (Fig. 5).

Fig. 3
figure 3

LULC map of 2001 (a) and 2021 (b) of Izeh city

Full size image

The area of wastelands in 2001 measured 219,369.05 km2, which slightly decreased to 238,147.04 km2 by 2021. Wetlands in the region also exhibited a decreasing trend, declining from 2499.78 km2 in 2001 to 2377.04 km2 in 2021. Over the past two decades, areas covered by dense vegetation, such as forests, have decreased from 69,302.65 to 67,175.09 km2 within the study area (Table 2, Figs. 5 and 6).

Results of the FCN Model

Subsequently, the land use detection maps generated through the integrated approach and the FCN model were validated using precision, recall, and standard score performance measures.

Table 4 shows the evaluation of the FCN method for the classification of regional images. As can be seen in Table, the highest resulting accuracy scores for the LULC map of 2001 and 2021 were 93% and 98%, respectively. These values were obtained using a window size of 128 × 128 for generating sample patches.

Table 4 Accuracy assessment of ResU-Net method for classification Images of the study area
Full size table

F1-score was used to evaluate the classification accuracy of the FCN method. This criterion is calculated based on the precision and recall of the classifier, where precision is the ratio of the number of true positive samples to the total number of samples predicted positive, and recall is the ratio of the number of true positive samples to the total number of positive samples. The highest possible value for F1-score is 1 and the lowest possible value for this criterion is 0. F1-score was calculated for each class showing that the number of positive samples correctly recovered is close to 1.

As a result, utilizing the FCN method and the theRes-UNet model yields higher accuracy compared to pixel-based methods such as Max-Likelihood. This supervised learning model can effectively differentiate between the characteristics of the six classes of the region and accurately separate (Fig. 4).

Fig. 4
figure 4

The LULC map of 2001 (a) and 2021 (b) of Izeh city using the FCN approach

Full size image

Change Detection of LULC

Table 5 illustrates the positive and negative changes within the six classes of the study area's LULC. Water bodies, urban areas, and wastelands have experienced positive changes, while forests, wetlands, and grasslands have undergone negative changes (Table 5). A positive change signifies an increase in the area of LULC, whereas a negative change denotes a decrease.

Table 5 Percentage changes of LULC from 2001 to 2021
Full size table

Residential areas have experienced a 4.28% increase due to urban expansion, resulting in a reduction in agricultural and wasteland areas (Fig. 6). Notably, the wetlands area, including Miangaran and Bandan wetlands, has decreased by 4.91% (Figs. 5, 6; Table 5). A graphical representation of the changes in each LULC class is provided in Fig. 6.

Fig. 5
figure 5

Map of LULC changes in Izeh city from 2001 to 2021

Full size image
Fig. 6
figure 6

The graph of changes in each LULC class from 2001 to 2021 in the study area

Full size image

Land Surface Temperature Changes

Land Surface Temperature (LST) maps were extracted for 2001 and 2021 (Fig. 7a and b). In 2001, the recorded LST ranged from a high of 58 °C to a low of 18 °C. In contrast, for 2021, there was a decrease in the maximum temperature and an increase in the minimum temperature. Specifically, the maximum temperature reached 55 °C, while the minimum temperature rose to 20 °C. Consequently, it was observed that the maximum temperature decreased by 3 °C, while the minimum temperature increased by 2 °C (Table 6).

Fig. 7
figure 7

LST map of the years 2001 (a) and 2021 (b) of Izeh city

Full size image
Table 6 Statistics of changes in the LST from 2001 to 2021
Full size table

Distinct patterns of Land Surface Temperature (LST) are closely linked to the thermal characteristics of different Land Use and Land Cover (LULC) classes (Weng Q., 2004). To comprehend how LULC influences LST, thermal values were obtained for each land use category. Figure 8 displays the temporal and spatial variations of LST for Izeh city.

Fig. 8
figure 8

Map of LST changes in the study area from 2001 to 2021 using Landsat8 images

Full size image

The minimum average LST levels were observed for water bodies, wetlands, forests and grasslands (in 2001: 32.02 °C for water bodies, 33.01 °C for wetlands, 33.85 °C for forests and 34.92 for grasslands,and in 2021: 30.11 °C for water bodies, 31.49 °C for wetlands, 33.06 °C for forests and 34.6 for grasslands).

Conversely, the maximum average LST levels were recorded for wastelands and urban areas (in 2001 37.54 °C for wastelands, and 36.53 °C for urban areas, and in 2021, 44.37 °C for wastelands, and 40.11 °C urban areas (Figs. 8, 9 and Table 7).

Fig. 9
figure 9

The graph of the average temperature changes of the land surface for the LULC classes

Full size image
Table 7 Average LST changes (°C) for each LULC class
Full size table

Figure 9 illustrates the average LST changes for each land use class over the twenty-year period, following this order: wasteland > urban > grassland > forest > wetland > water (as presented in Table 7, Figs. 8, and 9). The increase in LST between 2001 and 2021 can be attributed to urban development and wasteland expansion in Izeh city. Additionally, the analyses indicate that water bodies exhibit lower LST levels.

Interestingly, despite the high recorded LST being associated with forests, the digital elevation model (DEM) of the study area reveals that the forested regions are situated at higher altitudes. Moreover, observations suggest that the influence of altitude on LST outweighs that of vegetation (Aguilar-Lome, 2019). Consequently, the high LST observed in forested areas can be attributed to their elevated altitude within the study area.

The Influence of Water Bodies, Vegetation, and Urban Areas on LST

An analysis of correlations between Land Surface Temperature (LST) and factors such as vegetation, water bodies, and urban areas was conducted. The correlation between these factors and indices such as NDVI, NDWI, and NDBI was explored. Spatial distributions of NDVI in the study area are shown in Fig. 10a and b. Darker green regions indicate dense vegetation, while the purple color represents water bodies. In 2001, the northwestern and southwestern parts of Izeh city exhibited dense vegetation, while the western part displayed minimal vegetation due to urban and desert expansion. This relationship between NDVI and NDBI becomes evident.

Fig. 10
figure 10

Vegetation index map of the study area in 2001 (a) and 2021 (b)

Full size image

Figure 11 depicts a weak negative association between NDBI and NDVI (correlation coefficient of -0.47), indicating that urban expansion leads to reduced vegetation cover. Both NDVI and LST show a negative correlation for 2001 and 2021, as displayed in Fig. 12a and b. This correlation is due to abundant vegetation preventing higher surface temperatures.

Fig. 11
figure 11

Relation between NDBI and NDVI index

Full size image
Fig. 12
figure 12

Relation between LST and NDVI indices in 2001 (a) and 2021 (b)

Full size image

NDWI is another significant index negatively correlated with LST (correlation coefficient of -0.037, Table 8), primarily because water possesses a relatively high specific heat capacity (Moldoveanu & Minea, 2019) (Fig. 13).

Table 8 Correlation matrix between LST and other spatial indices for the year 2001
Full size table
Fig. 13
figure 13

Relation between LST and NDWI indices in 2001 (a) and 2021 (b)

Full size image

A noteworthy relationship emerges between NDBI and NDWI. While NDBI negatively impacts NDWI, suggesting a decline in water storage with increasing urban areas, the relationship shifts to a positive correlation as water areas expand, and dams are constructed (Fig. 14).

Fig. 14
figure 14

Relation between NDWI and NDBI index

Full size image

Additionally, NDBI significantly affects LST. Previous studies have demonstrated a strong linear relationship between LST and NDBI (Sun et al., 2012; Tariq et al., 2022). The correlation value between them is 0.322 for 2001 and 0.208 for 2021 (Tables 8 and 9), indicating a positive association between LST and NDBI (Fig. 15a and b). Given the urban development experienced by Izeh city over;the past two decades, an increase in LST with urbanization is anticipated.

Table 9 Correlation matrix between LST and other spatial indices for the year 2021
Full size table
Fig. 15
figure 15

Relation between LST and NDBI indices in 2001 (a) and 2021 (b)

Full size image

Conclusion

In this research, an attempt was made to determine the trend of LST changes on LULC in Izeh city. For LULC classification, two pixel-based and object-based methods were used. The ResU-Net model with the FCN approach with Landsat images had higher accuracy compared to the max-likelihood method despite covering a large area. Comprehensive knowledge and monitoring of land use-land cover, as well as multi-view analysis of the impact of LULC on the thermal environment, help achieve a deeper understanding of the effective mechanisms in increasing the LST in the study area. The trend of LST change in Izeh city shows that the minimum temperature has increased by about 2 °C, while the maximum temperature has decreased by 3 °C per year over 20 years. Spatial and temporal surveys of LST showed hot regions (areas with the highest LST) in the city and wastelands. Therefore, locating these points is essential for studies on sustainable development and environmental monitoring. By comparing the correlation between land use-land cover indices (NDVI, NDWI, and NDBI) and LST under different combinations, it was found that urban areas have a positive and significant correlation with LST. This indicates the effect of urban areas on the increase in LST in Izeh city. Industrial and commercial areas, as well as traffic, significantly influence the degree of human heat emission. Therefore, more trees and parks with thick vegetation should be established in urban areas, and more plants should be planted there.

Gao et al. (2019) utilized the FCN method to classify land cover in mountainous areas, achieving a classification accuracy of 90.6%. In a separate study, Chakhar et al. (2020) employed various classification algorithms alongside Landsat 8 and Sentinel 2 imagery for crop classification. They experimented with decision trees, diagnostic analysis, support vector machines, nearest neighbors, and group classifiers, yet none of these methods yielded an F1 score exceeding 90%. However, in our current research, the accuracy of classification using the fully convolutional neural network approach and the theRes-UNet model surged to 98%.

This research faced limitations as a result of utilizing Landsat 8 images for examining changes detection within the area. Each pixel's spatial resolution was 30 m, indicating the potential benefits of employing imagery featuring a greater spatial resolution, such as Sentinel-2, to enhance the analysis. Furthermore, the temporal resolution could be heightened by incorporating Landsat 9 images.

To enact these measures, meticulous planning is essential for Izeh city to mitigate the escalating temperature. Moreover, the findings of this research will play a pivotal role for urban policymakers and developers in evaluating the extent of land use-land cover alterations in the vicinity, aiming to enhance the future effectiveness and efficiency of the region, and bolster decision-making processes.