Defining a Method for Mapping Aeolian Sand Transport Susceptibility Using Bivariate Statistical and Machine Learning Methods—A Case Study of the Seqale Watershed, Eastern Iran

Abstract

Desert regions face unique challenges under climate change, including the emerging phenomenon of sand dune expansion. This research investigates aeolian sand transport in the Seqale watershed (eastern Iran) using geostatistical and machine learning methods to model and forecast dune spread, aiming to reduce the loss of sustainability in these valuable landscapes. Predictor variables (altitude, slope, climate, land use, etc.) and wind erosion occurrence were analyzed using classification algorithms (decision tree, random forest, etc.) and bivariate methods (information value, area density) in R software 4.5.0. Risk zoning maps were created and evaluated by combining these approaches. Results indicate a higher sand dune presence in regions with specific altitude (1200–1400 m), gentle northeast-facing slopes (2–5 degrees), moderate rainfall (250–500 mm), high evaporation (2500–3000 mm), outside flood plains, and far from roads (>3000 m) and water channels (>500 m). Dune expansion maps based on density area and information value methods showed substantial areas classified as high to very high movement risk. Machine learning analysis identified the Support Vector Machine (SVM) algorithm (AUC = 0.94) as the most effective for classifying sand dune zones. The study concludes that spatial forecasts, combined with tailored physical and biological measures, are essential for effective sand dune management in the region.

1. Introduction

Wind-blown sand deposits, which cover an estimated 6% of the Earth’s surface area, are predominantly found in the sand seas of arid regions, accounting for about 97% of these deposits [1]. Dry regions comprise a significant portion of the world, with about 20% of these areas experiencing wind-blown sand [2]. However, this ratio may vary significantly, ranging from 2% (North America) to over 30% (Australia) and more than 45%, e.g., in Central Asia [3]. Several regions in the Northern Hemisphere are part of the arid zone extending from North Africa to the Arabian Peninsula. This zone also encompasses specific territories in Iran and Pakistan, with representative examples of deserts situated in the mid-latitude deserts in Central Asia [4]. For instance, Southwest Africa [5] and Central Australia [6] register significant sand cover, but its spatial distribution is uneven and difficult to predict as it is transported across these regions. Various factors significantly affect sandy areas, such as parent material, weathering degree and rate, hillslope inclination, and wind direction or intensity [7].

From a geomorphological perspective, aeolian sediments play a crucial role in shaping desert landscapes, particularly in arid and semi-arid climates [8,9]. These sediments are vital for forming and maintaining dunes, ripples, and other unique features of desert topography and ecosystem services [10]. The study of aeolian sediments is of great interest to researchers and practitioners in various fields, including geology [11,12], geomorphology [13,14], ecology [15,16,17], and archaeology [18]. Ergs, also known as desert sand seas, are extensive land areas in arid regions that form through the accumulation of sediment and are of particular interest in the scientific literature [2]. These regions exhibit different levels of vegetation cover and typically encompass areas of at least 125 km² [19]. The term erg is derived from the African-Arabic and Arabic languages and has been used to describe aeolian deposits of any size located northwest of the Sahara. Asia has the largest area of ergs, accounting for 45.5% of the total surface. Africa follows with 34.2%, while the Americas contribute 20% and 0.3%, respectively [1]. Over the past few decades, the spread and growth of aeolian sedimentation in the Middle East [20], particularly in Iran [21], have posed a noteworthy risk to nearby populations, resulting in hazardous conditions and hindering development. The consequences of this phenomenon have caused irreparable damage to agriculture [22,23], transportation infrastructure [24,25], and the environment [26].

Iran, spanning approximately 164 million hectares, is situated within the world’s arid belt [27]. Its unique topographical features include the Alborz and Zagros Mountain ranges, enclosed internal lowlands, and the formation of playas [28].

Accurate and reliable information regarding the extent of erosion in Iran is scarce, and a notable disparity exists between current measurements and estimations [29,30]. Annually, a significant amount of the country’s topsoil is eroded, resulting in its transportation to both the Persian Gulf and the Caspian Sea [31]. Iran is experiencing a worrying trend in soil erosion [32], with rates far exceeding those observed in other Asian countries. Soil erosion in Iran registers five times higher rates than the average, affecting 75% of the country [33]. It is worth noting that the scientific literature has identified Iran as one of the most vulnerable countries globally in terms of soil erosion intensity and volume [34,35,36,37]. According to a survey conducted at the beginning of the century, a staggering 24 million hectares of land in Iran are affected by aeolian sediments [38]. This study identified three distinct areas of impact, including 12.8 million hectares in detachment areas where sand is detached from the surface, 6.8 million hectares in transport areas where the wind carries the detached particles, and 4.4 million hectares in sedimentation areas where the particles settle and accumulate, known as Ergs. A comprehensive study conducted in Iran [39] to evaluate the mobility of dune fields using the Lancaster Index [40] indicated that most of Iran’s dunes are either active or semi-active. Furthermore, this study reveals that Iran’s eastern and southeastern regions exhibit higher mobility rates than the rest.

Sand transport flux varies significantly across different underlying surfaces [41], and wind is the most important external force in shaping aeolian landforms [42]. Past studies have emphasized the complex challenge of measuring and managing aeolian sediment transport [43] and spatiotemporal variation in the sediment transport rate [44]. Meteorological factors, surface sediment grain size and its physical parameters, topographic factors and surface vegetation cover significantly influence blowing sand activities and sediment transport to the landform [44,45,46,47]. However, the study of each factor as an indicator of aeolian sediment transport, even at large sample sizes, showed uncertainty [48] in determining where the sand will be deposited. Therefore, land degradation factors, transport pathways and wind characteristics from source to sink, deposition factors of windblown sediments, and their interactions should be carefully examined to develop a better model.

Data mining techniques have been widely used to predict natural and geological hazards. As a method for recognizing functional patterns in data with minimal user intervention, data mining is vital. Efficient natural disaster management necessitates their identification and categorization. Machine learning algorithms can potentially enhance the precision and efficiency of natural disaster detection and classification [49]. They have become a viable strategy for improving natural disaster management in several ways in recent years [50]. Thus, knowing which classification methods have the best potential to obtain greater precision will improve natural disaster management in recent years [51] and enable stakeholders to mitigate and respond to natural disasters more effectively [50]. According to the literature review, data mining in large databases has been recognized by many researchers as a critical research topic and has been widely used in various fields. Hostile environmental conditions and various natural and anthropogenic factors characterize arid and semi-arid regions. Therefore, mitigation of land degradation depends on understanding the natural causes of degradation by developing appropriate techniques [52].

Studies are discussed in detail, providing insights into their capabilities in modelling the interactions between wind, sand, and vegetation. Integrating different perspectives into theoretical frameworks yields a more comprehensive understanding of aeolian sand transport. While the complexity of modelling this process is highlighted, there is a lack of research about the factors and their associated aeolian sand transport. Previous studies indicate that aeolian sand transport varies according to wind characteristics, climate, and regional geomorphology and topography. To address this gap, this study proposes a new framework based on numerical methods that identify aeolian sand transport risk from easily accessible features by incorporating machine learning algorithms and bivariate statistical methods. Accordingly, this paper focuses on additional approaches that can complement existing ideas on the technique’s accuracy, as well as providing a more detailed description of local features examined in relevant previous field surveys in arid lands.

2. Materials and Methods

2.1. Study Area

Seqala is a city in the northwest region of South Khorasan Province, located on the periphery of the Dasht-e Kavir of Iran and Sarayan city (Figure 1). Additionally, it falls within the watershed of Kavir Namak and experiences a hot and arid climate [53]. Due to its low altitude and distance from significant water sources and rain-bearing masses, the territory experiences a dry environment characterized by intense rainfall concentrated in a few events, high summer heat, and abundant evaporation [54]. Ferdows’ synoptic station (Figure 1) indicates that annual precipitation levels typically fall below 100 mm, leading to dry conditions that resemble droughts. August is the hottest month of the year, with temperatures reaching up to 48.4 °C, while January is considered the coldest month, with temperatures dropping to as low as 1.4 °C. The annual temperature absolute range varies from −16 to 44 °C, with daily temperature fluctuations. The annual precipitation in the region is marked by two peaks, one in March and another in September, with the highest and lowest rainfall, respectively. Over the 30-year period (1989–2018), mean monthly Class A pan evaporation and relative humidity values were derived (Figure 2). Annual evaporation averaged 2543 mm, with a mean relative humidity of 35.8%.

The Seqala region is located on Quaternary sediments, consisting mainly of limestone, chalk, and salt, separated from the other areas by geological faults, including the Kavir Bozorg, Kolmard, and Hariroud. The characteristic landscape shows the significant influence of this soil movement by erosion, with signs of water scarcity, vegetation adapted to the dry periods, and a frequent presence of dunes.

2.2. Dataset and Data Treatment

The dune expansion area was divided into two subsets, training and validation data, through random partitioning of the sandy region’s dataset. The training dataset was used to determine the weight and create an expansion map based on surface density and information value. The validation dataset was employed to validate the accuracy of the sensitivity map using a field study. This research involved modelling techniques using various types of spatial indicators, including topography (altitude, aspect, and slope), climate (temperature, precipitation, and evaporation), geomorphology, pedology, distance from roads and rivers, floodplains, and human-related indicators such as land use types (

Table 1. Indicators used in the research.

Indicators	Resolution	References
Altitude	30 m	Aster DEM, USGS (The United States Geological Survey)
Aspect	30 m	Aster DEM, USGS (The United States Geological Survey)
Slope	30 m	Aster DEM, USGS (The United States Geological Survey)
Temperature	100 m	Seqale watershed map (1:25,000), NRWMO (Natural Resources and Watershed Management Organization)
Precipitation	100 m	Seqale watershed map (1:25,000), NRWMO (Natural Resources and Watershed Management Organization)
Evaporation	100 m	Seqale watershed map (1:25,000), NRWMO (Natural Resources and Watershed Management Organization)
Geomorphology	100 m	Geology map (1:100,000), GSMEI (Geological Survey and Mineral Exploration of Iran)
Pedology	100 m	Pedology map (1:50,000), SWRI (Soil and Water Research Institute)
Distance from roads	100 m	Topography map (1:25,000), NCC (Iran National Cartographic Center)
Distance from rivers	100 m	Topography map (1:25,000), NCC (Iran National Cartographic Center)
Floodplains	100 m	Seqale watershed map (1:25,000), NRWMO (Natural Resources and Watershed Management Organization)
Land use types	100 m	Landsat satellite, USGS (The United States Geological Survey)

).

2.3. Modelling Techniques

The analysis assessed dune distribution based on various factors, quantifying the impact of each factor and assigning weights to different groups according to their statistical relationships. This was achieved using bivariate statistical methods: information value (IV) and density area (DA). The IV bivariate statistical method involves classifying each layer based on information layers and calculating the area and density of sand in each class. The study analyzed dunes across diverse factors to determine their extent. Bivariate statistical methods identified influential factors that quantified class weighting. The weight of each category was subsequently determined using the Van-Westan formula (Equation (1)) as described by [55].

$$W_i=\mathrm{Ln}\left[{\displaystyle \frac{Densclass}{Densmap}}\right]=\mathrm{Ln}\left[{\displaystyle \frac{\frac{NpixSi}{NpixNi}}{\frac{\sum NpixSi}{\sum NpixNi}}}\right]$$

(1)

where W_i is the weight value of each class, Densclass is the density of the dune areas in each category, Densmap is the total density of the regions in the zone, NpixSi is the number of pixels in each class, NpixNi is the number of areas in each category.

In the DA bivariate statistical method, the density of dunes is calculated by Equation (2) for each operating map or parameter [56].

$$W_{\mathrm{erea}}=1000\left[{\displaystyle \frac{NpixSi}{NpixNi}}-{\displaystyle \frac{\sum NpixSi}{\sum NpixNi}}\right]$$

(2)

where W_erea is the density of dunes.

With both methods, the layer of dune expansion was classified into five zones: very low risk, low risk, medium risk, high risk, and very high risk. Various classification algorithms, namely support vector machines (SVM), random forest (RF), decision tree (DT), boosting aggregate (BA), and neural network (NN), were employed using the Rattle package in the R software v. 4.5.0, similarly in some research [57,58]. Additionally, the algorithms were assessed using the receiver operating characteristic (ROC) curves based on the area under the curve.

A support vector machine is a robust algorithm for both linear and nonlinear classification. It can classify data that is not linearly separable by transforming it into a higher-dimensional space and finding an optimal separating hyperplane [59]. Boosting aggregate is an ensemble learning algorithm that enhances model accuracy by combining multiple weak learners into a strong one, focusing on previous iteration errors. In boosting, each training sample contributes to every decision tree subset, created through over-weighted sampling with replacement [60]. Decision trees are supervised learning algorithms for classification, prediction, interpretation, and data manipulation. Their flowchart-like structure simplifies understanding and interpretation, making them suitable for explaining decision-making processes and evaluating decision validity [61]. Random Forest, an ensemble method combining multiple decision trees, is widely used for classification due to its simplicity and diversity [62]. Artificial neural networks, inspired by human neurons, are machine learning algorithms designed to recognize patterns in large, diverse datasets [63]. Neural networks have shown slightly better performance than classification tree algorithms [64].

2.4. Validation

The system’s prediction accuracy was evaluated based on accurately estimating the occurrence and non-occurrence of events, where the dependent variable was the presence or absence of dunes at each point, and the independent variables were the informative layers. To evaluate the modelling accuracy, 70% of the points are used for model training, and 30% are considered for model testing. Additionally, it is common in many modelling studies using machine learning [30,57,65,66].

2.5. Wind Analysis and Sand Drift Potential

Hourly wind speed and direction data from the past two decades were collected from the Ferdows synoptic station, located at 34°10′ N, 57°50′ E, obtained from the Iranian meteorological organization [67]. This station is the closest to the study area’s dunes. A wind rose (WR) was generated using WRPLOT (Version 8.2) software. The Fryberger method [19] was employed to analyze sand drift potential using MATLAB 2018 software, as outlined in Equation (3):

$$\mathrm{D}\mathrm{P}={\mathrm{V}}^2(\mathrm{V}-{\mathrm{V}}_{\mathrm{t}})\times \mathrm{T}$$

(3)

where DP is the amount of sand drift based on the vector units (VU), V represents the average wind velocity (m s⁻¹) in eight ordinal and cardinal directions, V_t signifies the threshold wind velocity (m s⁻¹), and T denotes the time the wind blew, expressed as a percentage in a wind summary. Other parameters of interest include the resultant drift potential (RDP), the resultant drift direction (RDD), and the unit directional index of wind estimated by vector analysis [19], as expressed in Equations (4)–(8):

$$\mathrm{R}\mathrm{D}\mathrm{P}=\sqrt{({\mathrm{C}}^2+{\mathrm{D}}^2)}$$

(4)

$$\mathrm{R}\mathrm{D}\mathrm{D}=\mathrm{A}\mathrm{r}\mathrm{c}\tan ({\displaystyle \frac{\mathrm{C}}{\mathrm{D}}})$$

(5)

$$\mathrm{U}\mathrm{D}\mathrm{I}=\mathrm{RDP}/\mathrm{DP}$$

(6)

$$\mathrm{C}=\sum _{\mathrm{i}=1}^8\left(\mathrm{D}{\mathrm{P}}_{\mathrm{i}}\right)\sin {\mathsf{\theta }}_{\mathrm{i}}$$

(7)

$$\mathrm{D}=\sum _{\mathrm{i}=1}^8\left(\mathrm{D}{\mathrm{P}}_{\mathrm{i}}\right)\cos {\mathsf{\theta }}_{\mathrm{i}}$$

(8)

where θ_i represents the midpoint of wind speed class i direction, V_t is assumed as 6 m s⁻¹ [19,68,69,70] as well as C and D, are DP in directions.

3. Results and Discussion

3.1. Spatial Variability of the Dataset

Figure 3 presents a spatial pattern based on geomorphology, climate, distances, and human activities. Terrain height and slope decrease from the north and northeast towards the south and southwest (Figure 3a), with the southern region characterized by flat, low-lying areas. The temperature fluctuates between 12 and 18 °C, while central and southern areas experience higher rainfall and lower evaporation rates (Figure 3b). Land use analysis revealed weak and imbalanced rangeland cover as the dominant land type.

This can be attributed to the site’s natural geological and climatic conditions (Figure 3b,c), historical land use patterns, and current land management practices (Figure 3d) [71,72]. Despite their dominance in terms of land area, the productivity and ecological value of these rangelands may vary depending on several factors, such as soil quality, vegetation composition, and grazing intensity [73,74]. Iran suffers from inadequate precipitation levels, exacerbated by intense sunlight and strong winds that increase evaporation [75]. This process amplifies pre-existing moisture deficits, resulting in vegetation scarcity. Such conditions promote the aeolian sand movement, posing various ecological and environmental challenges [25]. Also, infrastructures have been damaged by the continuous accumulation of sand, causing expensive losses, such as the burial of railroad tracks [76]. The absence of soil moisture content and vegetation in these regions poses significant challenges for life forms to survive [77], affecting some primordial activities, such as agriculture or grazing. Additionally, the environmental balance is often disrupted, leading to severe consequences for the ecosystem [78].

3.2. Application of Information Value (IV) and Density Area (DA) Bivariate Methods

Analyzing various classes of each information layer using IV and DA bivariate methods revealed that some characteristics are more likely to correlate with a higher dune presence (

Table 2. Results of the value analysis of the most effective classes of each layer using both IV and DA methods.

Indicator	Effective Class	Density of Dune Areas (%)	Weight in Information Value Method	Weight in Density Area Method
Altitude	1200 to 1400 m	3.489	0.328	9.769
Slope	2 to 5°	3.442	0.3150	9.3020
Aspect	northeast	3.1561	0.2281	6.4373
Precipitation	250 to 500 mm	2.6328	0.0456	1.1750
Temperature	12 to 18 °C	2.5166	0.0001	0.0047
Evaporation	2500 to 3000 mm	3.0888	0.2050	5.7251
Flood plains	Outside	2.5346	0.0068	0.1711
Distance from the road	more than 3000 m	3.1584	0.2273	6.4224
Distance from the river	500 to 700 m	2.7217	0.0785	2.0554
Geomorphology	Quaternary	1.6955	−41.2914	144.3974
Land use	Active dunes	26.35	2.3491	238.3581
Pedology	XDL (Sabulous)	16.80	−41.2271	165.3120

). Based on class value calculations for each feature layer, the most probable areas for dune generation exhibit altitudes between 1200 and 1400 m, northeast-facing slopes of 2 to 5 degrees, total rainfall between 250 and 500 mm, evaporation rates between 2500 and 3000 mm, locations outside floodplains, distances exceeding 3000 m from roads and 500 to 700 m from rivers, an ultra-arid climate, and quaternary lithological units. These characteristics can individually increase aeolian sand transport. Among surface characteristics, vegetation, topography [79] and sand material attributes [79] primarily control dune formation and development. Surface moisture affects aeolian processes, which govern the shape, speed, and direction of mobile sand dunes. Also, the humidity of the surface affects the trapping of sand particles [80]. It is also concluded that particular types of sand dunes develop as a result of fluvial–aeolian interactions [81]. Generally, the roads in desert areas need to control measures actively because the sediment on the windward side of the road tends to increase and can be a source for forming road sand. Additionally, the physiognomy changes in wind erosion caused by road construction threaten its stability [82].

According to Han et al. [82], our results corroborate that the steep, wind-facing slopes primarily influence dune movement. In addition, erosion rates correlate positively with environmental dryness due to reduced vegetation and soil moisture [83]. Quaternary deposits, composed largely of unconsolidated particles, are susceptible to water and wind erosion [84]. Machine learning techniques revealed land use and geology as the most influential factors for distinguishing sand dunes and non-sand flats (Figure 4). Altitude and soil characteristics also contributed significantly, as demonstrated in Figure 5. The mean decrease in accuracy, as depicted in Figure 5, quantifies the reduction in model accuracy when feature values are permuted, or features are removed.

The Gini index is used to select features and calculates the probability of a specific feature being classified incorrectly when chosen randomly. The Gini index performs only binary splits, and if all the elements are associated with a single class, it can be considered “pure” [85]. The Gini index ranges from 0 to 1, with 0 representing a pure classification and 1 indicating a random distribution of elements across different classes. Hence, the features with a lower Gini Index are prioritized and can be placed higher up in the created tree closer to the root [86].

Recent research indicates that climate conditions such as low rainfall, increased evaporation and transpiration, and the number of drought periods, while human activities have also played an influential role in the expansion of dunes [87,88,89]. These findings underscore the crucial role of vegetation (often removed through human activities) and soil moisture content (influenced by rainfall) in dune stabilization. These factors reinforce dune structural integrity and protect against wind and water erosion. Preserving and promoting vegetation growth, along with maintaining adequate moisture, are essential for long-term dune stability and sustainability. Numerous studies in desert areas support these findings [90,91,92].

Regarding land slope, the highest values of dunes were for 2 to 8% and the northeast slopes. It is evident that on the high slopes, due to the lack of accumulation of aeolian sediments, it is not possible to develop dune areas. The direction exposed to danger is consistent with the source of the general threatening winds in the area and can be incorporated into wind erosion control schemes. Also, all dune areas are outside the flood zone and far from the watershed’s stream network (i.e., 500 to 700 m). The reason for this is the occurrence of seasonal floods, higher humidity, and coverage of these areas. However, areas beyond 3000 m are at risk at variable distances from the road (Figure 3).

Two different methods, the IV (Figure 5) and DA (Figure 6) were utilized to compute the area of each class, the density of regions in each category, and the value of each type for all the examined indicators. Table 2 displays the outcomes of these computations for the most impactful classes. According to the dune expansion zoning map, developed based on the DA method, approximately 28% and 42% of the dunes were classified as very high and high categories of movement dunes, respectively. Conversely, in the IV input zoning method, 43% and 27% of dune areas were categorized as high, respectively (

Table 3. Results of DA and IV methods on the expansion of the dunes.

Risk Class	Dunes Areas (%)		Density Within Each Class (%)		Area Covered by Dunes (km2)		Area Covered by Each Class (km2)
	IV	DA	IV	DA	IV	DA	IV	DA
Very high	43.32	28.40	1.33	1.22	3.84	2.51	289.42	205.68
High	26.98	42.00	0.48	0.62	2.39	3.72	497.20	597.02
Medium	19.83	19.48	0.25	0.25	1.76	1.72	695.67	698.85
Low	9.57	7.49	0.16	0.12	0.85	0.66	539.12	549.80
Very low	0.29	2.65	0.01	0.07	0.03	0.24	348.83	316.71

). Both methods show similar maps, situated the dunes in the same places. The medium, low, and very low values also register identical values, but the main differences are in the very high and high values.

The Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) metrics, used to measure the accuracy of the models, are determined based on the true/false classified cases by the classification algorithms. These values are displayed in

Table 4. The RMSE and MAE values for the classification algorithms.

Classification Algorithm	RMSE	MAE
Decision tree	0.23	0.11
Random forest	0.21	0.10
Boosting aggregate	0.27	0.13
Support vector machine	0.19	0.08
Neural network	0.31	0.15

.

3.3. Accuracy and Validation

Our study indicates that the SVM algorithm is the most effective model for classifying erosion zones, outperforming four other models (RF, DT, BA). This conclusion is based on the ROC curve analysis, which showed that the SVM algorithm has the highest area under the curve at 94% (Figure 7). The high accuracy of the SVM algorithm compared to other data mining methods was attributed to its speedy computational performance and its use of optimization rules to locate optimal boundaries between user classes [93]. As a result, they can be introduced as a viable alternative to other classification algorithms. However, it should be acknowledged that there is no set pattern for all issues. Determining the best model for each problem requires repeated trial and error, as well as testing the methods under consideration.

Additionally, algorithms are generally considered because they save time, enable easier control, reduce the cost of investigations and errors compared to field methods, and analyze unlimited solutions [94,95]. However, specific older algorithms require a robust search strategy to identify the optimal subset of computational properties. The Support Vector Machine (SVM) algorithm is a reliable technique for data classification that offers several advantages over other methods. SVM is known for its high accuracy in categorizing data, similar to the actual types [96,97]. In addition, SVM’s training process is relatively simple compared to other algorithms, such as neural networks. One of the critical advantages of SVM is its ability to avoid getting stuck in local maxima, which can be a common problem for high-dimensional data [91]. Overall, SVM is a powerful tool for data classification that can deliver accurate results with a simple and efficient training process.

Researchers have successfully demonstrated the ability to identify the relationships between natural hazards and the variables that impact them using two-variable statistical methods [98,99,100] and machine learning techniques [101,102,103]. This approach enables the recognition of complex patterns and the prediction of potential risks, providing valuable insights to various industries and decision-makers [30,65,94]. These findings have significant implications for hazard mitigation strategies and emergency management planning and can help minimize the associated risks and damages. Over the past few years, there has been a significant increase in the use of two variable statistical methods and machine learning techniques for creating accurate maps of various environmental hazards, including landslides [104], gully erosion [102,103,105], water pollution [106,107], subsidence [108,109], aeolian features [110,111] and dust events [66,103,112]. These advanced techniques have proven highly effective in accurately modelling and predicting environmental hazards, leading to better preparedness and mitigation strategies for these events. Additionally, numerous studies have shown that although specific techniques have demonstrated better results, every approach has advantages and limitations, making it unfeasible to suggest a singular model for every domain [100]. Conversely, the statistical techniques combined with Geographic Information Systems (GIS) exhibit commendable accuracy when forecasting natural phenomena [113].

A support vector machine is a supervised learning method for classification and multi-class regression [114]. This method is relatively new and outperforms older classification methods [115]. In this method, an optimization algorithm is used to obtain examples of class boundaries, which are then used to calculate an optimal boundary to separate classes. These examples are referred to as support vectors [116]; this algorithm has also been extended to non-linear mode [117]. This method supports several different linear and non-linear kernels by default. The advantages of the support vector machine include the ability to solve complex classification problems with many layers and few training samples. The most important feature is the ability to overcome the problem of the non-linear distribution of training data [118]. The reason for the high accuracy of the SVM algorithm compared to other data mining methods is that this method is computationally very fast and uses optimization rules to locate the optimal boundaries between groups. Consequently, it can be used and introduced as a suitable alternative to other classification algorithms [93].

The results of the support vector machine algorithm using different kernels have shown that the radial basis function kernel is the most appropriate choice. The Radial Basis Function kernel in the Support Vector Machine is a robust machine learning algorithm that can be used for classification and regression tasks. A comparison of performance evaluation parameters found that the SVM RBF technique works well in contrast to the random forest and SVM polynomial techniques [119]. Using polynomial functions in SVM predicts and models prone areas better than linear functions, which appears to be due to the complex and non-linear behaviour of the variables involved in the phenomenon’s occurrence in the study area [120].

Data mining is a commonly used analytical technique in research and involves exploring various model areas due to the increasing number of algorithms available [121,122]. Recently, in arid land for the Aeolian process, it has been recommended to use classification algorithms as they provide precise calculations and eliminate the need for expensive, time-consuming, and sophisticated tools [101,123,124]. Additionally, classification algorithms are recommended to accurately classify aeolian regions as they provide precise calculations [125,126].

We concluded that the influence of climate, geomorphology, and human factors on the sedimentation of basins has been thoroughly established in much research in the world [127,128,129], and especially, in this case study. This refers to the execution of initiatives aimed at safeguarding soil and controlling sediment, which are crucial aspects of environmental conservation and management [130]. Effective implementation of these programs is essential to ensure sustainable land use practices and prevent adverse impacts on ecosystems and human activities. Furthermore, acquiring comprehensive information about sediment development is crucial to identifying critical areas within a watershed [131,132,133]. This involves a meticulous evaluation of various factors that contribute to sedimentation, including but not limited to geological, hydrological, and anthropogenic factors. Identifying such critical areas is imperative to devise effective mitigation strategies and ensure sustainable management of the watershed [134]. The Seqala watershed, which extends from the western foothills of the northern highlands to the arid and desert areas of Iran, has a gradual slope and lower elevation than the surrounding mountains. Seqala’s inhabitants face the threat of migration due to concerns about desertification. The area is also experiencing significant damage to its facilities and agricultural lands owing to water shortages and the influx of dunes caused by desertification. These challenges have contributed to a regional agrarian crisis that requires urgent attention. The research indicates that improving the identification and tracking of the spatial patterns of dune movements is feasible. It is essential to adopt various reforestation methods to tackle this problem effectively. The depletion of these natural resources can lead to severe and irrevocable harm to the environment.

Decision trees excel in simplicity, interpretability, and classification accuracy [135]. However, they are susceptible to overfitting and instability, producing varying trees and predictions with slight training set changes. Random forest algorithms mitigate these issues through randomized feature selection, achieving higher accuracy than decision trees [136]. Nevertheless, RF is slower and less interpretable than a single decision tree. Neural networks effectively recognize complex patterns and make predictions but require extensive training data [137], similar to the present study, often too expensive, time-consuming, and complex. Boosting models can significantly improve weak models’ accuracy and enhance prediction accuracy [138]. It is sensitive to outliers, which are often unavoidable in field studies and usually interfere with and obstruct the data mining process [139].

According to Kavzoglu et al. [93], the reason for the high accuracy of the SVM algorithm compared to other data mining methods is that the SVM method is computationally speedy and uses optimization techniques to find the optimal separation boundary that maximally separates the classes. Consequently, it can be introduced as a suitable alternative for other classification algorithms. Of course, it should be stated that there is no predefined pattern for all issues, and determining the best model for each problem requires repeated trial and error and testing the methods in the desired conditions.

The applied classification algorithms only deal with the spatial investigation of the aeolian sand transport but do not give any estimate of the time of occurrence. Therefore, when the probability of an event is unpredictable, we face uncertainty. In the context of aeolian sand transport risk, it is impossible to determine all the quantities influencing the measurement result. Therefore, those with the most effects and their impact on measurement results can be identified. This limitation causes uncertainty in modelling. Also, uncertainty in measurement is a parameter related to measurement results such as testing and calibration, which determines the range of values related to the measured quantity. Components such as complexity and uncertainty in the natural environmental factors, inaccuracies, and dynamics in surrounding conditions [140] can also be a source of uncertainty in measuring and modelling the data in this research.

3.4. Wind Analysis and Sand Drift Potential

The results of the analysis of wind data in the annual scale over the last two decades showed the abundance of northeast (18.84%) and west (12.78%) winds, although, in the cold seasons, the west winds and in the warm seasons, the winds of the northeast and east have a higher incidence. Furthermore, it has been shown that south (3.96%), southwest (4.09%), and southeast (4.09%) winds have the least frequency and play no significant role in wind erosion processes (Figure 8a). The findings of the assessment of calm wind also indicate that it is equal to 31.72% on the annual scale, and its lowest and highest amount is in the summer (28.35%) and winter (38.41%) seasons, respectively (Figure 8b). The evaluation of average wind speed in Ferdows’ station has shown that the highest and lowest average wind speed are in summer (3.08 m s⁻¹) and winter (2.17 m s⁻¹) seasons, and it is equal to 2.63 m s⁻¹ on an annual scale (Figure 8c). The results of the drift potential show that it is quantitatively equivalent to 36 VU, Furthermore, the direction of movement of flowing sand in this region is from the western and northeastern areas to the eastern and southeastern regions (Figure 9).

We confirmed that various factors influence dune movement and expansion, including climate, topography, substrate, and human activity. Numerous studies have explored how these factors interact to drive changes in dune morphology over time [141,142,143]. Over the past few years, numerous studies have been conducted in aeolian environments, utilizing machine learning techniques and advanced statistical methods [101,122,123]. This scientific area is especially relevant in young aeolian environments that may hurt local communities. Therefore, it is essential to explore the potential consequences of such developments and develop strategies to mitigate their effects on the loss of sustainability. Due to the aeolian sand movement in Iran, a decision was made to couple two-variable statistical methods (information value, density area) and machine learning to predication sand dune movements based on climatic-environmental characteristics within the Seqale watershed in eastern Iran, in which sand movement has already brought significant detrimental effects to human societies.

4. Conclusions

Research in the Seqala watershed revealed significant aeolian dune transport towards the centre and southwest, affecting over half the area. Our study demonstrates the feasibility of improving spatial pattern identification and tracking dune movements to preserve sustainability. The Support Vector Machine (SVM) algorithm outperformed other models (RF, DT, BA) as the most effective for classifying erosion zones. Factors such as altitude (1200–1400 m), gentle northeast-facing slopes (2–5 degrees), rainfall (250–500 mm), high evaporation (2500–3000 mm), locations outside flood plains, and distances greater than 3000 m from roads and 500–700 m from water channels correlate with higher sand dune presence. Dune expansion maps, particularly those developed using the density area (DA) method, classified approximately 28 and 42% of dunes as very high and high movement categories, respectively. These machine learning-derived maps offer valuable tools for executive managers to implement targeted spatial actions for efficient problem mitigation in the region. The Seqala watershed, characterized by a gradual slope from northern highlands to central desert areas, faces a migration threat due to desertification impacting infrastructure and agriculture through water shortages and dune encroachment. The potential future agrarian crisis necessitates urgent attention and the adoption of reforestation and other ecosystem preservation methods. Machine learning algorithms can significantly enhance management efforts by enabling spatially targeted interventions.