Skip to content
BY 4.0 license Open Access Published by De Gruyter Open Access December 31, 2022

Agriculture and aquaculture land-use change prediction in five central coastal provinces of Vietnam using ANN, SVR, and SARIMA models

  • YuRen Wang and Nguyen Hong Giang EMAIL logo
From the journal Open Geosciences


Vietnam’s economy with agriculture and aquaculture still account for roughly 26% of the country’s gross domestic product, and nearly 70% of the Vietnamese population lives in rural areas; therefore, agriculture and aquaculture land use play a crucial role in the development process of Vietnam. Rapidly increasing population and infrastructure in rural areas and industrial zones lead to these land-use changes. Hence, these land-use change predictions are crucial for local authorities and the local people to make land-resource funds and set up planning. This article suggests support vector regression (SVR), artificial neural network (ANN), and seasonal autoregressive integrated moving average (SARIMA) methods to predict land-use change. By comparing the three models, the results indicate that almost all of the SVR models improve the accurate performance more than ANN and SARIMA in Quangtri, ThuThienHue, Danang, and Quảngnam provinces. Furthermore, the ANN model indicates more accurate forecasting than the SVR and SARIMA models in Quan Binh province. The result may be support for the Ministry of Natural Resources and Environment to conduct the land-use inventory and upgrade agriculture and aquaculture land-use change maps every 5 years. Afterward, the Department of Natural Resources and Environment’s provinces use the estimating database and update it manually.

1 Introduction

Spatial distributions of each land use, associated food security, environmental circumstances, socioeconomic conditions, and institutional and cultural issues may all significantly impact the evolution of regional land-use patterns [1,2,3,4,5,6]. Hence, several land, atmospheric, and environmental science applications are interested in the spatial distribution of land use as agriculture, aquaculture, forest, and urban [7,8]. Urbanization in Vietnam has extended residential and industrial zones into previously open, low-populated areas that were originally natural places for aquaculture and agricultural land uses during the last three decades [9,10]. On the other hand, agriculture and aquaculture still account for roughly 26% of the gross domestic product, and nearly 70% of the Vietnamese population lives in rural areas. As a result, land plays a critical role in Vietnam’s growth [11,12].

Land-use change prediction methodologies include three main groups: Geographic Information System (GIS), statistical, and artificial intelligence (AI) methods [13,14,15,16,17,18,19,20,21,22]. Apart from the GIS method, the boundary between the statistical and artificial neural network (ANN) methods is quite complicated. The literature uses several techniques that consist of a statistical technique, namely, seasonal autoregressive integrated moving average (SARIMA), or ARIMA models [23,24], and two AI techniques, namely, ANN and support vector machines [25,26,27,28,29,30]. Recently, many studies have applied machine learning to analyze land-use changes, as shown below. Wang et al. [28] used various kernels to forecast the correlations between cultivated land area and impact land-use change factors, including radial basis function, polynomial, and sigmoid kernels. The result indicated that it could be best used to estimate the cultivated land areas. Jiang et al. [31] used three models, named SARIMA, dynamic harmonics regression (DHR), and seasonal-trend decomposition procedure based on loss (STL) to simulate the moderate resolution imaging spectroradiometer leaf area index (LAI) from 2001 to 2006 and predict short-term LAI values in the future (2007). The study result showed that the SARIMA model was better at estimating R 2 values in all land cover types than the DHR or STL models. Li et al. [32] employed three different models of the previous 5 years’ weekly average, seasonal decomposition (SD) and Box–Jenkins multiplicative SARIMA models to predict the annual cycle of the weekly soil dryness index (SDI) series in the southwest of Western Australia. The result showed that the SARIMA model outstripped the other models in estimating the annual cycle of SDI data. Kong et al. [33] used GIS and ANN methodologies to assess geo-environmental suitability for agricultural land use in Hangzhou, China’s rural–urban periphery. The result showed the two models successfully criticized agricultural land-use suitability based on the geomorphology, soil conditions, basic infrastructure, and geological environment. Zhang and Chen [34] evaluated the suitability of urban land use in Hefei using the ANN model. The research suggests that the model might be used to assess the feasibility of urban land use. Hence, these support vector regression (SVR), ANN, and SARIMA models are suitably deployed for land-use change prediction in Vietnam’s five central coastal provinces (Quangbinh, Quangtri, ThuaThienHue, Danang, and Quang Nam).

The study aims to describe the forecast of agriculture and aquaculture land-use change using SVR, ANN, and SARIMA algorithms. The models’ input vectors are based on 44 quarters of land usage across 11 years in five central coastal provinces in Vietnam (Quangbinh, Quangtri, ThuaThienHue, Danang, and Quang Nam). This study also highlights a comparison among SVR, ANN, and SARIMA that is based on the results of statistical accuracy parameters such as mean (M), root mean square error (RMSE), mean absolute error (MAE), standard deviation (St Dev), Pearson correlation coefficient (R), kurtosis coefficient (Kurt), skewness coefficient (Skew), minimum (Min), maximum (Max), and correlation of determination (R 2). The results of these three models may show the working efficiency of the models for land-use prediction. In addition, by projecting these types of land-use changes, future scenarios that can aid land-use planning and decision-making can be developed.

The following is the content of the article. The first section of the article is an introduction. The SVR, ANN, and SARIMA models are introduced in Section 2 and explained in detail for use throughout this study. The study findings and discussions are explained in Sections 3 and 4. Finally, the conclusions are presented in Section 5.

2 Study area and methodology

2.1 Study area

Quangbinh, Quangtri, ThuaThienHue, Danang, and Quảngnam provinces are located in the central coastal area of Vietnam (see Figure 1). This area’s topography has a lower elevation, starting from the mountainous range down to the midland hills, to the plains inside the coastal dunes, and finally to the coastal islands. The delta area has a coastal-mountainous nature, and it is divided by mountain branches close to the sea as Hoanh Son mountain range – Ngang pass, Bach Ma mountain range – Hai Van pass [35]. These provinces’ total land area is 2,950,669 ha, in which, Quangbinh, Quangtri, ThuaThienHue, Danang, and Quảngnam occupy 799,876, 470,123, 494,711, 128,473, and 1,057,486 ha, respectively.

Figure 1 
                  Five provinces’ locations.
Figure 1

Five provinces’ locations.

2.2 ANN model

An ANN model proposed in this study was created using Keras Neural Network, a highly modular neural network library. TensorFlow or Theano may be chosen as the backend [36]. Hence, Keras also contains plentiful modules that include activation function, layer, preprocessing, objective function, optimization method selection, etc. Deploying these modules indicates that network models may be easily set up, and core parameters of neural networks may be better enhanced [37]. The structure of the Keras Neural Network consists of a three-layer neuronal architecture as an input layer, hidden layer, and output layer [38]. The hidden layer may contain one or multilayers. The backpropagation learning algorithm of Levenberg–Marquardt has been used for the data feature extraction contained in the model’s inputs [39,40]. Figure 2 shows a structural view of the ANN model, as the below-input layer has m input nodes from i 1 to i m , one neuron of the output layer. There are n hidden layers as the first hidden layer contains neurons from H 11 to H 1m , the second one is from H 21 to H 2m , and the last one is from H n1 to H nm . Each neuron has a resemble Weight and Bias in the hidden layer, and the output layer, such as W 11 ( 2 ) , B 1 ( 1 ) and W n 1 ( n + 1 ) , B m ( n ) are the Weight and Bias to communicate with neuron H 11, and neuron H nm , respectively. The Sigmoid or Linear functions are commonly utilized equations for the activation functions for the hidden layers [41,42].

Figure 2 
                  The architecture of ANN for land-use prediction.
Figure 2

The architecture of ANN for land-use prediction.

2.3 SVR model

Regarding a set of training data {(x 1 , y 1), (x 2 , y 2) (x n , y n )}, where x i, y i, and n are an input, a target output, and a number of samples, respectively. Assuming SVR is to demonstrate a function of f(x) that may estimate accurate future values [43], and it is supposed by:

(1) f ( x ) = i = 1 n ( α i α i ) K ( x i x ) + c ,

where K is the kernel function (named linear, sigmoid, radial basis, and polynomial functions), which nonlinearly maps the input data into higher feature space, c is the bias, α* and α are the Lagrange multipliers that are optimized by a maximum of the quadratic function as follows:

(2) W ( α i , α i ) = 1 2 i , j = 1 n ( α i α i ) ( α j α j ) K ( x i , x j ) + i = 1 n y i ( α i α i ) + ε i = 1 m y i ( α i α i ) ,

s . t . i = 1 n ( α i α i ) α i , α i [ 0 , C ] ,

where C is called the penalty indicator, which commands the sanctioned degree on samples whose errors are beyond ε. Normally, only certain parts of the samples may suit the property: α i α i 0 , they are named support vectors since they only contribute to the model.

The kernel-based SVR is a powerful tool for solving regression problems. First, the kernel is chosen in a data-dependent way. Then, the applied kernel is a commonly linear function. They then describe different mappings from lower space to higher feature space [44,45].

2.4 SARIMA model

If the time series X t is considered as the SARIMA (p, d, q) × (P, D, Q, s) model with (p, d, q) nonseasonal order terms, (P, D, Q) seasonal order terms have the structure as follows [46,47,48]:

(3) p ( B ) Φ P ( B s ) W t = θ q ( B ) Θ Q ( B s ) ω t ,

where p and seasonal P denote several autoregressive terms (lags of the stationary series); d and seasonal D denote differencing that must be done to stationary series, q and seasonal Q demonstrate a number of moving average terms (lags of the forecast errors), s demonstrates the seasonal length in the data, ω t and B denote the backward shift operator and the white noise value at time t.

Equation (3) may be rewritten mathematically after replacing the value of W t = d ( B ) s D ( B ) X t .

(4) p ( B ) Φ P ( B s ) ( 1 B ) d ( 1 B s ) D X t = θ q ( B ) Θ Q ( B s ) ω t .

The SARIMA components may be rewritten as follows:

(5) Non seasonal auto regression ( A .R . ) : p ( B ) = 1   1 B   2 B 2   3 B 3   p B p ,

(6) Non seasonal moving average ( M .A . ) : θ q ( B ) = 1 θ q B θ q B 2 θ q B 3  θ q B q ,

(7) Seasonal auto regression ( A .R . ) : Φ P ( B s ) = 1 Φ 1 B s Φ 2 B 2 s Φ 3 B 3 s Φ P B P s ,

(8) Seasonal moving average ( M .A . ) : Θ Q ( B s ) = 1 Θ 1 B s Θ Q B 2 s Θ Q B 3 s Θ Q B Q s ,


(9) ( B ) X t = X t s ,

(10) s X t = s ( B ) X t = ( 1  B s ) X t = X t B s X t = X t X t s ,

(11) d ( B ) X t = ( 1 B ) d X t ,

(12) s D ( B ) X t = ( 1 B s ) d X t .

The d and D point out the order of the nonseasonal and seasonal differencing, and their values are usually less than 1 and 2 totals of seasonal difference, respectively (i.e., 0 d; D 1) [49]. Fitting the data of the SARIMA model begins with selecting the normal and periodic differencing schemes. Auto-correlation function selects the best model orders, partial auto-correlation function, and Akaike’s information criteria [50,51,52].

2.5 Performance metrics

Predicting outputs rely most on estimating and comparing the actual values to the predicted values. These indicators of the accuracy measurement parameters include the MAE, RMSE, R 2, and R. In addition, the error metrics are defined as follows [53,54,55]:

(13) MAE  = t = 1 n x t x t n ,

(14) RMSE = t = 1 n ( x t x t ) 2 n ,

(15) R 2 = 1 t = 1 n ( x t x t ) 2 t = 1 n x t 1 n t = 1 n x t 2 ,

(16) R = t = 1 n ( x t x ¯ ) ( x t x ¯ ) t = 1 n ( x t x ¯ ) 2 t = 1 n ( x t x ¯ ) 2 ,

where x t and x t are estimated and observed outputs in the period time t, and n is the number of the observed values in the testing data. x ¯ and x ¯   are the mean of the predicted and observed outputs. The R 2 and R should be approaching 1 to indicate strong model performance, and the MAE and RMSE should be as close to zero as possible.

The diagram describes the methodology of this study in Figure 3. First, the collected data on agriculture and aquaculture land uses are preprocessed and checked by statistical methods in the input layer. Next, the data is separated into first quarter to thirty-second quarter for the training phase, and from thirty-third quarter to forty-fourth quarter for the testing phase; simultaneously, the SVR, ANN, and SARIMA algorithms are applied to learn the training samples and choose the optimal network parameters in the process of the methods. Last, the three models’ implementations showed the basis functions and were compared using the accuracy measurement parameters, namely RMSE, MAE, R, and R 2, in the possible result stage and looking for the most fitted estimation model for the study in the output layer.

Figure 3 
                  Diagram of the research steps used in this study.
Figure 3

Diagram of the research steps used in this study.

3 Results

3.1 Database collection

This study uses a database containing land-use change in 44 quarters from 2010 to 2020. Two input variables, agriculture (Agri) land use and aquaculture (Aqua) land use, were gathered from five provinces at the Department of Natural Resources and Environment (DONRE). The typical statistical outcomes are also highlighted in Table 1. The Mean, Min, Max, St Dev, Skew, and Kurt values were calculated from the recorded quarters per year. Five provinces’ input data patterns are chosen in two stages. The first section is utilized for the training phase, which lasted from the first quarter to the thirty-second quarter. The testing phase, which lasted from the thirty-third quarter to the forty-fourth quarter, is the second section. Moreover, this division supports the models to optimize the accuracy parameters, and the univariate time series data forecasting fits the SARIMA model.

Table 1

Descriptive statistics of land use in five provinces

Parameter Quangbinh Quangtri ThuaThienHue Danang Quảngnam
Agri Aqua Agri Aqua Agri Aqua Agri Aqua Agri Aqua
Kurtosis −0.73 −0.51 −0.62 −0.22 −0.61 −0.87 0.85 1.19 −0.42 −0.52
Skewness −1.23 −1.51 −1.64 −1.50 −1.66 0.07 −0.88 0.01 −1.71 −1.63
St Dev 4,631 263 14,931 150 4,418 81 605 33 46,793 102
Mean (ha) 87,133 3,157 109,946 2,851 65,561 6,012 7,222 147 173,836 3,582
Min (ha) 79,222 2,763 88,282 2,651 58,815 5,835 6,641 117 113,495 3,431
Max (ha) 91,858 3,468 122,583 3,058 69,258 6,133 8,295 211 220,104 3,692

3.2 Experimental results

Find the best ANN, SVR, and SARIMA models for the land use of five provinces after many tests. With the values of the fundamental parameters shown in Table 2, this investigation discovered the best model.

Table 2

Basic configuration of ANN, SVR, SARIMA models

Model Item Quangbinh Quangtri ThuaThienHue Danang Quảngnam
Agri config Aqua config Agri config Aqua config Agri config Aqua config Agri config Aqua config Agri config Aqua config
ANN Model type Seq Seq Seq Seq Seq Seq Seq Seq Seq Seq
Layer type Dense Dense Dense Dense Dense Dense Dense Dense Dense Dense
Number of inputs 32 32 32 32 32 32 32 32 32 32
Number of hidden 2 2 2 2 2 2 2 2 2 2
Hidden layer size 200/200 100/100 200/200 100/100 200/200 100/100 200/200 100/100 200/200 100/100
Number of output 1 1 1 1 1 1 1 1 1 1
Activation function Relu Relu Relu Relu Relu Relu Relu Relu Relu Relu
Optimizer control Adam Adam Adam Adam Adam Adam Adam Adam Adam Adam
Epoch 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
SVR Kernel Linear Linear Linear Linear Linear Linear Linear Linear Linear Linear
C 100 1,000 100 1,000 100 1,000 100 1,000 100 1,000
Gamma 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05
Epsilon 1.98 2.98 1.98 2.98 1.98 2.98 1.98 2.98 1.98 2.98
SARIMA p, d, q 0,1,0 0,1,0 0,1,0 0,1,0 0,1,0 0,1,0 0,1,0 0,1,0 0,1,0 0,1,0
P, D, Q, s 0,1,0,4 0,1,1,4 1,1,1,4 0,1,1,4 0,1,0,4 1,1,1,4 0,1,1,4 1,1,0,4 1,1,1,4 0,1,1,4
Akaike information criterion 682 461 777 411 700 332 558 302 866 391
Bayesian information criterion 681 464 782 415 701 337 561 305 863 395

Note: Seq, Sequential; config, configuration.

3.3 Analysis and comparison of simulation results among ANN, SVR, and SARIMA models

With regard to the obtained results from the data prediction for land use in Table 3 and Figure 4, it can be verified that the best forecasting models are in Quangtri province. The second-best models are ThuaThienHue province. Danang, Quangbinh, and Quảngnam provinces show the predicted reliabilities to gain third, fourth, and fifth positions, respectively. The specific forecast results for two types of land use in each location are as follows. The SVR models for agriculture land use (with R 2 = 0.95, MAE = 1.03 [ha], and RMSE = 1.03 [ha]) and aquaculture land use (with R 2 = 0.93, MAE = 4.87 [ha], and RMSE = 5.01 [ha]) supply a better fit than the other models for Quangtri. In the implementation of the other models, the hierarchy shows to consider the order of SVRAgriculture_land_use > ANNAgriculture_land_use > SARIMAAgriculture_land_use, and SVRAgriculture_land_use > ANNAquacultute_land_use > SARIMAAquaculture_land_use. The land-use prediction for ThuaThienHue province shows that the SVRAgricultute_land_use and SVRAquacultute_land_use models are the best forecasting accuracy, and the hierarchical order of other models are ANNAgriculture_land_use > SARIMAAquaculture_land_use and ANNAquacultute_land_use > SARIMAAquaculture_land_use. The forecasting land-use models for Quangbinh and Danang indicating hierarchical order are SVRAgricultute_land_use > ANNAgriculture_land_use > SARIMAAgriculture_land_use and SVRAquaculture_land_use > ANNAquacultute_land_use > SARIMAAquaculture_land_use for Quangbinh, and SVRAgricultute_land_use > ANNAgriculture_land_use > SARIMAAgriculture_land_use and SVRAquaculture_land_use > ANNAquacultute_land_use > SARIMAAquaculture_land_use for Danang, respectively. At the same time, the model of SARIMAAquaculture_land_use with MAE = 39.66 (ha), RMSE = 85.44 (ha), and R 2 = 0.67 for Quangbinh is also the worst prediction compared to others. The prediction result of land use in Quảngnam shows that the order of hierarchical models with the accuracy of performance is SVRAgricultute_land_use > ANNAgriculture_land_use > SARIMAAgriculture_land_use and SVRAquaculture_land_use > ANNAquacultute_land_use > SARIMAAquaculture_land_use. In addition, the forecasting of SARIMAAgriculture_land_use model for Quảngnam gives the lowest accurate level with MAE, RMSE, R 2, and R values are 480 (ha), 610 (ha), 0.61, and 0.75, respectively.

Table 3

Accuracy parameters for agriculture land-use and aquaculture land-use prediction using three models across five provinces

Province Parameter Agriculture land use Aquaculture land use
Quangbinh MAE (ha) 3.41 85.75 36.97 3.87 39.66 13.40
RMSE (ha) 3.45 51.58 37.1 3.96 85.44 13.42
R 2 0.91 0.72 0.93 0.92 0.67 0.93
R 0.92 0.85 0.94 0.92 0.73 0.92
Quangtri MAE (ha) 1.03 293.94 67.16 4.87 10.41 9.91
RMSE (ha) 1.03 649.75 67.51 5.01 25.32 9.91
R 2 0.95 0.71 0.94 0.93 0.81 0.93
R 0.94 0.68 0.92 0.94 0.89 0.94
ThuaThienHue MAE (ha) 1.73 61.64 43.95 3.44 9.39 10.12
RMSE (ha) 1.72 166.84 43.94 3.63 25.01 15.29
R 2 0.94 0.72 0.92 0.94 0.87 0.88
R 0.95 0.68 0.93 0.92 0.84 0.91
Danang MAE (ha) 3.01 178.31 42.23 8.75 9.01 12.52
RMSE (ha) 3.09 502.31 45.35 10.43 21.45 13.8
R 2 0.93 0.82 0.94 0.86 0.83 0.84
R 0.94 0.73 0.95 0.89 0.79 0.81
Quangnam MAE (ha) 1.35 480 44.03 2.31 9.21 13.08
RMSE (ha) 1.38 610 45.15 2.45 21.27 13.11
R 2 0.94 0.61 0.91 0.93 0.72 0.91
R 0.94 0.75 0.93 0.94 0.81 0.87

Furthermore, the forecasting implementation of the models is also visualized and considered by the Taylor diagram. Taylor's diagram summarizes the St Dev and correlation coefficient (CC) that are comprised concomitantly in the assessment of the respective model [56,57]. The St Dev and CC between the observed and predicted datasets for all the land-use models of the provinces are described in the Taylor diagram, as shown in Figure 5. It may be observed for ANN, SVR, and SARIMA models in Quangbinh (CCSVR_agricuture_land_use = 0.92, CCSARIMA_agriculture_land_use = 0.85, CCANN_agriculture_land_use = 0.94, CCSVR_aquacuture_land_use = 0.92, CCSARIMA_aquaculture_land_use = 0.79, and CCANN_aquaculture_land_use = 0.92), in Quangtri (CCSVR_agricuture_land_use = 0.94, CCSARIMA_agriculture_land_use = 0.68, CCANN_agriculture_land_use = 0.92, CCSVR_aquacuture_land_use = 0.94, CCSARIMA_aquaculture_land_use = 0.89, and CCANN_aquaculture_land_use = 0.94), in ThuaThienHue (CCSVR_agricuture_land_use = 0.95, CCSARIMA_agriculture_land_use = 0.68, CCANN_agriculture_land_use = 0.93, CCSVR_aquacuture_land_use = 0.92, CCSARIMA_aquaculture_land_use = 0.84, and CCANN_aquaculture_land_use = 0.91), in Danang (CCSVR_agriculture_land_use = 0.94, CCSARIMA_agriculture_land_use = 0.73, CCANN_agriculture_land_use = 0.95, CCSVR_aquaculture_land_use = 0.89, CCSARIMA_aquaculture_land_use = 0.73, and CCANN_aquaculture_land_use = 0.81), and in Quảngnam (CCSVR_agriculture_land_use = 0.94, CCSARIMA_agriculture_land_use = 0.75, CCANN_agriculture_land_use = 0.93, CCSVR_aquaculture_land_use = 0.94, CCSARIMA_aquaculture_land_use = 0.81, and CCANN_aquaculture_land_use = 0.87). This diagram demonstrates that these models were optimal accuracies of almost all models’ outcomes were significantly closer to 1. Moreover, it can result in overestimation when the SD of the observation values is lower than the SD of estimation values and vice versa. On the other hand, the SARIMA model for agriculture land-use prediction with CCSARIMA_agriculture_land_use = 0.68 in Quangtri, CCSARIMA_agriculture_land_use = 0.68 in ThuaThienHue point out that the level of gaining accuracy is only above medium; at the same time, the lowest estimations for agriculture and aquaculture land-use changes are SARIMAAgriculture_land_use model with CC for Quảngnam, and model of SARIMAAquaculture_land_use for Quangbinh, respectively.

4 Discussions

The central coastal area has a slower socioeconomic development rate than that of Hanoi and Ho Chi Minh cities. Therefore, agricultural land use and aquatic land use in agriculture and aquaculture are still crucial sources of income for the people living in the region. The urbanization and industrialization process gradually narrow down the land-use squares in recent years. Therefore, the assessment, analysis, and forecasting of land use are very important for local authorities and stakeholders.

This study used SVR, SARIMA, and ANN models for agriculture and aquaculture land-use change in the Central Coast Region of Vietnam to assess the impact of urbanization speed and industrialization processing on the two types of land use. Three algorithms showed highly accurate results for estimating the land-use change, in which the SVR and ANN models indicated more accuracy for Quangbinh, Quangtri, ThuaThienHue, Danang, and Quảngnam, respectively, compared with the SARIMA model. In addition, the predicted values of land-use changes created by the SARIMA model also supply acceptable results.

In order to evaluate the level of predicted accuracy of these results, this study is compared with other studies based on the accurate parameters. Jamali (2019) [58] used ANN to predict land-use/land-cover mapping using Landsat 8 O.L.I. in the northern region of Iran. The RMSE and MAE values for the testing phase were 6.0 and 6.0, respectively, and the values of the two parameters are lower than this study’s results for ANN models. Talukdar et al. [59] employed ANN and SVR models to predict the land-use land-cover classification of the river Ganga from Rajmahal to Farakka barrage in India. The RMSE values of the two models were 0.09 and 0.11 for ANN and SVR, respectively. The parameter values are lower than this study’s results, in which the lowest values of accurate parameters are 9.91 and 1.03 for ANN and SVR models. Soltani et al. [60] deployed the SARIMA model to estimate lake surface area in the Tashk-Bakhtegan lakes, Iran. The simulation result showed that the RMSE = 70.217, MAE = 49.425, and the indicator values are equivalent to this study result.

Although classification accuracies for land use are still important for the provinces, these results may support the provinces’ authorities and other stakeholders in decision-making and planning regarding two types of land use. The result can help the Ministry of Natural Resources and Environment to carry out the land-uses inventory and upgrade agriculture and aquaculture land-use change maps every 5 years. At the same time, the DONREs use the predicting database and update the kinds of land-use change manually. Moreover, evaluating the land-uses change will give policymakers regarding labor reorganization in the agriculture and aquaculture fields, also changing production methodology.

5 Conclusions

In this study, the SVR, SARIMA, and ANN models are presented to forecasting land-use change in Quangbinh, Quangtri, ThuaThienHue, Danang, and Quảngnam provinces of Vietnam. The observed data consisting of agriculture and aquaculture land use were collected from the start of 2010 to the end of 2020 for prediction. The data is also divided into two phases as the training and testing phases occupied from first quarter to thirty-second quarter and from thirty-third quarter to forty-fourth quarter, respectively. The studied and observed values are compared using four widely used statistical indicators (i.e., RMSE, MAE, R, and R 2). The study results show that almost SVR models improve the accurate performance more than ANN and SARIMA in Quangtri, ThuaThienHue, Danang, and Quảngnam provinces. The ANN model indicates more accurate forecasting implementation than the SVR and SARIMA models in Quangbinh province. Hence, the SVR and ANN models may be successfully precited for better land-use change. Furthermore, predicting land-use change can support land planning and decision-making for the local authorities.

tel: +8490-515-3333


We are thankful to the Department of Civil Engineering at the National Kaohsiung University of Science and Technology, Taiwan, and Thu Dau Mot University in Vietnam.

  1. Author contributions: Conceptualization, material and methods, writing, original draft preparation: Nguyen Hong Giang; writing, review, editing: YuRen Wang.

  2. Conflict of interest: The authors declare no conflict of interest.

  3. Data availability statement: The data used to support the findings of this study are available from the corresponding author upon request.


[1] Verburg PH, Soepboer W, Veldkamp A, Limpiada R, Espaldon V, Mastura SS. Modeling the spatial dynamics of regional land use: the CLUE-S model. Environ Manag. 2002;30:391–405.10.1007/s00267-002-2630-xSearch in Google Scholar PubMed

[2] Tayyebi AH, Delavar MR, Tayyebi A, Golobi M. Combining multi criteria decision making and Dempster Shafer theory for landfill site selection. Int Arch Photogramm Remote Sens Spat Inf Sci. 2010;38:1073–8.Search in Google Scholar

[3] Geist H, McConnell W, Lambin EF, Moran E, Alves D, Rudel T. Causes and trajectories of land-use/cover change. Land-use and land-cover change. Berlin, Heidelberg: Springer; 2006. p. 41–70.10.1007/3-540-32202-7_3Search in Google Scholar

[4] Godfray HCJ, Beddington JR, Crute IR, Haddad L, Lawrence D, Muir JF, et al. Food security: the challenge of feeding 9 billion people. Science. 2010;327(5967):812–8.10.1126/science.1185383Search in Google Scholar PubMed

[5] Li MW, Wang YT, Geng J, Hong WC. Chaos cloud quantum bat hybrid optimization algorithm. Nonlinear Dyn. 2021;103(1):1167–93.10.1007/s11071-020-06111-6Search in Google Scholar

[6] Veldkamp A, Fresco LO. Exploring land use scenarios, an alternative approach based on actual land use. Agric Syst. 1997;55(1):1–17.10.1016/S0308-521X(95)00079-KSearch in Google Scholar

[7] National Research Council, & Climate Research Committee. Radiative forcing of climate change: Expanding the concept and addressing uncertainties. National Academies Press; 2005.Search in Google Scholar

[8] Board SS, National Research Council. Earth science and applications from space: national imperatives for the next decade and beyond. Washington DC, USA: National Academies Press; 2007.Search in Google Scholar

[9] Grădinaru SR, Fan P, Iojă CI, Niță MR, Suditu B, Hersperger AM. Impact of national policies on patterns of built-up development: an assessment over three decades. Land Use Policy. 2020;94:104510.10.1016/j.landusepol.2020.104510Search in Google Scholar

[10] Fan P, Ouyang Z, Nguyen DD, Nguyen TTH, Park H, Chen J. Urbanization, economic development, environmental and social changes in transitional economies: Vietnam after Doimoi. Landsc Urban Plan. 2019;187:145–55.10.1016/j.landurbplan.2018.10.014Search in Google Scholar

[11] Rutten M, Van Dijk M, Van Rooij W, Hilderink H. Land use dynamics, climate change, and food security in Vietnam: a global-to-local modeling approach. World Dev. 2014;59:29–46.10.1016/j.worlddev.2014.01.020Search in Google Scholar

[12] Khanh Nguyen HT, Nang Thu TT, Lebailly P, Azadi H. Economic challenges of the export-oriented aquaculture sector in Vietnam. J Appl Aquaculture. 2019;31(4):367–83.10.1080/10454438.2019.1576568Search in Google Scholar

[13] Wang SW, Gebru BM, Lamchin M, Kayastha RB, Lee WK. Land use and land cover change detection and prediction in the Kathmandu district of Nepal using remote sensing and G.I.S. Sustainability. 2020;12(9):3925.10.3390/su12093925Search in Google Scholar

[14] Liping C, Yujun S, Saeed S. Monitoring and predicting land use and land cover changes using remote sensing and G.I.S. techniques – A case study of a hilly area, Jiangle, China. PLoS One. 2018;13:e0200493.10.1371/journal.pone.0200493Search in Google Scholar PubMed PubMed Central

[15] Saputra MH, Lee HS. Prediction of land use and land cover changes for north Sumatra, Indonesia, using an artificial neural network based cellular automaton. Sustainability. 2019;11(11):3024.10.3390/su11113024Search in Google Scholar

[16] Veldkamp A, Lambin EF. Predicting land-use change. Agriculture, ecosystems & environment. 2001;85(1–3):1–6.10.1016/S0167-8809(01)00199-2Search in Google Scholar

[17] Viney NR, BormANN H, Breuer L, Bronstert A, Croke BFW, Frede H, et al. Assessing the impact of land use change on hydrology by ensemble modelling (L.U.C.H.E.M.) II: Ensemble combinations and predictions. Adv Water Resour. 2009;32(2):147–58.10.1016/j.advwatres.2008.05.006Search in Google Scholar

[18] Su L, Zhu JH, Wang W, Liu M. Application of A.R.M.A. Model in prediction of land use demand take Farmland in Jin-Hu Coastal Area as Example. Hunan Agric Sci. 2012;5:61–3Search in Google Scholar

[19] Zhang P, Ke Y, Zhang Z, Wang M, Li P, Zhang S. Urban land use and land cover classification using novel deep learning models based on high spatial resolution satellite imagery. Sensors. 2018;18(11):3717.10.3390/s18113717Search in Google Scholar PubMed PubMed Central

[20] Luus FP, Salmon BP, Van den Bergh F, Maharaj BTJ. Multiview deep learning for land-use classification. IEEE Geosci Remote Sens Lett. 2015;12(12):2448–52.10.1109/LGRS.2015.2483680Search in Google Scholar

[21] Mu L, Wang L, Wang Y, Chen X, Han W. Urban land use and land cover change prediction via self-adaptive cellular based deep learning with multisourced data. IEEE J Sel Top Appl Earth Obs Remote Sens. 2019;12(12):5233–47.10.1109/JSTARS.2019.2956318Search in Google Scholar

[22] Azad A, Wang X. Land use change ontology and traffic prediction through recurrent neural networks: A case study in Calgary, Canada. ISPRS Int J Geo-Information. 2021;10(6):358.10.3390/ijgi10060358Search in Google Scholar

[23] Malarvizhi K, Kumar SV, Porchelvan P. Urban sprawl modelling and prediction using regression and Seasonal A.R.I.M.A.: A case study for Vellore. India Model Earth Syst Environ. 2022;8(2):1597–615.10.1007/s40808-021-01170-zSearch in Google Scholar

[24] Rizeei HM, Pradhan B, Saharkhiz MA. Surface runoff prediction regarding L.U.L.C. and climate dynamics using coupled LTM, optimized A.R.I.M.A., and GIS-based SCS-CN models in tropical region. Arab J Geosci. 2018;11:1–16.10.1007/s12517-018-3397-6Search in Google Scholar

[25] Tayyebi A, Pijanowski BC. Modeling multiple land use changes using ANN, CART and MARS: Comparing tradeoffs in goodness of fit and explanatory power of data mining tools. Int J Appl Earth Obs Geoinf. 2014;28:102–16.10.1016/j.jag.2013.11.008Search in Google Scholar

[26] Bounouh O, Essid H, Farah IR. Prediction of land use/land cover change methods: A study. 2017 International Conference on Advanced Technologies for Signal and Image Processing (A.T.S.I.P.). Fez, Morocco: IEEE; 2017, May. p. 1–7.10.1109/ATSIP.2017.8075511Search in Google Scholar

[27] Lin C, Wang K, Wu D, Gong B. Passenger flow prediction based on land use around metro stations: A case study. Sustainability. 2020;12(17):6844.10.3390/su12176844Search in Google Scholar

[28] Wang X, Wang Z, Jin G, Yang J. Land reserve prediction using different kernel based support vector regression. Trans Chin Soc Agric Eng. 2014;30(4):204–11.Search in Google Scholar

[29] Huang B, Xie C, Tay R, Wu B. Land-use-change modeling using unbalanced support-vector machines. Environ Plan B: Plan Des. 2009;36:398–416.10.1068/b33047Search in Google Scholar

[30] Zhang Z, Hong WC. Application of variational mode decomposition and chaotic grey wolf optimizer with support vector regression for forecasting electric loads. Knowl Syst. 2021;228:107297.10.1016/j.knosys.2021.107297Search in Google Scholar

[31] Jiang B, Liang S, Wang J, Xiao Z. Modeling MODIS LAI time series using three statistical methods. Remote Sens Environ. 2010;114(7):1432–44.10.1016/j.rse.2010.01.026Search in Google Scholar

[32] Li Y, Campbell EP, Haswell D, Sneeuwjagt RJ, Venables WN. Statistical forecasting of soil dryness index in the southwest of Western Australia. For Ecol Manag. 2003;183(1–3):147–57.10.1016/S0378-1127(03)00103-8Search in Google Scholar

[33] Kong C, Lan H, Yang G, Xu K. Geo-environmental suitability assessment for agricultural land in the rural–urban fringe using BPNN and GIS: a case study of Hangzhou. Environ Earth Sci. 2016;75(15):1–14.10.1007/s12665-016-5956-zSearch in Google Scholar

[34] Zhang XR, Chen G. Application of neural network in urban land use suitability evaluation. Key Engineering Materials. Vol. 474. Chengdu, China: Trans Tech Publications Ltd; 2011. p. 681–6.10.4028/ in Google Scholar

[35] Nguyen TV, Tran TQ. Forestland and rural household livelihoods in the North Central Provinces, Vietnam. Land Use Policy. 2018;79:10–9.10.1016/j.landusepol.2018.07.046Search in Google Scholar

[36] Ang LI, Li YX, Li XH. Tensorflow and Keras-based convolutional neural network in cat image recognition. DEStech Transactions on Computer Science and Engineering, (CMSAM); 2017. p. 5.10.12783/dtcse/cmsam2017/16428Search in Google Scholar

[37] Jiang Z, Shen G. )Prediction of house price based on the back propagation neural network in the keras deep learning framework. 2019 6th International Conference on Systems and Informatics (I.C.S.A.I.). Shanghai, China: IEEE; 2019, November. p. 1408–12.10.1109/ICSAI48974.2019.9010071Search in Google Scholar

[38] Samarawickrama AJP, Fernando TGI. A recurrent neural network approach in predicting daily stock prices an application to the Sri Lankan stock market. 2017 IEEE International Conference on Industrial and Information Systems (I.C.I.I.S.). Peradeniya, Sri Lanka: IEEE; 2017, December. p. 1–6.10.1109/ICIINFS.2017.8300345Search in Google Scholar

[39] Lv C, Xing Y, Zhang J, Na X, Li Y, Liu T, et al. Levenberg–Marquardt backpropagation training of multilayer neural networks for state estimation of a safety-critical cyber-physical system. IEEE Trans Ind Inform. 2017;14:3436–46.10.1109/TII.2017.2777460Search in Google Scholar

[40] Du YC, Stephanus A. Levenberg-Marquardt neural network algorithm for degree of arteriovenous fistula stenosis classification using a dual optical photoplethysmography sensor. Sensors. 2018;18:2322.10.3390/s18072322Search in Google Scholar PubMed PubMed Central

[41] Sharma S, Sharma S. Activation functions in neural networks. Towards Data Sci. 2017;6(12):310–6.10.33564/IJEAST.2020.v04i12.054Search in Google Scholar

[42] Feng J, Lu S. Performance analysis of various activation functions in artificial neural networks. Journal of Physics: Conference Series. Vol. 1237, Issue 2. Xi’an, China: I.O.P. Publishing; 2019, June. p. 022030.10.1088/1742-6596/1237/2/022030Search in Google Scholar

[43] Xiang Y, Gou L, He L, Xia S, Wang W. A SVR–ANN combined model based on ensemble E.M.D. for rainfall prediction. Appl Soft Comput. 2018;73:874–83.10.1016/j.asoc.2018.09.018Search in Google Scholar

[44] Dash CSK, Sahoo P, Dehuri S, Cho SB. An empirical analysis of evolved radial basis function networks and support vector machines with mixture of kernels. Int J Artif Intell Tools. 2015;24(4):1550013.10.1142/s021821301550013xSearch in Google Scholar

[45] García-Floriano A, López-Martín C, Yáñez-Márquez C, Abran A. Support vector regression for predicting software enhancement effort. Inf Softw Technol. 2018;97:99–109.10.1016/j.infsof.2018.01.003Search in Google Scholar

[46] Kardakos EG, Alexiadis MC, Vagropoulos SI, Simoglou CK, Biskas PN, Bakirtzis AG. Application of time series and artificial neural network models in short-term forecasting of P.V. power generation. 2013 48th International Universities’ Power Engineering Conference (U.P.E.C.). Dublin, Ireland: IEEE; 2013, September. p. 1–6.10.1109/UPEC.2013.6714975Search in Google Scholar

[47] Mao Q, Zhang K, Yan W, Cheng C. Forecasting the incidence of tuberculosis in China using the seasonal auto-regressive integrated moving average (SARIMA) model. J Infect public health. 2018;11(5):707–12.10.1016/j.jiph.2018.04.009Search in Google Scholar PubMed PubMed Central

[48] Vagropoulos SI, Chouliaras GI, Kardakos EG, Simoglou CK, Bakirtzis AG. Comparison of SARIMAX., SARIMA, modified SARIMA and ANN-based models for short-term P.V. generation forecasting. 2016 IEEE International Energy Conference (ENERGYCON). Leuven, Belgium: IEEE; 2016, April. p. 1–6.10.1109/ENERGYCON.2016.7514029Search in Google Scholar

[49] Farsi M, Hosahalli D, Manjunatha BR, Gad I, Atlam ES, Ahmed A, et al. Parallel genetic algorithms for optimizing the SARIMA model for better forecasting of the NCDC weather data. Alex Eng J. 2021;60(1):1299–1316.10.1016/j.aej.2020.10.052Search in Google Scholar

[50] He X, Xu H, Sabetamal H, Sheng D. Machine learning aided stochastic reliability analysis of spatially variable slopes. Comput Geotech. 2020;126:103711.10.1016/j.compgeo.2020.103711Search in Google Scholar

[51] Maxwell O, Chinedu IU, Ifeanyi AC, Chinonso NC. On modeling murder crimes in Nigeria. Sci Rev. 2019;5:157–62.10.33552/SJRR.2019.02.000528Search in Google Scholar

[52] Li HL, Zheng RJ, Zheng Q, Jiang W, Zhang XL, Wang WM, et al. Predicting the number of visceral leishmaniasis cases in Kashgar, Xinjiang, China using the ARIMA-EGARCH model. Asian Pac J Trop Med. 2020;13(2):81.10.4103/1995-7645.275416Search in Google Scholar

[53] Kardani N, Zhou A, Nazem M, Shen SL. Estimation of bearing capacity of piles in cohesionless soil using optimised machine learning approaches. Geotech Geol Eng. 2020;38(2):2271–91.10.1007/s10706-019-01085-8Search in Google Scholar

[54] Touzani S, Granderson J, Fernandes S. Gradient boosting machine for modeling the energy consumption of commercial buildings. Energy Build. 2018;158:1533–43.10.1016/j.enbuild.2017.11.039Search in Google Scholar

[55] Yang JH, Yang MS. A control chart pattern recognition system using a statistical correlation coefficient method. Comput Ind Eng. 2005;48(2):205–21.10.1016/j.cie.2005.01.008Search in Google Scholar

[56] Taylor KE. Summarizing multiple aspects of model performance in a single diagram. J Geophys Res Atmos. 2001;106(D7):7183–92.10.1029/2000JD900719Search in Google Scholar

[57] Ghorbani MA, Deo RC, Yaseen ZM, Kashani MH, Mohammadi B. Pan evaporation prediction using a hybrid multilayer perceptron-firefly algorithm (MLP-FFA) model: case study in North Iran. Theor Appl Climatol. 2018;133:1119–31.10.1007/s00704-017-2244-0Search in Google Scholar

[58] Jamali A. Evaluation and comparison of eight machine learning models in land use/land cover mapping using Landsat 8 OLI: A case study of the northern region of Iran. SN Appl. Sci. 2019;1:1448.10.1007/s42452-019-1527-8Search in Google Scholar

[59] Talukdar S, Singha P, Mahato S, Pal S, Liou YA, Rahman A. Land-use land-cover classification by machine learning classifiers for satellite observations – A review. Remote Sens. 2020;12(7):1135.10.3390/rs12071135Search in Google Scholar

[60] Soltani K, Azari A, Zeynoddin M, Amiri A, Ebtehaj I, Ouarda TB, et al. Lake surface area forecasting using integrated satellite-SARIMA-long-short-term memory model. Res. Sq. 2021.10.21203/ in Google Scholar

Received: 2021-09-02
Revised: 2022-07-02
Accepted: 2022-10-05
Published Online: 2022-12-31

© 2022 the author(s), published by De Gruyter

This work is licensed under the Creative Commons Attribution 4.0 International License.

Downloaded on 2.12.2023 from
Scroll to top button