Assessment of Land Suitability Potential Using Ensemble Approaches of Advanced Multi-Criteria Decision Models and Machine Learning for Wheat Cultivation

Abstract

Land suitability assessment, as an important process in modern agriculture, involves the evaluation of numerous aspects such as soil properties, climate, relief, hydrology and socio-economic aspects. The aim of this study was to evaluate the suitability of soils for wheat cultivation in the Gavshan region, Iran, as the country is facing the task of becoming self-sufficient in wheat. Various methods were used to evaluate the land, such as multi-criteria decision-making (MCDM), which is proving to be important for land use planning. MCDM and machine learning (ML) are useful for decision-making processes because they use complicated spatial data and methods that are widely available. Using a geomorphological map, seventy soil profiles were selected and described, and ten soil properties and wheat yields were determined. Three MCDM approaches, including the technique of preference ordering by similarity to the ideal solution (TOPSIS), gray relational analysis (GRA), and simple additive weighting (SAW), were used and evaluated. The criteria weights were extracted using Shannon’s entropy method. Random forest (RF) model and auxiliary variables (remote sensing data, terrain data, and geomorphological maps) were used to represent the land suitability values. Spatial autocorrelation analysis as a statistical method was applied to analyze the spatial variability of the spatial data. Slope, CEC (cation exchange capacity), and OC (organic carbon) were the most important factors for wheat cultivation. The spatial autocorrelation between the key criteria (slope, CEC, and OC) and wheat yield confirmed these results. These results also showed a significant correlation between the land suitability values of TOPSIS, GRA, and SAW and wheat yield (0.74, 0.72, and 0.57, respectively). The spatial distribution of land suitability values showed that the areas classified as good according to TOPSIS and GRA were larger than those classified as moderate and weak according to the SAW approach. These results were also confirmed by the autocorrelation of the MCDM techniques with wheat yield. In addition, the RF model showed its effectiveness in processing complex spatial data and improved the accuracy of land suitability assessment. In this study, by integrating advanced MCDM techniques and ML, an applicable land evaluation approach for wheat cultivation was proposed, which can improve the accuracy of land suitability and be useful for considering sustainability principles in land management.

1. Introduction

One of the most important issues facing humanity is food and health security [1]. The growing world population requires more food, while land resources are decreasing due to population growth and some destructive environmental problems such as climate change, land degradation, and urbanization. Therefore, concern about land capacity is increasing [2].

In many developing countries, land managers use the land for different crops without paying attention to the possibilities and principles of sustainability [3,4]. The unplanned changes in land use have led to destructive phenomena such as deforestation, desertification, global warming, and soil erosion. Therefore, soils in developing countries should be utilized according to their potential in order to increase yield production and conserve soil resources.

Land suitability is determined by the land use planning process [5], and land evaluation is an important step in this process. Land suitability assessment requires a potential evaluation of environmental characteristics such as climate, soil, relief, water, social, and economic factors [6]. Through a detailed land suitability assessment, the most suitable crops or land uses for specific regions can be identified [7,8,9,10].

Until 1973, the systems introduced to classify land suitability or capability were generally used to evaluate land for different uses [11]. Some of these methods have been applied in industrialized countries, but the accessible information on these land uses is sometimes unrelated to regional knowledge and conditions [12,13]. In recent decades, research has focused on the use of multi-attribute models for complex decision-making related to land suitability. Existing multi-criteria decision-making (MCDM) and geographic information system (GIS) techniques allow the combination of different information to manage and design applications.

MCDM, which was developed in the 1960s, is widely used in various fields [14]. MCDM is a framework to generate the best choice for complicated circumstances and helps developers and policymakers select the best solutions based on their constraints [15,16,17]. Determining the priority of effective factors in evaluation and decision-making is always considered one of the basic principles of planning and management. In these decisions, multiple measurement attributes can be used instead of one attribute. This method involves a series of techniques that allow a number of independent attributes to be evaluated, weighted, and then ranked.

MCDM methods have a great potential to reduce costs and increase accuracy in spatial decisions, as well as to provide an applicable solution for problem-solving [16,18]. In this context, techniques such as the technique of preference ordering by similarity to the ideal solution (TOPSIS), gray relational analysis (GRA), and simple additive weighting (SAW) are commonly used [19,20,21]. These methods are a good tool for solving challenging problems with complicated decision conditions in a variety of sectors.

On the other hand, machine learning (ML) methods have been increasingly developed as part of the knowledge of artificial intelligence in the age of technology and information in various scientific fields, especially in digital mapping [22,23,24]. One of the main applications of ML methods is the identification and prediction of existing patterns in several large datasets, such as satellite data, parameters derived from digital elevation models, spectroscopy data, or other sources [25]. Specifically, in the area of land suitability, several studies have successfully applied ML techniques to assess and predict land suitability for various purposes. For example, Ahmed and Hussain [26] applied ML to predict wheat production in northern Pakistan. Their study shows the potential of machine learning in improving agricultural forecasting for developing countries like Pakistan. Similarly, Naghdyzadegan Jahromi [27] used ML models to predict wheat yields based on environmental variables. Their models showed promising results.

The integration of ML algorithms and MCDM methods as an applicable and useful approach has recently been used by researchers in various fields [28,29,30], although there is little research in this regard in the field of land evaluation [28,29]. In determining site suitability, a combination of ML and MCDM can be used to evaluate various parameters, make appropriate land use decisions, and increase the accuracy of the assessment. Saha and Mondal [28] evaluated the suitability of agricultural land in India using MCDM and machine learning techniques. They used the Analytic Hierarchy Process for weighting the criteria and compared the effectiveness of methods such as Random Forest (RF) and FCOPRAS for analyzing land suitability. RF demonstrates high model accuracy through ROC curves and accuracy measurements.

Wheat has the largest acreage of any cereal in the world and, as one of the most important agricultural crops, has an impact on global food security and human nutrition. It is an important source of food for a large part of the world’s population. It provides food for more than 33 percent of the world’s population and is considered a strategic commodity. Iran is one of the largest wheat producers in the Middle East region [30]. However, wheat cultivation in Iran faces difficulties as the lands have not been tested for their suitability, and the plan for “food independence” in Iran has been postponed [31]. Therefore, proper land management is crucial for crop planning and requires a close examination of actual characteristics such as soil quality, relief, water, and climate. It can help to increase wheat yield and minimize undue impacts on the environment. Few studies have been conducted to compare advanced MCDA techniques for land suitability assessment. MCDA methods are particularly considered due to their robust and effective algorithms [32]. Furthermore, ML approaches were combined with the MCDA approach for land suitability mapping of wheat to increase the accuracy of the assessment. The aim of this study was to (1) determine the suitability of land for wheat cultivation in western Iran by combining MCDA and ML, (2) compare the advanced MCDA techniques (TOPSIS, SAW, and GRA), and (3) determine the most important factors for land suitability for wheat cultivation in the studied area.

2. Materials and Methods

2.1. Studied Area Description

The research area, whose size is estimated at 5341 hectares, is in the Gavshan region in western Iran (Figure 1). The area is categorized as a cold, semi-arid climate (BWk) according to the Köppen–Geiger climate classification system. Winters are characterized by cold temperatures and heavy rainfall, while summers are quite dry and hot. The average annual precipitation is 370 mm, with more than 90% of it falling between mid-autumn and the end of winter. The average annual temperature is rather cool at 10.8 °C (Figure 2). The soil temperature and humidity regimes are described as Mesic and Xeric [33], respectively. The minimum and maximum altitudes are 1520 and 2366 m above sea level, respectively. The most common soils in the study area belong to the soil reference groups Leptosols, Lithosols, Cambisols, and Calcisols according to the FAO World Reference Base for Soil Resources (WRB 2014) [34] as well as Entisols and Inceptisols according to the USDA soil taxonomy [33]. Cropland (about 80% of the total area) and pastureland are the two main land use categories in this region. Geomorphological features of the area include piedmonts, alluvial fans, hills, and mountains with gentle to steep slopes (Figure 1 and

Table 1. Geomorphology map legend.

Code	Landscape	Landform	Lithology	Geomorphological Surface
Hi111	Hill	Highly eroded hill	Shale and sandstone	Hill with high topography
Hi211		Moderately eroded hill	Limestone and sandstone	Hill with high topography
Mo111	Mountain	Rock outcrop	Limestone	Eroded rock surface
Mo121			Sandstone and conglomerate	Eroded rock surface
Pi111	Piedmont	Alluvial fan	Alluvial fan	Moderate to high slop
Pi121				Moderate slope
Pi211		Pediment	Medium-level alluvial deposits	Low slope
Pi212				Moderate slope
Pl111	Plain	Flat	Recent alluvium	Low slope
Pl211		Alluvial terrace	Medium-level alluvial deposits	Cultivated (moderate to few slopes)

).

2.2. Soil and Yield Sample Collection

The methodological framework applied to this study is shown in Figure 3. Considering that different parameters have an influence on the spatial distribution of soil, a map of geomorphology was created based on the four-level hierarchical method developed by Tommanian [35]. Then, according to the layers of the geomorphology map at the landform level, the locations of 70 soil profiles were determined using a stratified random sampling method. They were then excavated and described using standard methods and relevant soil properties such as soil structure, soil color, flooding, and drainage characteristics [36]. Soil samples were carefully collected from all soil horizons and then transported to the laboratory. In the next step, the organic carbon content was quantified by a wet combustion method [37]. Soil pH and electrical conductivity (EC) were determined by measurements in a saturated paste using a pH electrode [38] and a conductivity meter [39], respectively. Cation exchange capacity (CEC) was determined using the 1-N-ammonium acetate method at a pH of 7.0 [40]. The calcium carbonate equivalent (CCE) was determined volumetrically [41]. The particle size distribution was determined using the Bouyoucos hydrometer method [42]. The percentage of exchangeable sodium (ESP) was calculated as the ratio of sodium to CEC. In addition, the available phosphorus (P) was determined with the extraction agent Bray-1, a solution of 0.03 M ammonium fluoride (NH₄F) and 0.025 M hydrochloric acid (HCl) [43], the available potassium (K) with 1 M ammonium acetate (NH₄OAc) as an extraction agent [44], and the gravel content. The ratio of sodium to CEC was used to calculate the percentage of exchangeable sodium in the soil. The gypsum content was considered negligible and, therefore, excluded from further investigation. In addition, slope and climate data from the Gavshan synoptic weather station [45] and a digital elevation model [46] were used.

2.3. Yield Samples Collection

Wheat samples were systematically taken from each soil profile within a 1 m² to ensure a uniform plant density within the defined area on 23 August 2019. All plant samples were harvested manually and dried at a controlled temperature of 75 °C for 48 h, and the wheat grains were then weighed. The wheat yield was determined for each of the 70 sampling points. The yield (Y) is calculated using the following formula (Equation (1)):

$$\mathrm{Y}=\frac{\left(N_1\times \mathrm{10,000}\right)\times N_2\times \left(\frac{N_3}{1000}\right)}{1000}$$

(1)

Y stands for tons per hectare (t ha⁻¹), N₁ for the number of ears per square meter, N₂ for the number of grains per ear, which are the components of wheat yield, and N₃ for the weight of 1000 grains (in grams).

As a rule, the use of machinery with conventional tillage systems (moldboard plow and disk harrow) [47] on wheat farms ends between 15 October. Wheat was planted on 15 October. In the study area, burning of crop residues is a regular activity. Before sowing, 50 kg ha⁻¹ of triple superphosphate fertilizer was applied to the wheat fields. Before sowing and during the growth of the stalks, 100 kg ha⁻¹ of nitrogen fertilizer (urea) was applied to the wheat fields. The usual herbicides and insecticides are 24D and deltamethrin, respectively. Similar management parameters are used throughout the research region, including machinery, crop rotation, planting density, sowing technique, plant spacing, fertilizer rate, and winter wheat variety.

2.4. Land Suitability Assessment

2.4.1. Diagnostic Attribute

When evaluating land, the first step, which has to do with care, is the selection of the associated attributes. In this study, the suitability of soils for wheat cultivation was assessed on the basis of the following attributes: soil texture, gravel content, soil depth, gypsum, CEC, CCE, ESP (exchange sodium percentage), pH, EC, available phosphorus, available potassium, slope, climatic conditions, susceptibility to flooding, and drainage characteristics. Finally, following the principles for land suitability evaluation formulated by Sys [11], eight attributes were selected for the evaluation of wheat croplands based on a review of the literature and a detailed study of the spatial variability of the parameters along the studied area [12,48,49,50,51,52]. It is worth mentioning that our direct investigations in the studied area showed that this relevance exists. Moreover, these variables proved to be the main constraints for wheat cultivation and are closely related to the specific conditions, soil properties, and soil requirements of the study area.

2.4.2. Diagnostic Attributes Weights

In this phase, the weights of the individual diagnostic attributes were determined using Shannon’s entropy. This method provides a suitable framework for calculating the attribute weights. It was developed by Claude Shannon [53] and is often used to solve the challenge of load determination in MCDM. It assesses the degree of uncertainty or variation within a group of attributes and facilitates the assignment of relative loads to each attribute. In subjective assessment, such as the entropy weighting technique, the subjective weighting is taken directly from the actual data, while the subjective weighting is usually determined by an expert’s judgment based on experience. The advantage of the entropy weighting method is that it increases objectivity while reducing the subjective influence of the decision-makers. Originally, the entropy concept was used to calculate the disorder of a system in the field of thermodynamics. Shannon used it as a strategy to deal with uncertainty by applying it to problems in information theory. Entropy theory states that information can be supplied in proportion to the entropy value. As a result, more weight can be given to the criterion [54]. The idea of entropy weighting has been widely applied in a variety of fields [55,56]. In the following, Shannon’s entropy weight calculation is introduced.

Suppose the decision problem has m options (A1, A2, …, Am) and n criteria (C1, C2, …, Cn).

The first decision matrix is as follows:

Step 1: Standardize the decision matrix.

$$\mathrm{A}=\left[\begin{array}{ccc}{\mathrm{a}}_{11}& {\mathrm{a}}_{12}& {\mathrm{a}}_{1\mathrm{n}}\\ {\mathrm{a}}_{21}& {\mathrm{a}}_{22}& {\mathrm{a}}_{2\mathrm{n}}\\ {\mathrm{a}}_{\mathrm{m}1}& {\mathrm{a}}_{\mathrm{m}2}& {\mathrm{a}}_{\mathrm{m}\mathrm{n}}\end{array}\right]=\left[{\mathrm{a}}_{\mathrm{i}\mathrm{j}}\right]\mathrm{m}\times \mathrm{n}$$

(2)

where a_ij donate i_th options of j_th criterion.

The standardized value of the j_th attribute for the i_th answer, Ai is given by r_ij.

$${\mathrm{r}}_{\mathrm{i}\mathrm{j}=}\frac{{\mathrm{a}}_{\mathrm{i}\mathrm{j}}}{{\sum }_{\mathrm{i}=1}^{\mathrm{m}}{\mathrm{a}}_{\mathrm{i}\mathrm{j}}^2}\left(\mathrm{i}=\mathrm{1,2},\dots ,\mathrm{m}\right);$$

(3)

Step 2: Calculate the entropy.

$${\mathrm{e}}_{\mathrm{j}=}-\mathrm{K}{\sum }_{i=1}^m{r}_{ij}ln{r}_{ij, }\left(\mathrm{j}=\mathrm{1,2},\dots ,\mathrm{n}\right);$$

(4)

where K = 1/ln m.

Step 3: The weights of each attribute are calculated.

$${\mathrm{r}}_{\mathrm{i}\mathrm{j}=}\frac{{1-\mathrm{e}}_{\mathrm{j}}}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}(1-\mathrm{e}\mathrm{j})}\left(\mathrm{i}=\mathrm{1,2},\dots ,\mathrm{n}\right)$$

(5)

2.4.3. Rating Values for Selected Attributes

The rating values for the specific parameters of wheat were determined according to the methods described in

Table 2. Soil and topography demands for wheat farming [11].

Property	Rating Scale
	100–95	95–85	85–60	60–40	40–25	25–0
Slope (%)	0–2	2–5	5–8	8–16	16–25	>25
EC (dS m−1)	0–4	4–8	8–12	12–16	16–24	>24
Soil texture	SiCL, SiC, SiL, C	SCL, LSC, CL	SL, Cm, C, SiCm	LS		S
Gravel (%)	0–3	3–15	15–35	35–55	-	>55
CCE (%)	3–20	<320–35	35–50	50–60		>60
OC (%)	>0.6	0.6–0.4	<0.4	-	-	-
CEC (Cmol + kg−1)	24>	24–16	16>	-	-	-
Soil depth (cm)	>100	100–60	30–60	20–30		<20

SiCL: silty clay loam; SiC: silty clay; SiL: silty loam; SCL: sandy loam; m: massive (compact); LS: loamy sand; S: sandy.

. These values were formulated by compiling information from scientific references [11,52]. These scores were then used in the MCDA methods to produce a land suitability rating for wheat cultivation.

2.4.4. An Overview of the Proposed Method

Three MCDM methods, namely SAW, TOPSIS, and GRA, were used to determine the land potential for wheat cultivation. Using these methods, each soil profile was ranked by determining the land suitability based on the weights and scores of the diagnostic attributes.

SAW: In this approach, decision-makers directly weigh each attribute to express its relative importance. The assigned importance for each attribute is then multiplied by the standardized value for wheat cultivation for that attribute to obtain an overall score. The resulting products are then aggregated across all attributes for each soil profile [57,58]. The following methods (3 and 4, respectively) are used to determine the adjusted value for both positive and negative attributes.

$$n_{ij}=\frac{g_{ij}}{g_{max}} i=1,\dots .,m j=1,\dots ,n$$

(6)

$$n_{ij=}\frac{g_{min}}{g_{ij}} i=1,\dots ,m j=1,\dots ,n$$

(7)

$$Final score=\sum w_{gij}\times n_{ij} \sum w_{gij}=1$$

(8)

where g_ij is the value of the criterion, g_max is the highest possible score for each positive requirement, g_min is the lowest possible score for each negative requirement, w_g is the weighting, and n_ij is the normalized value.

TOPSIS: TOPSIS is an MCDM method for evaluating and prioritizing choices regarding attributes depending on their distance to the best positive and negative answer. This method was proposed by Hwang and Yoon [59] and soon found its place in the field of MCDM. The term TOPSIS refers to preference methods based on equality for the best answer. The logic of the method is the definition of the best answer. The best answer is the one that increases the advantage of the attribute and decreases the disadvantage of the attribute. The best choice is the one that has the smallest distance to the best answer and, at the same time, the furthest distance to the negative best answer [57,58]. The TOPSIS process comprises the following phases.

Step 1: Create a decision matrix.

$$r_{ij=}\left[\begin{array}{ccc}{x}_{11}& x_{12}& x_{1n}\\ x_{21}& x_{22}& x_{2n}\\ x_{m1}& x_{m2}& x_{mn}\end{array}\right]$$

(9)

Step 2: Create a standardized decision matrix.

The standardized value of the j_th attribute for the i_th answer, Ai is given by r_ij.

$$r_{ij=}\frac{x_{ij}}{{\sum }_{i=1}^m{x}_{ij}^2}\left(i=1,2,\dots ,m\right); \mathrm{j}=1,2,\dots \mathrm{n}$$

(10)

Step 3: The weighted and standardized decision matrix is calculated.

$$v_{ij}{=w}_j{r}_{ij,} i=1,\dots ,m; j=1,\dots ,n$$

(11)

where w_j is the weight of the j_th attribute.

Step 4: Determine the negative and the best positive answer.

$$A^{-}=\left\{v_1^{-},\dots ,v_n^{-}\right\}$$

(12)

$$A^{+}=\left\{v_1^{+},\dots ,v_n^{+}\right\}$$

(13)

where A⁻ denotes the best negative answer and A⁺ is the best positive answer. If the j_th attribute is an advantageous attribute, then

$$v_j^{+}=max\left\{v_{ij},i=1,\dots ,m,\right\}or\left\{v_j^{+}\right\} and v_j^{-}=min\left\{v_{ij},i=1,\dots ,m,\right\}$$

(14)

If, on the other hand, the j_th attribute is a cost attribute, then

$$v_j^{+}=min\left\{v_{ij},i=1,\dots ,m,\right\} or \left\{v_j^{+}\right\} and v_j^{-}=max\left\{v_{ij},i=1,\dots ,m,\right\}$$

(15)

Step 5: The distances between each answer and the best positive and negative answer are calculated.

$$d_i^{+}=\sqrt{\sum _{j=1}^n{(v_{ij}-v_j^{+})}^2} i=1,\dots ,m$$

(16)

$$d_i^{-}=\sqrt{\sum _{j=1}^n{(v_{ij}-v_j^{-})}^2} j=1,\dots ,n$$

(17)

di⁺ and di⁻ donate the distance between the i_th answer and the best positive and negative answer, respectively.

Step 6: The relative estimate of the best answer is calculated.

$$C_i\frac{d_i^{-}}{d_i^{+}+d_i^{-}}; 0\le C_i\le 1; i=1,2,\dots ,m$$

(18)

Step 7: Ranking preference orders are done based on the C_i values; the choices are ranked in descending order.

GRA: This method has shown that it has effectiveness in various challenges in MCDM [60,61]. GRA is founded on the fundamental concepts of the grey systems theory. GRA principles are founded on the grey systems theory and theory that is applied to solve difficult problems [58]. GRA is used to assess the correlation between two series of alternatives, and references use grey relational grades. The procedure for GRA includes the following computational steps.

Steps 1–4 are like Steps 1–4 of the TOPSIS section.

Step 5: The grey relationship coefficients are calculated using the weighted standardized decision matrix. The grey relationship coefficient represents the relationship between the i_th answer and the best positive answer for the j_th attribute:

$${\gamma }_{\mathrm{i}\mathrm{j}}^{+}=\gamma \left({\mathrm{r}}_{ij},{\mathrm{v}}_{\mathrm{j}}^{+}\right)=\frac{{\mathrm{m}}^{+}\mathsf{\xi }{\mathrm{M}}^{+}}{{\Delta }_{\mathrm{i}\mathrm{j}}^{+}\mathsf{\xi }{\mathrm{M}}^{+}},\mathsf{\xi }\in \left(0,1\right)$$

(19)

${\Delta }_{ij}^{+}\mathsf{\xi }{M}^{+}=\left|\begin{array}{cc}{r}_{ij}& v_j^{+}\end{array}\right|,m^{+}={min}_i{min}_j{\Delta }_{ij}^{+},m^{+}={max}_i{max}_j{\Delta }_{ij}^{+},\mathsf{\xi }$ is the coefficient of discrimination.

Using the matrix of grey relationship coefficients of the best positive answer, the grey relationship coefficients between different choices are created.

$${\gamma }^{+}={\left[{\gamma }_{\mathit{i}\mathit{j}}^{+}\right]}_{\mathit{m}\times \mathit{n}}$$

(20)

By comparing with the negative best answer, the grey relationship coefficient for the i_th answer of the j_th attribute can be calculated as follows:

$${\gamma }_{ij}^{-}=\gamma \left(r_{ij},v_j^{-}\right)=\frac{m^{-}\mathsf{\xi }{M}^{-}}{{\Delta }_{ij}^{-}\mathsf{\xi }{M}^{-}},\mathsf{\xi }\in \left(0,1\right)$$

(21)

where ${\Delta }_{ij}^{-}\mathsf{\xi }{M}^{-}=\left|\begin{array}{cc}{r}_{ij}& v_j^{-}\end{array}\right|,m^{\pm }={min}_i{min}_j{\Delta }_{ij}^{-}, \mathrm{a}\mathrm{n}\mathrm{d} m^{\pm }={max}_i{max}_j{\Delta }_{ij}^{-}$.

If you consider the negative best answer, the grey coefficient matrix is shown here:

$${\gamma }^{-}={\left[{\gamma }_{\mathit{i}\mathit{j}}^{-}\right]}_{\mathit{m}\times \mathit{n}}$$

(22)

Step 6: the grey relational grade is determined.

To create an overall evaluation of the solutions based on all attributes, the grey relational score is used. It is defined as the mean of the relational coefficients for the distinctive attributes. For the i_th difference, the grey relational score is expressed in terms of the best positive difference as follows:

$$G_i^{+}=\frac{1}{n}\sum _{j=1}^n{\gamma }_{ij}^{+}$$

(23)

Similarly, the degree of the grey relation for the i_th difference in relation to the negative difference is as follows:

$$G_i^{-}=\frac{1}{n}\sum _{j=1}^n{\gamma }_{ij}^{-}$$

(24)

Step 7: The relative grey value of the relationship is calculated.

$$C_i\frac{D_i^{-}}{D_i^{+}+D_i^{-}}$$

(25)

Step 8: the alternatives are arranged in descending order based on their Ci.

2.4.5. Comparison and Validation of Methods for Determining Land Suitability Values

In this step, the results of the previous step (i.e., the determination of land suitability values for 70 soil profiles by three MCDM methods, including TOPSIS, SAW, and GRA) were evaluated. In evaluating the effectiveness of the MCDM methods, the analysis focused primarily on identifying the notable correlations between the values calculated by each method and the observed wheat yield. We used the statistical metric of coefficient of determination (R²) to evaluate how well the methods predicted the value of land suitability.

2.4.6. Mapping and Modeling of Land Suitability Values

To assess the spatial variations of land suitability values in our study area, we used a random forest (RF) ML, digital soil mapping (DSM) and a set of environmental covariates. Within the DSM, environmental covariates are used to predict soil attributes and the relationships between soil attributes and environmental covariates are established by ML models. RF is a great machine learning model whose predictive ability has been demonstrated in various fields [62,63,64]. A literature review on RF can be found in Antoniadis [65] and Belgiu and Dragut [66].

The variables were a combination of categorical data (e.g., a geomorphological map), terrain features generated from a model of digital elevation with a resolution of 10 m, and remote sensing data obtained from Landsat imagery with a resolution of 30 m. Together, these data sources contributed to the predictive modeling of suitability values for each geographic unit.

SAGA software 9.5.1 [67] was used to calculate terrain features such as valley floor flatness, slope, LS factor, topographic wetness, aspect, plane curvature, and elevation. The spectral bands and some indices, such as the normalized difference vegetation index, were created from a Landsat 8 Operational Land Imager (OLI) image (23 August 2019). All environmental covariates were resampled to achieve a uniform grid size (30 m) and were then used to predict the determined land suitability values by three MCDM methods.

All data were divided into categories, including train data (70%) and test data (30%). To validate the results of the random forest ML models, a rigorous evaluation strategy was applied. A k-fold cross-validation procedure with 100-fold replication was applied. The statistical indices, including coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE), were applied.

2.4.7. Classification of the Land Suitability Value

To allow a better visual comparison of the map produced by the MCDM methods, the land suitability values calculated by each method were categorized into three classes: high, medium and low. This categorization process included the following steps:

-

Determine the range: The range of land suitability values for each method has been determined. This range is the difference between the maximum and minimum suitability values obtained with each method.

-

Divide into intervals: The total range of values was divided into three equal intervals. These intervals correspond to the categories of land suitability: high, medium, and low.

-

Define interval boundaries: The first interval (low suitability) starts at the minimum score and extends to the value that marks the end of the first third of the range. The second interval (medium suitability) starts at the end of the first interval and extends to the end of the second third of the range. The third interval (high suitability) starts at the end of the second interval and extends to the maximum score.

-

Assign categories: each land suitability value was then assigned to one of the three categories based on the interval in which it fell [68,69].

2.4.8. Spatial Autocorrelation Analysis

We used spatial autocorrelation analysis to further analyze the spatial variability of key attributes and to create a map of land suitability values obtained by the three proposed MCDM approaches (i.e., TOPSIS, SAW, and GRA). A statistical technique used to measure and evaluate the degree of dependency between spatial data is the spatial autocorrelation analysis. In other words, it facilitates the delineation of whether the values are randomly distributed or whether they are close to each other in space [70]. R 4.3.2 software was used to calculate the focal correlation between two raster maps of land suitability assessment produced by the three proposed MCDM approaches and wheat yield.

3. Results

3.1. The Statistical Description of Soil Properties

Table 3. Descriptive statistics for soil properties.

	Sand	Clay	Silt	OC	Gravel	CCE	EC	CEC	pH
	%						dS m−1	cmol + kg−1
Number	70	70	70	70	70	70	70	70	70
Mean	42.52	28.78	28.67	0.81	27.54	14.3	0.59	11.49	7.52
Minimum	19.92	12.00	9.20	0.15	7.47	2.37	0.20	2.17	7.00
Maximum	66.19	45.00	43.00	1.60	47.42	38.00	1.50	23.63	7.90
Skewness	0.17	0.20	−0.29	0.20	0.49	0.77	1.30	0.38	−0.53
Kurtosis	−0.29	−0.14	1.11	−0.27	−1.09	0.82	2.35	−0.27	1.24
CV	44.17	21.02	24.72	38.51	40.48	49.95	41.05	41.02	2.15

contains a statistical description of the soil properties. The amounts of CCE range from 2.37% to 38%, with a mean of 14.3%; the pH values range from 7 to 7.9, with a mean of 7.52; the amounts of OC range from 0.15 to 1.6%, with a mean of 0.81%; gravel values range from 7.47 to 47.42%, with a mean of 27.54%; EC values range from 0.2 to 1.5 dS m⁻¹, with a mean of 0.59 dS m⁻¹; and CEC values range from 2.17 to 23.63 Cmol + kg⁻¹, with a mean of 11.49 Cmol + kg⁻¹. The main soil textural classes are SC (sandy clay), C (clay), SCl (sandy clay loam), CL (clay loam), SL (sandy clay), and L (loam). EC, CCE, sand, and clay have a low positive skewness. pH and silt, on the other hand, have a negative skewness. The coefficient of variation (CV) for EC, CEC, OC, gravel, sand, and CCE is greater than 30% (Table 3). In the study area, silt and clay have a moderate coefficient of variation (10–30%), while pH has a low coefficient of variation (10%).

3.2. Calculated Weighting of the Attributes

Because each attribute has a specific and significant effect on plant growth, decision-makers must assign different importance to these requirements when evaluating land suitability.

Table 4. Results of attribute weighting using Shannon’s entropy method.

	Slope	OC	CEC	Texture	EC	CCE	Gravel	Soil Thickness
Weight	0.439	0.135	0.108	0.010	0.086	0.090	0.077	0.054

shows the results of using Shannon’s entropy to determine the final weights for all attributes. The results show that slope, OC, and CEC are the most significant variables in evaluating the suitability of land for wheat cultivation, with loadings of 0.439, 0.135, and 0.108, respectively (Table 4). The loadings for CCE, EC, gravel, soil thickness, and soil texture are 0.090, 0.086, 0.077, 0.054, and 0.010, respectively. The spatial variability of key attributes and wheat yield in the studied area is shown in Figure 4.

3.3. Determination and Validation of the Land Suitability Value

Table 5. Values for land suitability, land suitability class, and key attributes for wheat cultivation.

	TOPSIS	SAW	GRA	Slope %	OC %	CEC (cmol + kg−1)	Yield (ton ha−1)
Mean	0.31	0.15	0.27	15	0.81	11.49	1.14
Maximum	0.51	0.21	0.44	79	1.60	23.63	1.75
Minimum	0.06	0.06	014	0	0.15	2.17	0.5
Good class range	0.34–0.44	0.38–0.52	0.34–0.44	<5	1<	15	1.4<
Medium class range	0.24–0.34	0.24–0.38	0.24–0.34	5–10	0.5–1	10–15	1–1.4
Weak class range	0.14–0.24	0.10–0.24	0.14–0.24	5<	<0.5	<10	0.5–1

shows the results of the MCDA methods used to determine the land suitability value of the study region for wheat cultivation. This table also shows the results of the categorized land suitability value. With a mean of 0.31, TOPSIS results range from 0.06 to 0.51, GRA range from 0.14 to 0.44, with a mean of 0.27, and SAW range from 0.06 to 0.21, with a mean of 0.15, and were used to determine the land suitability values.

As can be seen in Figure 5, the performance of each model is evaluated by determining the significant relationship with wheat yield. There are differences between the MCDA approaches, as our comparison shows. The TOPSIS approach gives the best results with a value of 0.74, as shown in Figure 6, followed by GRA (0.72) and SAW.

3.4. Land Suitability Value and Key Attribute Mapping

We used the RF algorithm and a DSM method to assign the suitability values determined by the MCDA approaches.

Table 6. Results of RF models for mapping land suitability values for wheat using three MCDM methods.

	GRA	TOPSIS	SAW
RMSE	0.072	0.081	0.061
MAE	0.005	0.017	0.003
R2	0.83	0.80	0.88

shows the result of the RF algorithm for assigning the suitability of wheat cropland using three MCDM methods. The SAW approach (0.003, 0.061, and 0.88) gives the best results as measured by the MAE, RMSE, and R² statistical indices. This is followed by the GRA (0.081, 0.017 and 0.83) and TOPSIS (0.005, 0.072, and 0.80) methods. Figure 6 shows the results of this study, including maps of machine learning-based land suitability values for wheat cultivation in the study area using the RF model.

3.5. Spatial Autocorrelation Analysis

Figure 7a shows the results of a spatial autocorrelation study comparing wheat yield with three important criteria: slope, CEC, and OC. Spatial autocorrelation can be broadly categorized into three types: positive, negative, and absent. The data from the study area show that slope, CEC, and OC correlate strongly with wheat yield. Within the study area, slope inclination and yield show a negative and significant spatial autocorrelation. On the other hand, the spatial autocorrelation between yield and OC and CEC is positive and high in the study area (Table 2, Figure 3 and Figure 7a). In addition, Figure 7b shows the map of spatial variability of land suitability assessment as determined by the three proposed MCDM approaches (i.e., TOPSIS, SAW, and GRA). In the study area, the spatial autocorrelation of the MCDA methods with the yields is also positive, with the TOPSIS method showing the strongest correlation.

4. Discussion

4.1. Key Attributes for Wheat Cultivation

The results show that soil properties, especially CEC and OC, have undeniable effects on soil productivity and the promotion of agricultural sustainability in the study area (Table 4 and Figure 4). The coefficients of variation (CV) of slope, OC, and CEC (106.75, 49.95, and 49.02, respectively) (Table 2) support this assertion. This study emphasizes the critical importance of topographic factors in developing a reliable land evaluation model, with slope having a particularly strong influence on yield outcomes. As an important regulator of water distribution, slope can directly influence drainage conditions, erosion and water retention in the landscape. On gentle slopes, it is more important for agriculture as it controls the infiltration rate and prevents waterlogging. On steep slopes, however, it can lead to soil erosion and a reduction in soil productivity [71,72].

As an important soil property, CEC has a strong influence on soil productivity and, thus, on crop yields. OC was identified as another important soil property in the study area that significantly influences soil productivity via pH and soil fertilization. It is known that the content of OC in the soil has a positive influence on soil productivity [73]. This variation in OC, CEC, and slope was largely due to variability in topography and parent material. This shows that topography and parent material are the most important factors for soil formation in the study area, as Taghizadeh-Mehrjardi [12] also found.

Dedeoğlu and Dengiz [74] calculated the land suitability index for wheat by integrating AHP and GIS in Turkey, and their results also showed that the most effective attributes were soil strength, soil texture, and slope indicators.

These results are also confirmed by evaluating the spatial autocorrelation of MCDA methods in terms of crop yields (Figure 7a). It is evident that slope, CEC, and OC have a high correlation with crop yields. The spatial autocorrelation of slope and yield within the study area is remarkably negative and high and is observed in most of the region. There is a positive spatial autocorrelation in low slope areas, which include Pi111, Pi121, Pi211, Pi212, Pl111, and pl211. In contrast, spatial autocorrelation is negative and high in high slope areas, which mainly include Hi111, Hi121, Hi131, Hi211, Mo111, and Mo121. Regarding OC, we observed different trends. In areas with high OC and CEC content, there is a positive correlation between yield and OC (Figure 7a).

4.2. Evaluation of the MCDM Techniques

Depending on the validation results (Figure 5), the best results were observed for the TOPSIS method, followed by the GRA and SAW methods. The results show that the TOPSIS and GRA methods have a stronger significant correlation with yield than the SAW method, indicating that the TOPSIS and GRA methods efficiently determined the land suitability assessment in most of the studied areas. The methods (TOPSIS and GRA) have also been used in the context of land suitability by other researchers, such as Ekmekcioglu [75] and Zhang and Duan [76]. These studies have shown that these techniques are a useful tool for decision makers in evaluating land suitability.

The methods help decision-makers make long-term land use and management decisions by considering various parameters [77]. The differences between these methods (TOPSIS and GRA) and SAW lie in the assignment of weights, the use of different mathematical functions, and standardization. Although the SAW method is easy to use and understand, it may not capture the differences in land suitability assessment as well as the other two methods. Seyedmohammadi [57] also evaluated the application of three models, including the fuzzy TOPSIS, TOPSIS, and SAW models, to determine the land suitability values for the cultivation of three crops (soybean, canola, and corn) in the northern part of Iran. Their results proved that the fuzzy TOPSIS method was the best approach for determining land suitability. These results are further supported by examining the spatial autocorrelation of MCDA methods with respect to crop yields (Figure 7b). The spatial autocorrelation of MCDA methods with yields within the study area is positive and high and can be found throughout the region. However, the spatial autocorrelation is lower in areas with a lower number of samples, especially in the northern and western parts.

4.3. Mapping and Modeling the Land Suitability Values

The result of the RF algorithm for mapping the land suitability of wheat is acceptable (Table 6). Random forest provides the best estimate by integrating the results of many decision trees. This method can model complicated spatial patterns and is well suited for mapping land suitability levels. The RF algorithm has also been successfully used in other land suitability studies [12,77].

According to the maps of land suitability for wheat cultivation (Figure 6), suitable land units (good class) are mainly concentrated in the center of the study area. These areas are characterized by a slope of less than 5%, a yield of more than 1.4 t ha⁻¹, a CEC of more than 15 cmol + kg⁻¹ and an OC of more than 1% (Figure 4 and Table 5). These areas account for 33%, 20%, and 6% of the suitability values assessed using the TOPSIS, GRA, and SAW methods, respectively. They correspond to the geomorphological land units Pi111, Pi211, and Pl111 (Figure 1). Moderately suitable land (medium class) is characterized by a slope of 5% to 10%, a yield in the range of 1–1.4 t ha⁻¹, a CEC in the range of 10–15 cmol + kg⁻¹, and an OC of more than 0.5 to 1% (Figure 4 and Table 5). These land categories are concentrated in the central region and partly extend to the southern and eastern parts of this study (Figure 6). They account for 16%, 22%, and 29% of the suitability values determined with the TOPSIS, GRA, and SAW methods, respectively. They are assigned to the geomorphological units Pi121, Pi212, and Pl211 (Figure 1). Overall, the TOPSIS and GRA methods classify more areas as good compared with the SAW method, which classifies more areas as moderately or poorly suitable.

4.4. Practical Applications

Overall, this study offers practical advice with a variety of implications for land use planning and agricultural management. The results of the study have a major impact on the establishment of land use regulations in the area studied. A solid basis for optimal decisions is the understanding and identification of the different soil classes suitable for wheat cultivation. Finding suitable land with certain topographic and soil characteristics, often found in the center of the study area, can provide a great opportunity for wheat cultivation, and this tactic will likely increase wheat yields.

However, information on areas that are neither well nor moderately adapted helps in decision-making, e.g., when it comes to nature conservation or other land uses. Precision agriculture approaches tailored to local conditions are supported by research by identifying critical variables such as soil properties (CEC and OC) and topographic parameters (slope). Precision agriculture builds on this expertise to help farmers manage their land more prudently and reduce the risk of soil degradation. Fertilization and irrigation techniques are adapted based on soil fertility and the ability to store water. This study’s conclusions benefit sustainable land use planning by identifying techniques to preserve the environment in environmentally fragile areas and optimize crop selection to increase productivity. Policymakers, investors and farmers can make informed decisions about agricultural investments, policies and climate change adaptation strategies with the help of decision support tools developed based on land suitability maps. In the future, the use of cutting-edge technologies such as continuous monitoring and remote sensing will further improve the accuracy and relevance of land suitability assessments.

5. Conclusions

In this study, site suitability was determined by advanced MCDM techniques and machine learning for wheat cultivation in the Gavshan region of western Iran. The effects of various factors (such as soil and environment) on wheat yield were evaluated, and the importance of sustainable land management for optimizing agricultural productivity was considered. In this study, three MCDM techniques, including TOPSIS, SAW, and GRA, were presented for land suitability evaluation and their application for wheat cultivation was discussed. The weighting of diagnostic attributes was determined using Shannon’s entropy. It was found that factors such as slope, CEC, and OC were the most important attributes and had the greatest influence on the evaluation of wheat land suitability. The results proved that the TOPSIS and GRA methods were more efficient than the SAW method in determining land suitability and had a higher correlation with the observed wheat yield data. The spatial distribution of suitability values was mapped with RF and categorized into three levels, including good, moderate, and poor areas for wheat cultivation. In general, this study shows the importance of combining advanced MCDM techniques and ML to assess the suitability of wheat areas. It underlines the need for a comprehensive analysis of soil and environmental attributes to optimize agricultural productivity and sustainability. The results provide important insights for land use policy and agricultural management decisions and contribute to more informed and sustainable land use practices. In general, this study has shown the importance of combining advanced MCDM techniques and ML to assess the suitability of wheat land. It underlines the need for a comprehensive analysis of soil and environmental characteristics to optimize agricultural productivity and sustainability. The results provide important insights for decision-making in land use policy and agricultural management and contribute to more informed and sustainable land use practices. In other words, even though this study provides insightful results for wheat cultivation, there are some limitations that need to be considered. First, the models were calibrated using local data, so their direct transferability to other locations with different conditions is limited. The results are unique to the Gavshan region in western Iran. Furthermore, this study is not applicable to temporal variations in environmental conditions that may have a significant impact on wheat cultivation, such as seasonal variations or long-term climatic changes. Finally, additional validation with independent datasets from other locations or time periods could improve the generalizability of the results, even if the MCDM and machine learning models were validated with local wheat production data.