Evaluation of Prediction Performance of Vegetation Biomass Density for Two Different Case Study Areas in Turkey with Hybrid Wavelet and Artificial Neural Network Method

doi:10.21203/rs.3.rs-2589401/v1

Deterioration of natural resources such as vegetation, due to urbanization and increasing population density is evident in many areas of the world. For land use planning, it is vital to assess the plant density and forecast its future changes in light of vegetation-climate interactions given the current trend of global climate change. The purpose of this article is to show how we can detect the variation of vegetation density and forecast its future values with the Artificial Intelligence (AI) methods. As case studies, we selected 2 districts namely Alanya in Antalya province and Iznik in Bursa province of Turkey, that showed the highest and lowest land cover change between 2006 and 2018 respectively, according to the CORINE land cover classification. In the analysis, we have used satellite data (Enhanced Vegetation Index (EVI) and Land Surface Temperature (LST) data from MODIS/Terra satellite) and atmospheric data (archive precipitation and temperature measurements at meteorological stations) to define vegetation changes up to 2030.We have used ANN with original data and with the data obtained by the wavelet transform application (W-ANN). The average EVI value for 2030 was calculated as 0.22 with a 5.4% error probability for Iznik, and 0.28 with a 2% error probability for Alanya. By comparing the predicted values of W-ANN for 2030 with respect those of 2018, vegetation biomass density will decrease by 21.4% in Iznik, and 6.6% in Alanya. The results were also compared with the Landsat Normalized Difference Built-up Index (NDBI).

Remote Sensing

Wavelet-ANN Modeling

Spectral Indexes

LST

NDBI

LULC

Land cover, which expresses the cover covering the earth's surface, is constantly changing over time due to both natural and human reasons (Hassan et al. 2016). Continental movements can be shown as an example of a land cover changing due to natural causes i.e., tsunamis, erosion, landslides, floods (Breuste et al. 2021). The human, who determines how the land will be used, causes changes in the land cover due to reasons such as urbanization and industrialization. (Stephens et al., 2019). It is very important to control and map the change over time, to determine the current situation in the region, to analyze the changes and to have information about the ecological processes that cause this change to be able to project on the future. Especially after the industrial revolution, the rapid and uncontrolled change of land cover/land use (LULC) seriously affects the climate, biological diversity and many other vital systems. Therefore, analyzing LULC is very important in terms of ensuring sustainability on earth (Lambin et al. 2001; Thuiller et al. 2015).

The studies on determining LULC using conventional methods take a long time, the LULC data based on the satellite images have become an important data source for the environmental studies. In recent years observing the large areas with satellite data or monitoring the change over time provides the opportunity to determine the spatial distribution of LULC classes (Caldas et al. 2007; Turner et al. 2007). In addition, ease of access to the data, reliability, synoptic view, broad spectrum and extraction of high-level feature information enable remote sensing to play an important role in mapping in many applications (Karaca 2018).

Land use changes are dynamic in nature. It is influenced by a multitude of factors operating on a local, regional and global scale (Hassan et al. 2016; Rahman et al. 2012). Rapid population growth in the urbanization process, along with industrial development and economic growth, causes the land use model to change constantly (Friedberger 2000; Kijas 2002; Merson et al. 2010; Yue et al. 2013; Dutta et al. 2019). As a result, a significant number of green spaces has become urbanized over the years (Dur and Yigitcanlar 2015) and the ecosystem balances specific to these areas have been permanently lost (Sinclair 2003). In other words, the deterioration of natural resources, the destruction of vegetation, urbanization, increasing population density, and climate change due to human influence are giving signals for the future of humanity (Ng and Marsalek 1989; Brun and Band 2000; Weng 2001; Lin et al. 2007; Li and Wang 2009; Krellenberg 2013; Wang and Kalin 2018).

Academic studies on urbanization in Turkey have generally shown that economic, technological, political and socio-psychological factors are effective on migration (Keleş 2016). Studies reveal that there is a parallelism between the increase in the urban population and the level of industrialization. For example, Bursa is one of the most important industrially developed provinces of Turkey. Iznik, chosen as the study area, is one of the districts of Bursa with the lowest urbanization rate with 0.39% from 2006 to 2018. According to MODIS data, it was determined that forest areas decreased by 14.66% from 2006 to 2018 in Iznik (Isler and Aslan 2021). On the other hand, Antalya is a very popular developing province, whose economy is dominated by tourism. Alanya, chosen as the study area, is one of the districts of Antalya with the highest urbanization rate with 7.14% from 2006 to 2018. In addition, according to MODIS data, it was determined that forest areas decrease by 13.6% from 2006 to 2018 in Alanya (Isler and Aslan 2021).

The most common spectral indices used in the literature for the detection of LULC change are NDVI (Normalized Difference Vegetation Index), EVI (Enhanced Vegetation Index), BRBA (Band Ratio for Built-up Area), Soil index, BCI (Biophysical Mix Index), NDBI (Normalized Difference Built-up Area Index), NDISI (Normalized Difference Impervious Surface Index) and SVI (Soil and Vegetation Index) (Kayman and Sunar 2015; da Silva et al. 2020). These spectral indices are used significantly in various applications like LULC mapping, forest biomass mapping, agricultural monitoring, hazard mapping, urban heat island mapping, and monitoring (Yue et al. 2016; Narayan et al. 2016; Alves and Lopes 2017; Shatnawi and Qdais 2019; Kamali et al. 2021).

Artificial neural network model has been used frequently in the literature because it gives reliable results in simulation and estimation of satellite data (Shatnawi and Qdais 2019; Tu and et al 2020). NARX, which is a feedback artificial neural network that gives successful results in nonlinear system modeling and prediction applications in time series, is a frequently used method for simulating satellite data and forecasting. Shatnawi and Qdais applied NARX ANN model to simulate and estimate of LST values for the next 10 years using between 2000 and 2006 data (Shatnawi and Qdais 2019). The results of another study show that the NARX model is very successful and has very strong fault tolerance capability in LAI time series estimation based on MODIS data (Wang et al. 2019). Although NARX ANN is a reliable and successful method for estimating nonlinear data, it is possible to increase the estimation success by developing hybrid methods (Alves and Lopes 2017). For example, in some literature studies, it is seen that the performance can be increased with a hybrid model created by combining the ANN and wavelet analysis method (Terzi and Barak 2013; Anshuka et al. 2019).

The aim of the paper is to investigate regional urbanization with statistical and artificial intelligence methods using satellite data and to forecast the vegetation biomass density until 2030. For this purpose, NARX artificial neural network model was used to determine the changes in EVI. Satellite data were analyzed together with air temperature and precipitation data from Turkish State Meteorological Service (TSMS) between 2005–2018. In addition, CORINE LULC data has been used to monitor LULC changes over the years. Then, the findings were validated together with the NDBI (Normalized Difference Built-up Index) data obtained from the Landsat 8 OLI satellite sensor and LULC changes in the area. Vegetation change scenarios for the year 2030 are interpreted according to the result of the developed hybrid model and the statistical relationships between the parameters. Since there are not many studies in the literature on long-term estimation of vegetation biomass density using EVI data, this study will be important in terms of understanding the relationship between vegetation biomass density, LST, air temperature, precipitation, and NDBI at a local scale.

2.1. Study Area

We have chosen as study areas 2 districts (Iznik and Alanya) of two regions of Turkey (Fig. 1) since according to CORINE LULC data, the two districts, resulted to have had respectively the lowest and the highest LULC change in these last decades. The reasons for this choice, as stated in the introduction, are to test the capacity of the selected methods to detect and predict changes of different amplitudes.

Iznik, with an area of 73 km², is in the Marmara region, in the northeast of Lake Iznik, on a plain surrounded by mountains from the north and south. In addition, it is surrounded by the Samanlı Mountains in the north and the Katırlı Mountains in the south and located in the temperate climate zone. The population growth rate, which was 0.2% in 2017, increased to 1.68% in 2018 (Url-2). Iznik has very rich vegetation and it is located in the contact area of Mediterranean flora and Pale boreal European flora. The amount of precipitation varies depending on the altitude and it has an important effect on the distribution of vegetation. Forest, maquis and pseudo maquis plant group dominate in the region.

Alanya, with an area of 1.582 km² is one of the most crowded districts of Antalya, is located in the Mediterranean Region. It is surrounded by the Mediterranean Sea in the south, Gazipasa in the east, Manavgat in the west, and the Taurus Mountains in the north. The population growth rate was 1.67% in 2017, it was increased to 4.29% in 2018 (Url-7). Since Alanya has the most fertile soils of the Mediterranean region in terms of both its climate and location, the plant species grown also show diversity. In addition, 750 hectares of land are afforested every year in order to increase forest areas in Alanya (Url-5).

2.2 Data Used

2.2.1. Meteorological Data

Vegetation density, distribution and boundaries of vegetated areas are constantly changing according to the change of external factors such as meteorological data (especially precipitation and air temperature) (Kaya and Aladağ 2009). Since the number of missing temperature and precipitation data is less since 2005, the air temperature and precipitation data to be used in the study were obtained from the TSMS archive for the years 2005–2018.

According to 2005–2018 TSMS data, the annual average temperature is 16.1 ºC in Iznik and 21 ºC in Alanya (Fig. 2). When the monthly average temperature values are analyzed, August is the warmest month of the year and the temperature is 25.7 ℃ in Iznik and 30.0 ℃ in Alanya. January is the coldest month with 6.6 ℃ in Iznik and 12.6 ℃ in Alanya. The average rainfall in January is 56.6 kg/mm in Iznik and 180 kg/mm in Alanya. The total annual average precipitation is 484.5 kg/mm in Iznik and 893.7 kg/mm in Alanya between 2005 and 2018. As seen in Fig. 2, Alanya has higher temperature and precipitation than Iznik, depending on its geographical location.

2.2.2 Satellite Data

In this study, MODIS data were used to obtain EVI and LST values. MODIS is a key instrument that collecting data at 250 m, 500 m and 1 km spatial resolution twice daily since 2000, with global coverage by NASA's Terra and Aqua satellites. Raw data from MODIS sensors is combined into daily (8 days and 16 days) and annual images. The combined images provide data as different MODIS products for atmospheric, ocean or land processing applications. The sensor specifications of MODIS are given in Table 1.

Table 1

Main sensor specifications of MODIS (Url-9).
Specifications	MODIS
Spatial resolution	0.25km x 0.25km (bands 1–2) 0.50km x 0.50km (bands 3–7) 1.00km x 1.00km (bands 8–36)
Radiometric resolution	12 bits
Temporal resolution	Twice daily
Spectral resolution	36 bands
Orbit	Sun-synchronous, polar
Launch data	First instrument 1999 (Terra) Second instrument 2002 (Aqua)
Altitude	705 km

Atmospheric aerosol, clouds, or other atmospheric conditions, causing incomplete data, often affect satellite images. The TERRA satellite on the MODIS system, which makes EVI and LST measurements, has transmitted these immeasurable values to us as -3000 or N/A data (Url-3). In other words, 0.4% of EVI data and 2.4% of LST data were transmitted as missing data. To overcome this problem, EVI and LST missing data were recalculated by averaging the previous and subsequent data.

As seen in Fig. 3, the highest and lowest annual average EVI values were measured as 0.33 and 0.27 in 2011 and 2016 in Iznik, respectively. The lowest annual average LST value was recorded 18.7 ℃ in 2011. The highest and lowest annual average EVI values were measured as 0.32 and 0.28 in 2013 and 2007 in Alanya, respectively. The lowest annual average LST value was recorded 23.2 ℃ in 2006. As seen in Fig. 3, it can be concluded that the annual average EVI values of both regions are quite similar, whereas the LST in Alanya is higher than in Iznik due to the higher annual average temperature shown in Fig. 2.

2.2.3 LULC data

CORINE LULC data were also evaluated to detect temporal LULC types are changes. According to the CORINE land cover classification system Level 1, the land cover divided into 5 classes. (Table 2). Between 2006 and 2018, artificial surfaces increased by 6.25% in Iznik. Although agricultural areas increased by 3.04%, forest and semi-natural areas decreased by 2.67%. In Alanya, forest and semi-natural areas increased by 34% from 2006 to 2012 but 3% decrease was observed in 2018. For all that, artificial surfaces increased by 35% until 2018.

Table 2

Changes in CORINE LULC classes in Iznik and Alanya between 2006–2018 (Url-6).
	Iznik			Alanya
	2006	2012	2018	2006	2012	2018
Artificial surfaces (1)	0.80	0.85	0.85	1,98	2.22	2.33
Agricultural areas (2)	35.52	36.60	36.60	26.24	25.34	25.25
Forest and semi-natural areas (3)	42.58	41.44	41.44	71.32	71.66	71.63
Wetlands (4)	0.42	0.42	0.42	-	-	-
Water bodies (5)	20.69	20.68	20.68	0.46	0.79	0.79

2.3 Methodology

The flowchart of the methodology used in this paper is presented in Fig. 4.

As seen in Fig. 4, not only satellite data but also daily meteorological data were used as input features for building future scenarios. In this context, EVI, LST, air temperature and precipitation were analyzed together to define vegetation biomass density up to 2030. As a first step, data must be preprocessed (noise suppression) to replace or delete obvious incorrect values. The approach provides nonlinearity (i.e., normalization) of data. Normalization is done before the data is used in the network. With the normalization process, excessive oscillations in the data are prevented and the system performance is increased. In this study, the data were scaled between [-1,1] by min-max normalization. After the preprocessing of the data was completed, prediction simulation was performed with NARX ANN modeling. A hybrid model was developed with W-ANN to improve the model performance. Afterwards, the EVI scenarios obtained from the best performing model were examined for their compatibility with CORINE LULC and Landsat NDBI data. As a result, temporal variables of biomass density until 2030 was simulated.

2.3.1 Spectral Indices

Spectral indices are combinations of pixel values from two or more spectral bands in a multispectral image, created to draw attention to pixels that show the relative abundance or absence of a particular type of land cover in the image (for example, vegetation or urban). The spectral indexes EVI and NDBI, which are two separate ones, were utilized in the analysis.

EVI

It is a common index used to measure vegetation greenness, as is the NDVI. An important reason for its widespread use is the correction of some atmospheric conditions and canopy background noise, which makes it more accurate in areas with dense vegetation. The calculation of EVI is given by Eq. 1 (Huete et al. 2002; Galvão et al. 2011; Çelik 2016).

$$EVI=G.\frac{\left(NIR-RED\right)}{\left(NIR+C1.RED-C2.BLUE+L\right)} \left(1\right)$$

where G = a gain factor (2.5),

NIR = reflection rate of near infrared bands,

RED = reflection rate of red bands,

C1, C2 = aerosol resistance parameters (C1 = 6 and C2 = 7.5),

L = canopy (vegetation) value (1).

In vegetation pixels, EVI values range from − 1 to 1, with healthy vegetation often falling between the ranges of 0.20 and 0.80. (Sinergise 2020).

For the years 2005–2018, EVI, the second vegetation layer of the MODIS Vegetation Indices (MOD13Q1) Version 6 data produced every 16 days at a spatial resolution of 250 meters as a Level 3 product, was used. Using low cloud, low view angle, and highest EVI criteria, the system chooses the best possible pixel value from all of the acquisitions made over the 16-day period. (URL-3). The MODIS Vegetation Index is derived from daily, bidirectional surface reflectance corrected for the atmosphere. Accuracy of the EVI is within ± 0.015 (Jarchow et al. 2018; Skakun et al. 2018).

NDBI

Another significant indicator that makes use of the near-infrared (NIR) and mid-infrared (MIR) bands to track changes in urban areas over time (Zha et al. 2003). As shown in Eq. 2, the index is obtained by rationing the difference of the spectral reflectance values in these two bands.

$$NDBI=\frac{\left(MIR – NIR\right)}{ \left(MIR + NIR\right)} \left(2\right)$$

where MIR = reflection rate mid-infrared and NIR = reflection rate near-infrared.

The NDBI values are between − 1 and 1. The greater NDBI value represents a higher proportion of the built-up area. NDBI value is low in areas where vegetation is dense (Bostan et al. 2016). The NDBI data used in this study were obtained from the Landsat 8 OLI/TIRS satellite sensor for the years 2014–2018. The relationship between NDBI and other parameters was analyzed by machine learning. NDBI data was also used to support the validation of the model results.

2.3.2 Land Surface Temperature (LST) Retrieval

LST is a crucial data source that aids in our comprehension of the climatic, environmental, and physical characteristics of the Earth's surface (Çolak and Sunar 2020; Jimenez-Munoz et al. 2009). LST can be estimated from satellite data or directly ground-truth measurements. In this study, LST dataset estimated from the infrared radiations measured on channels 31 and 32 with the MODIS-TERRA sensor, is obtained from the MOD11A2 product (with 8-day temporal, 1 km spatial resolution) for the years 2005–2018. LST can only be measured in clear sky conditions. The equation used for the definition of LST from MODIS Terra MOD11A2 is given below (Artis and Carnahan 1982).

$$LST=\frac{\text{T}\text{c}}{1+\left(\lambda *\frac{Tc}{p}\right)In\left(e\right)} \left(3\right)$$

where Tc = satellite brightness temperature

λ = wavelength of emitted radiance (10.6µm)

p = h.c/ σ (1.438.10^− 2 m. K)

σ = (1.38.10–23 JK^− 1)

h = Planck’s constant (6.626.10^− 34 J.s)

c = Velocity of light (2.998.10⁸ ms^-1)

2.3.3 Modeling

Artificial neural networks are a method that gives successful results in model studies and is frequently encountered in the literature. In simple terms, neural networks train the network using past and present data, find hidden dependencies between data and use them to future scenarios forward. NARX model is one of the most widely used feed forward model types among artificial neural network models. In the study, NARX feedback ANN model are used for definition of vegetation biomass density for 2030. In order to increase the NARX ANN model performance, a hybrid model was developed by combining wavelet and NARX ANN.

2.3.3.1 NARX (Nonlinear Autoregressive Exogenous Neural Network) Model

NARX networks are an artificial neural network model proposed by Lin et al. (1996). It is used to define where the desired output depends on historical data. NARX is a feedback and forward computation looped dynamic neural network with different layers such as input layer, hidden layer, delay layer and output layer. The NARX model was created by taking the linear ARX (Autoregressive Exogenous) model as an example, which is widely used in time series modeling. As with many time series models, in the NARX model, the dependent variable in the dataset is a function of its historical value. NARX model can be expressed with the following function (Url-4);

$$y\left(t\right)= f\left(y\left(t - 1\right), y\left(t - 2\right), \dots , y\left(t - ny\right), x\left(t - 1\right), x\left(t - 2\right), \dots , x\left(t - nx \right)\right) \left(4\right)$$

where

y(t-1), y(t-2), … ,y(t-ny); network outputs

x(t-1), x(t-2), … ,x(t- nx); network inputs.

ny and nx; the number of historical outputs and historical inputs to be used for feedback, respectively.

The y(t) dependent on the output signal is calculated by the signal of the previous output value and the rotation of the previous independent input signal. NARX networks can be designed using parallel or serial-parallel architectures (Xie et al. 2009; Jyothi 2015).

It is seen that there are many different techniques when the future predictions are examined in the literature. Examples of some models employed are Back Propagation, ART (Adaptive Reverberation Theory), simple Recurrent Networks, and NARX networks (Nonlinear Autoregressive Exogenous). In this study, a feedback artificial neural network model called NARX networks was selected because it is thought to be effective for time series estimation and modeling nonlinear systems. In addition, it should be highlighted that when compared to conventional feedback network architectures, it converges faster and offers more effective learning (Sani et al. 2014; Gündoğdu 2020).

In this study, NARX artificial neural networks are used for modelling of EVI values since they need to be estimated depending on other variables. NARX network with serial-parallel architecture was first used for training in the network, and after completing the network training, a feedback loop was added and was used parallel architecture (Url-2). In the model, LST, air temperature and precipitation values were taken into account as input and EVI as output data.

2.3.3.2 Wavelet Transformation (WT)

More than one interconnected parametric variable (such as vegetation cover, leaf area index, precipitation, population density, number of buildings, etc.) should be analyzed in order to assess the climate change and urbanization process in order to choose the best course of action. More than one nested parametric variable is looked at to determine the urbanization process within the study's parameters. Because the variables are chaotic, it is less likely that events based on a specific time period may be predicted (Labat et al. 2005). To improve the performance of the model in this work, a hybrid model using wavelet transform and ANN was developed. Since detailed information about the signal can be obtained in the frequency and time domain, the wavelet transform, which is widely used in many areas, is preferred in the hybrid model.

In a wavelet analysis that shifts the time domain and scales the wavelet function with time, a wavelet function (ψ (x)) should be utilized. The mathematical representation of a signal given by WT is as follows (Percival and Walden 2002; Aslan et al. 2019):

$${ {\psi }}_{\left(a,b\right)}\left(t\right)=\frac{1}{\sqrt{a}}{\psi }\left(\frac{\text{t}-\text{b}}{a}\right) \left(5\right)$$

where ψ_ (a, b) = continuous wavelet transforms,

a = scaling parameter,

b = conversion parameter,

Ψ = wavelet function (main wavelet).

In the study, LST, temperature and precipitation data are divided into sub-detail components using discrete wavelet transform. Parameters with a correlation value greater than 0.01 with the original EVI data were given as input to the ANN.

In the ANN model, there is one input layer, one hidden layer and one output layer. Three processing elements (neurons) are used in the input layer for LST, temperature and precipitation parameters. EVI data is used in the output layer. The number of neurons in the hidden layer is one of the most important factors affecting the learning success of the network. If the number of neurons in the hidden layer is less, learning will not be as successful. Even if the number of neurons in the hidden layer is more than it should be, it may decrease the performance of the network as it will cause the models to memorize the data (Çakır 2018). In the paper, number of neurons was gradually increased until the best number of hidden layer neurons was obtained.

In the model, 2005–2015 data (264 data) were used to train, 2016–2018 data (72 data) were used to test and validation. In other words, 60% of the data was used for training, 40% for validation and testing. The model performance was calculated by Mean Absolute Percent Error (MAPE), Mean Square Error (MSE), Root Mean Square Error (RMSE), and correlation coefficient (R) statistics.

In the NARX model created, the best result was obtained when 6 hidden layer neurons were selected in both stations. The model performance according to the number of selected hidden layer neurons is given in Table 3.

Table 3

Comparison of the MSE and R values of the number of neurons in the ANN system.
	Model	ANN
Station	Hidden Neuron Number	3		6		9		12
		MSE	R	MSE	R	MSE	R	MSE	R
Iznik	Training	3.8e-4	0.74	2.4e-4	0.82	3.5e-4	0.75	3.2e-4	0.75
	Validation	2.5e-4	0.84	4.1e-4	0.78	3.3e-4	0.8	3.3e-4	0.79
	Test	2.1e-4	0.74	4.7e-4	0.77	4.7e-4	0.72	3.6e-4	0.76
	Output	3.2e-4		3.7e-4		2.8e-4		3.3e-4
Alanya	Training	2.5e-4	0.82	2.7e-4	0.84	2.4e-4	0.82	2.5e-4	0.83
	Validation	3.2e-4	0.82	2.5e-4	0.83	1.9e-4	0.83	2.7e-4	0.83
	Test	2.7e-4	0.75	3.2e-4	0.79	2.0e-4	0.81	3.2e-4	0.72
	Output	4.3e-4		4.1e-4		2.1e-4		4.6e-4

The relationship between the target values of training, validation and test results is as in Table 3. The training values are 82% and %84 for Iznik and Alanya, respectively. In general, 70% and 80% R values are accepted for good training in ANN scans (Sağık et al. 2021). As a result of the training with the ANN, the estimated percentage of all data are as 80% and 82% for Iznik and Alanya, respectively.

The performance results of ANN model, which has the most successful performance according to the number of hidden layer neurons, according to the MAPE and RMSE performance indications are given in Table 4.

Table 4

ANN Model performance for Iznik and Alanya.
Station	Model	Training		Test		All
Station	Model	R	R²	R	R²	R	R²	MAPE	RMSE	MSE
Iznik	ANN	0.82	0.67	0.77	0.59	0.80	0.64	15.9	0.29	0.002
Alanya	ANN	0.84	0.70	0.79	0.62	0.82	0.67	5.1	0.02	0.001

As a result of ANN analysis, MAPE values were calculated to be 15.9% and 5.1% for Iznik and Alanya, respectively, and 0.29 and 0.02 for RMSE. Furthermore, the MSE values were 0,002 and 0,001 for Iznik and Alanya, respectively. The low RMSE and MAPE values indicate that the proposed model is successful, in other words, the proposed model results are successful of both stations.

The literature review shows that the success ratio of model based on hybrid models by preprocessing of the data with ANN increases (Partal and Kişi 2007; Zhang et al. 2014; Rout et al. 2014). In this hybrid model, LST, temperature and precipitation data were divided into 8 sub-components with Discrete Wavelet Transform (DWT). The data was analyzed by using different wavelets and, the best model performance was obtained by using the 'd4' wavelet. This step is of critical importance as the selection of the appropriate main wavelet and the determination of the appropriate level of decomposition according to the characteristics of the data affect the performance analysis (Emhan 2013). The attributes of the outputs were obtained with the statistical operations and reanalyzed as NARX ANN input parameters. In this study, the sum of the components with a correlation coefficient greater than 0.1 was used as the input parameter of W-ANN to define the subcomponents. The W-ANN model performance according to the number of selected hidden layer neurons is given in Table 5.

Table 5

Comparison of the MSE and R values of the number of neurons in the W-ANN system.
	Model	W-ANN
Station	Hidden Neuron Number	3			6		9		12
			MSE	R	MSE	R	MSE	R	MSE	R
Iznik	Training		3.7e-4	0.77	4.2e-4	0.96	2.08e-4	0.95	2.3e-4	0.92
	Validation		2.1e-4	0.82	2.7e-4	0.94	2.8e-4	0.91	2.3e-4	0.93
	Test		2.7e-4	0.76	1.8e-4	0.94	4.4e-4	0.94	2.4e-4	0.91
	Output		2.8e-4		2.9e-4		3.06e-4		2.3e-4
Alanya	Training		2.6e-4	0.81	2.7e-4	0.81	1.9e-4	0.91	2.2e-4	0.85
	Validation		3.3e-4	0.84	2.1e-4	0.86	2.3e-4	0.85	2.3e-4	0.86
	Test		2.6e-4	0.85	2.4e-4	0.83	3.5e-4	0.81	2.6e-4	0.82
	Output		2.8e-4		2.5e-4		2.4e-4		2.6e-4

The best result was obtained when the number of hidden layer neurons was 6 for Iznik and 9 for Alanya. As a result of the model developed by using W-ANN, the training values are 96% and %91 and the estimated percentage of all data increase to 95% and 88% for Iznik and Alanya, respectively. With the hybrid model developed, it was determined that the network provided an 18.7% and 7.3% increase in learning success in Iznik and Alanya, respectively. The performance evaluation of the network developed using the W-ANN model is given in Table 6.

Table 6

W-ANN Model performance in Iznik and Alanya.
Station	Model	Training		Test		All
Station	Model	R	R²	R	R²	R	R²	MAPE	RMSE	MSE
Iznik	W-ANN	0.96	0.92	0.94	0.88	0.95	0.90	5.4	0.02	0.001
Alanya	W-ANN	0.91	0.80	0.85	0.72	0.88	0.77	2	0.02	0

The EVI results made with both models are given in Fig. 5. As seen in Fig. 5, the average value of the 2018 EVI was calculated as 0.28 and 0.30 in Iznik and Alanya, respectively. On the other hand, according to the W-ANN model results, the average EVI for 2030 is expected to be 0.22 and 0.28 in Iznik and Alanya, respectively. EVI value for the Iznik W-ANN model results will decrease by 21.4% until 2030 with a 5.4% error probability compared to 2018 (Fig. 5a). Similarly, with a 2% error probability in Alanya, the EVI value will decrease by 6.6% until 2030 compared to 2018 (Fig. 5b).

It is also important to explore the relationship between NDBI, LST and EVI to better understand the impact of the results on urbanization (Mathew et al., 2015; Malik, 2019) (Fig. 6). According to the EVI and NBDI data observed from 2014 to 2018, it was determined that there was a strong negative relationship between the variables (R = 0.87 and R = 0.98, alpha = 0.01 in Iznik and Alanya, respectively). When the data between 2005 and 2018 were examined, it was determined that a moderately negative correlation between EVI and daily average air temperature (R = 0.63 and R = 0.57, alpha = 0.01 in Iznik and Alanya, respectively) was slightly stronger than that between EVI and LST (R = 0.54 and R = 0.55, alpha = 0.01 in Iznik and Alanya, respectively). EVI was shown a strong positive relationship with precipitation during the whole period (R = 0.88 and R = 0.80, alpha = 0.01 in Iznik and Alanya, respectively) (Fig. 7).

As seen in Fig. 6, the strong relationships between these parameters can be a very good indicator for the future urbanization rates of the selected regions.

As a result of the statistical research, it was determined that relationship between EVI and LST is statistically a moderately downhill relationship (R = 0.54 and R = 0.55, alpha = 0.01 in Iznik and Alanya, respectively). As the density of vegetation biomass on the land surface increases, a decrease in land surface temperature is observed since the evaporation it causes is greater. For this reason, it causes cooling because the amount of energy lost is high in areas where the vegetation is high and dense. The weakening of the negative relationship between LST and vegetation biomass density shows the effect of continuous expansion of impermeable surfaces on increasing land surface temperature. According to this situation, it can be predicted that the land surface temperatures are high in areas where the vegetation biomass density is low and the land surface temperatures will increase as the vegetation biomass density decreases. The relation between EVI and air temperature is a moderately negative (R = 0.63 and R = 0.57, alpha = 0.01 in Iznik and Alanya, respectively). Numerous studies have found that the air temperature is lower in green spaces than in cities. It is anticipated that a rise in air temperature will result in a drop in plant density.

It is important for LULC studies to determine the time-dependent variation of vegetation biomass density in a region. In this study, some preliminary results were presented regarding the future effects of vegetation biomass density change and the problems arising from urbanization in two different districts with the NARX ANN method.

Within the scope of this study, two different models, ANN and W-ANN, were used to define the change in vegetation biomass density until 2030 in Iznik and Alanya. The model results of Iznik showed that when the EVI value was estimated only by ANN, the test result of the network was 0.77 and the overall success R-value was 0.80. When it was estimated by W-ANN, the test result of the network was 0.94 and the overall success R-value was 0.95, which shows that the W-ANN network increased the overall learning success by 18.7% in 2018 EVI average value was calculated as 0.28. According to W-ANN results, the average value of EVI until 2030 was calculated to 0.22 with a 5.4% error probability. According to these results, vegetation biomass density will decrease by 21.4% until 2030 compared to 2018 in Iznik. On the other hand, the simulation results of Alanya showed that when the EVI was estimated only by ANN, the test result of the network was 0.81 and the overall success R-value was 0.82. Analyzing W-ANN, the test result of the network was 0.81 and the overall success R-value was 0.88, which shows us that the W-ANN network increased the overall learning success by 7.3%. 2018 EVI average value was calculated as 0.30. According to W-ANN results, the average value of EVI in 2030 was estimated to be 0.28 with a 2% error probability. According to these results, vegetation biomass density will decrease by 6.6% until 2030 compared to 2018 in Alanya. As a conclusion, the model results showed that the hybrid model increased the learning success ratio by 18.7% and %7.3 in Iznik and Alanya, respectively. Therefore, considering the W-ANN model, which gives better results, there would be a decreasing trend in vegetation biomass density in both stations.

The relationship between parameters (i.e., LST, air temperature, precipitation, NDBI and EVI) for urbanization potential was also investigated. Very strong negative correlation defined between EVI and NDBI (R = 0.87 and R = 0.98, alpha = 0.01 in Iznik and Alanya, respectively). Therefore, it was concluded that urbanization has a direct relationship with the decrease in vegetation biomass density. Besides it was determined that there is a strong positive relationship between EVI and precipitation (R = 0.88 and R = 0.80 in Iznik and Alanya, respectively). Consistency of the results with those of previous literature studies in the selected study areas is of great importance, in terms of determining the policies to be applied in the urbanization process and improving the welfare and health of future generations.

Some additional research topics by planned for the future scope of the present work. First, the EVI and LST values obtained using other LST retrieval algorithms or from other satellite sensors (with different spatial resolutions) will be considered to investigate their impact on model performances. Second, it will be tested in other districts with in situ measurements to verify the validity of the model. Third, learning performances will be compared using other methods and larger amounts of data in the analysis. Finally, in order to generalize the applied method, it will be tested by applying it to other regions with different climatic conditions and physiographic characteristics.

Author contributions All authors contributed to the conception and design of the study. Material preparation, data collection and analysis were done by B. Isler. B. Isler wrote the first draft of the manuscript which was later critically reviewed and edited by Z. Aslan and F. Sunar. The final check was done by A. Güneş, E. Feoli, and D. Gabriels, and they gave their opinion on critical issues. All authors have read and approved the final manuscript.

Funding No funding was received for this study.

Data availability Data sharing not applicable to this article as no datasets were generated or analysed during the current study.

Conflicts of interest All authors declare that they have no conficts of interest

Alves EDL, Lopes A (2017) The Urban Heat Island Effect and the Role Vegetation to Address the Negative Impacts Climate Changes in a Small Brazilian City, Atmosphere, 8(2), 18.
Anshuka A, van Ogtrop FF, Vervoort RW (2019) Drought Forecasting Through Statistical Models Using Standardized Precipitation Index: A Systematic Review and Meta-regression Analysis, Natural Hazards, 97, 955-977.
Artis DA, Carnahan WH (1982) Survey of Emissivity Variability in Thermography of Urban Areas, Remote Sensing of Environment, 12(4), pp.313-329.
Aslan Z, Dökmen F, Feoli E, Sıddıqı, AH (2019) Editor, Mathematical Modelling of Real-World Problems, Interdisciplinary Studies in Applied Mathematics, Mathematics Research Developments, NOVA Science Publishers Inc, New York, ss.103-118.
Atıcı R, Sağır S, Emelyanov LY, Lyashenko M (2021) Investigation of Ionospheric Electron Density Change During Two Partial Solar Eclipses and Its Comparison with Predictions of NeQuick 2 and IRI2016 Models, Wireless Personal Communications.
Breuste J, Haase D, Pauleit S, Sauerwein M (2021) How Vulnerable Are Urban Ecosystems and How Can Urban Resilience Be Developed with Them? In Urban Ecosystems (pp. 209-262). Springer, Berlin, Heidelberg.
Brun SE, Band LE, (2000) Simulating Runoff Behavior in an Urbanizing Watershed, Computers, Environment and Urban Systems, 24: 5-22.
Bostab S, Şanlı FB, Kaya Ş (2016) Nüfusa Bağlı Yerleşim Alanı Değişiminin Uzaktan Algılama Yöntemleri ile Analizi: Küçükçekmece Örneği, 6. Uzaktan Algılama-CBS Sempozyumu (Uzal-Cbs 2016), 5-7 Ekim 2016, Adana.
Caldas MDP, Ambrosano GMB, Haiter Neto F (2007) New Formula to Objectively Evaluate Skeletal Maturation Using Lateral Cephalometric Radiographs. Brazilian Oral Research, 21(4), 330-335.
Çakır FS (2018) Yapay Sinir Ağları (1. b.). Ankara: Nobel.
Çelik MA (2016) Bitki İndeks Modelleri (NDVI, EVI, VCI) Kullanılarak Akdeniz Bölgesi’nde Kuraklık Analizi (2000-2014). Phd, T.C. Kahramanmaraş Sütçü İmam University, Sosyal Bilimler Enstitüsü, Kahramanmaraş.
Çolak E, Sunar F (2020) Spatial Pattern Analysis of Post-fire Damages in the Menderes District of Turkey. Front. Earth Sci., 14, 446 – 461. https://doi.org/10.1007/s11707-019-0786-4.
Da Silva VS, Salami G, da Silva MIO, Silva A, Monteiro Junior JJ, Alba E (2020) Methodological Evaluation of Vegetation Indexes in Land Use and Land Cover (LULC) Classification. Geology, Ecology, and Landscapes, 4(2), 159-169.
Dur F, Yigitcanlar T (2015) Assessing Land-use and Transport Integration via a Spatial Composite Indexing Model. International Journal of Environmental Science and Technology, 12(3), 803-816.
Dutta D, Rahman A, Paul SK, Kundu A (2019) Changing Pattern of Urban Landscape and Its Effect on Land Surface Temperature in and around Delhi. Environmental Monitoring and Assessment, 191(9), 1-15.
Emhan Ö (2017) Yukarı– Aşağı İmleç Hareketine İlişkin EEG Kayıtlarının Ayrık Dalgacık, Knn ve Dvm ile Sınıflandırılması. MSc, Dicle University.
Friedberger M (2000) The Rural-urban Fringe in the Late Twentieth Century. Agricultural History, 74(2), 502-514.
Galvão LS, Santos JR, Roberts D, Breunıg DA, Toomey F M, Moura MYM (2011) On Intra-Annual EVI Variability in The Dry Season of Tropical Forest: A Case Study with MODIS and Hyperspectral Data, Remote Sensing of Environment, 115, Ss. 2350–2359.
Gündoğdu S (2020) Comparison of static MLP and dynamic NARX neural networks for forecasting of atmospheric PM10 and SO2 concentrations in an industrial site of Turkey. Environmental Forensics, 21(3-4), 363-374.
Hassan Z, Shabbir R, Ahmad SS, Malik AH, Aziz N, Butt A, Erum S (2016) Dynamics of Land Use and Land Cover Change (LULCC) Using Geospatial Techniques: A Case Study of Islamabad Pakistan, Springer plus, 5 (1), p. 812.
Huete A, Didan K, Miura T, Rodriguez EP, Gao X, Ferreira LG (2002) Overview of the Radiometric and Biophysical Performance of the MODIS Vegetation Indices. Remote Sensing of Environment. 83, 195–213.
Isler B, Aslan Z (2021) Prediction of Regional Urbanization and Climate Change Effects with Artificial Intelligence Methods, PhD. Thesis, Istanbul Aydın University, Institute of Science and Technology, Istanbul.
Jarchow CJ, Didan K, Barreto-Muñoz A, Nagler PL, Glenn EP (2018) Application and comparison of the MODIS-derived enhanced vegetation index to VIIRS, landsat 5 TM and landsat 8 OLI platforms: A case study in the arid colorado river delta, Mexico. Sensors, 18(5), 1546.
Jimenez-Munoz JC, Cristobal J, Sobrino JA, Soria G, Ninyerola M, Pons X, (2009) Revision of the Single-Channel Algorithm for Land Surface Temperature Retrieval from Landsat Thermal-Infrared Data. IEEE Geoscience and Remote Sensing Letters, (47), 339-349. https://doi.org/10.1109/TGRS.2008.2007125.
Jyothi MN, Dinakar V, Teja NSSR, Kishore KN (2015) NARX Based Short Term Wind Power Forecasting Model. TELKOMNIKA Indonesian Journal of Electrical Engineering, 15(1), 20-25.
Karaca AV (2018) Menderes İlçesindeki Orman Yangınının Süperpiksel Bölütleme Temelli Arama Yöntemiyle Tespiti. Journal of the Faculty of Engineering & Architecture of Gazi University, 18(2), 1-22.
Kamali M E, Hashemi H, Berndtsson R, Daneshkar Arasteh P, Kazemi M (2021) Impact of Spatiotemporal Land-use and Land-cover Changes on Surface Urban Heat Islands in a Semiarid Region Using Landsat Data. International Journal of Digital Earth, 14(2), 250-270.
Kaya B, Aladağ C (2009) Precipitation, Temperature and Vegetation Relations in the Conditions of Konya. Selçuk Üniversitesi Sosyal Bilimler Enstitüsü Dergisi, 22: 265, 78.
Kayman Ö, Sunar F (2015) Spektral İndekslerin Landsat TM Uydu Verileri Kullanılarak Arazi örtüsü/kullanımı Sınıflandırmasına Etkisi: İstanbul, Beylikdüzü İlçesi, Arazi Kullanımı Değişimi, TUFUAB VIII, Teknik Sempozyumu, Konya-Türkiye, 52-59, 21-23 May.
Keleş R (2016) Kentleşme Politikası (15. Baskı), İmge Kitabevi, Ankara.
Kijas J (2002) A place at the coast: Internal Migration and the Shift to the Coastal Countryside. Transformations, 2, 1Á12.
Krellenberg K, Müller A, Schwarz A, Höfer R, Welz J (2013) Flood and Heat Hazards in the Metropolitan Region of Santiago de Chile and the Socio-economics of Exposure. Applied Geography, 38, 86-95.
Labat D, Ronchail J, Guyot JL (2005) Recent Advances in Wavelet Analyses: Part 2-Amazon, Parana, Orinoco and Congo Discharges Time Scale Variability, Journal of Hydrology, 314(1-4), 289-311.
Lambin EF, Turner BL, Geist H, Agbola S, Angelsen A, Bruce J, Coomes O, Dirzo R, Fischer G, Folke C, George PS, Homewood K, Imbernon J, Lemmans R., Li X, Moran E, Mortimore M, Ramakrishnan PS, Richards JF, Dkanes H, Steffen W, Stone GD, Svedin U, Veldkamp T, Vogel C, Xu, J (2001) The Causes of Land-Use and Land-Cover Change: Moving Beyond the Myths, Global Environmental Change, 11 261–269.
Lin T, Horne B G, Tino P, Giles CL (1996) Learning Long-term Dependencies in NARX Recurrent Neural Networks. IEEE Transactions on Neural Networks, 7(6), 1329-1338.
Lin YP, NM Hong, PJ Wu, CJ Lin (2007) Modeling Responses of Hydrologic Processes to Future Watershed Land Use Scenarios and Climate Change in an Urbanized Watershed in Taiwan, Environmental Geology, 53:623–634.
Li Y, C Wang (2009) Impacts of Urbanization on Surface Runoff of the Dardenne Creek Watershed, St. Charles County, Missouri, Physical Geography, 30(6): 556–573.
Malik F, Umar Z (2019) Dynamic Connectedness of Oil Price Shocks and Exchange Rates. Energy Economics, 84, 104501.
Mathew A, Chaudhary R, Gupta N, Khandelwal S, Kaul N (2015) Study of Urban Heat Island Effect on Ahmedabad City and Its Relationship with Urbanization and Vegetation Parameters. Int. J. Comput. Math. Sci, 4, 126-135.
Merson J, Attwater R, Ampt P, Wildman H, Chapple R (2010) The Challenges to Urban Agriculture in the Sydney Basin and Lower Blue Mountains Region of Australia. International Journal of Agricultural Sustainability, 8(1-2), 72-85.
Narayan K, Khanindra P, Abhisek C, Sahoo S (2016) Urban Heat Island Explored by Correlationship Between Land Surface Temperature vs Multiple Vegetation Indices. Spat. Inf. Res. http://dx. doi.org/10.1007/s41324-016-0049-3.
Ng HYF, Marsalek J (1989) Simulation of the Effects of Urbanization on Basin Stream Flow, Water Resources Bulletin, 25: 117-124.
Partal T, Kişi Ö (2007) Wavelet and Neuro-fuzzy Conjunction Model for Precipitation Forecasting. Journal of Hydrology, 342(1-2), 199-212.
Percival DB, Walden AT (2002) Wavelet Methods for Time Series Analysis. Cambridge University Press, second edition.
Rahman A, Kumar S, Fazal S, Siddiqui MA (2012) Assessment of Land use/land cover Change in the North-West District of Delhi Using Remote Sensing and GIS Techniques. Journal of the Indian Society of Remote Sensing, 40(4), 689-697.
Rout M, Majhi B, Majhi R, Panda G (2014) Forecasting of Currency Exchange Rates Using an Adaptive ARMA Model with Differential Evolution-based Training. Journal of King Saud University-Computer and Information Sciences, 26(1), 7-18.
Sağık M, Altıntaş O, Ünal E, Özdemir E, Demirci M, Çolak Ş, Karaaslan M (2021) Optimizing the Gain and Directivity of a Microstrip Antenna with Metamaterial Structures by Using Artificial Neural Network Approach. Wireless Personal Communications, 118(1), 109-124.
Sani MG, Abdul Wahab N, Sam YM, Samsudin SI, Jamaludin I W (2014) Comparison of NARX neural network and classical modelling approaches. In Applied mechanics and Materials (Vol. 554, pp. 360-365). Trans Tech Publications Ltd.
Shatnawi N, Qdais HA (2019) Mapping Urban Land Surface Temperature Using Remote Sensing Techniques and Artificial Neural Network Modelling, International Journal of Remote Sensing, 40(10), 3968-3983.
Sentinel Hub by Sinergise (Sinergise) (2020) https://www.sentinel-hub.com/eoproducts/evi-enhanced-vegetation-index-0 (28 February 2020).
Sinclair I, Bunker R, Holloway D (2003) From the Outside Looking in–Planning and Land Management in Sydney’s Fringe. In State of Australian Cities National Conference.
Skakun S, Justice CO, Vermote E, Roger JC (2018) Transitioning from MODIS to VIIRS: an analysis of inter-consistency of NDVI data sets for agricultural monitoring. International journal of remote sensing, 39(4), 971-992.
Stephens L, Fuller D, Boivin N, Rick T, Gauthier N, Kay A, Ellis E (2019) Archaeological Assessment Reveals Earth's Early Transformation Through Land Use. Science, 365(6456), 897-902.
Terzi Ö, Barak M (2013) Dalgacık-Sinir Ağı Yaklaşımı ile Yağış-Akış Tahmini: Kızılırmak Nehri Örneği. Tarım Bilimleri Dergisi. Sayfa/pp: 546-557, Cilt 21, Sayı 4.
Thuiller W, Lavorel S, Araujo MB, Sykes MT, Prentice C (2015) Climate Change Threats to Plant Diversity in Europe, Proceedings of the National Academy of Sciences, 102(23), 8245-8250.
Tu Y, Zhou H, Lang W, Chen T, Li X, Xu B (2020) A Novel Cross-sensor Calibration Method to Generate a Consistent Night-time Lights Time Series Dataset. International Journal of Remote Sensing, 41(14), 5482-5502.
Turkish State Meteorological Service (TSMS) (2018) Data Normals of Provinces (1929-1028). Retrieved from https://www.mgm.gov.tr/veridegerlendirme/il-ve-ilceler istatistik.aspx?m=ISTANBUL, (12 January 2019).
Turner BL, Lambin EF, Reenberg A (2007) The Emergence of Land Change Science for Global Environmental Change and Sustainability. Proceedings of the National Academy of Sciences, 104(52), 20666-20671.
Wang R, Kalin L (2018) Combined and Synergistic Effects of Climate Change and Urbanization on Water Quality in the Wolf Bay Watershed, Southern Alabama. Journal of Environmental Sciences, 64, 107-121.
Wang J, Wang J, Shi Y, Zhou H, Liao L (2019) A Recursive Update Model for Estimating High-resolution LAI Based on the NARX Neural Network and MODIS Times Series. Remote Sensing, 11(3), 219.
Weng Q (2001) Modeling Urban Growth Effects on Surface Runoff with the Integration of Remote Sensing and GIS, Environmental Management, 28(6): 737-748.
Xie H, Tang H, Liao YH (2009) Time Series Prediction Based on NARX Neural Networks: An Advanced Approach. International Conference on Machine Learning and Cybernetics, China, 3, 1275–1279.
Yue W, Liu Y, Fan P (2013) Measuring Urban Sprawl and Its Drivers in Large Chinese Cities: The Case of Hangzhou. Land Use Policy, 31, 358-370.
Yue Y, Li M, Zhu A, Xinyue Ye Rui Mao, Jinhong W, Jin Dong (2016) Land Degradation Monitoring in the Ordos Plateau of China Using an Expert Knowledge and BP-ANN-Based Approach, Sustainability 2016, 8(11), 1174.
Zha Y, Gao J, Ni S (2003) Use of Normalized Difference Built-up Index in Automatically Mapping Urban Areas from TM Imagery. International Journal of Remote Sensing, 24(3), 583–594.
Zhang O, Cheng YB, Lyapustin AI, Wang Y, Xiao X, Suyker A, Verma S, Tan B, Middleton, EM (2014) Estimation of Crop Gross Primary Production (GPP): I. Impact of MODIS Observation Footprint and Impact of Vegetation BRDF Characteristics. Agricultural and Forest Meteorology 191, 51–63.

Url-1< (https://data.tuik.gov.tr/Search/Search?text=n%C3%BCfus%20art%C4%B1%C5%9F%20h%C4%B1z%C4%B1&dil=1) >, (Access date:07.06.2019).

Url-2< (https://www.nufusu.com/ilce/iznik_bursa-nufusu) >, (Access date:07.06.2019).

Url-3 <(https://lpdaac.usgs.gov/products/mod13q1v006/)>

Url-4<https://www.mathworks.com/help/deeplearning/ug/design-time-series-narx-feedback-neural-networks.html>, (Access date: 05.04.2018).

Url-5< http://www.antalya.gov.tr/Alanya>, (Access date:17.06.2019).

Url-6 < https://corine.tarimorman.gov.tr/corineportal/>

Url-7 < https://www.nufusu.com/il/antalya-nufusu>

Url-8 < https://earthengine.google.com/ >

Url-9< https://gisgeography.com/modis-satellite/>

No competing interests reported.

Evaluation of Prediction Performance of Vegetation Biomass Density for Two Different Case Study Areas in Turkey with Hybrid Wavelet and Artificial Neural Network Method

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials And Methods