Introduction

Locating the landfill sites as the disposal location of waste materials that are of no use for recycling has been a pivotal issue for municipal and industrial waste management agencies (Ngoc and Schnitzer 2009; Khan et al. 2018). Despite the official reports concerning the considerable waste generation per capita in rural areas accounting for about 220–340 g per day in 2016 as well as the regulations designed to govern waste management in Iran (Khayamabshi 2016; Tehran times 2017), the latter has come to a head and municipalities have begun to identify safe landfilling locations and propose modern disposal techniques such as digesters, incinerators, and biogas (del Valle-Zermeño et al. 2015; Khayamabshi 2016). Landfill site locating is substantial in developing countries where environmental restrictions and organizational benefits may occasionally stand on opposite sides. The countless poorly located landfill sites at or nearby the rural areas have been threatening the surrounding environment including its inhabitants by producing an unpleasant odor, diminishing esthetic values, and contaminating water resources (Rathi 2006; Mondelli et al. 2007; Li’ao et al. 2009; Khorram et al. 2015). Open garbage dumps and uncontrolled landfills have been significantly affecting the quality of the streams in the Ardak watershed in the north of Khorasan Razavi province, Iran. According to the official reports of Water Company and Forest, Range and Watershed Management Organization of Khorasan Razavi province, effluents from landfill leachates have contaminated nearby surface waters as well as groundwater resources by injecting different pollutants such as xenobiotic organic compounds, heavy metals, and nutrient salts (RWCKRFootnote 12018). Hence, precise and effective landfill site selection techniques are essential for water quality management to address downstream pollution concerns.

Rating, weighting, and ranking have been the most common basis of landfill site selection so far with which many novelties have been offered in terms of model integration, data inclusion/configuration, and geographical toolboxes with setback spatial criteria go forth. Developing a GIS tool underpinned by the ensemble of Not-In-My-Backyard (NIMBY) syndrome and nuisance score (Simsek et al. 2014a, b) and incorporating different multi-criteria decision-making methods such as simple additive weighting (SAW), analytical hierarchy process (AHP), and TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) (Rahmat et al. 2017, 2018; Chabuk et al. 2017; Spigolon et al. 2018; Khan et al. 2018; Șener et al. 2010; Kamdar et al. 2019; Chabuk et al. 2019; Özkan et al. 2019), remote sensing studies with employing aerial photography, airborne and spaceborne sensors (Manzo et al. 2016; Mahmood et al. 2017), novel Thiessen polygon–based ranking schemes integrated with remotely examined environmental criteria (Richter et al. 2019), and electrical resistivity combined with a MCDM method (Goulart Coelho et al. 2017; Yousefi et al. 2018; Akintorinwa and Okoro 2019) are some of these endeavors.

Although such methodologies have expanded the knowledge on the ways to better concentrate on the desired landfill sites which are most commonly termed as areal constraining/filtering, the current trail of literature has not considered the element of data mining and pattern recognition nor acknowledged their merits. Here we proposed a novel data driven-probabilistic ensemble model that bears not only on-site selection but also recognition of similar patterns with different degrees of similarities to the ideal conditions. The latter, in particular, is a key principle when authorities require a highly differentiated spatial map that includes the transient zone which may be transferred into either suitability or unsuitability conditions over the course of time. Such a technique would simply improve traditional binary classifications. In this process, we attempted to employ the most easy-to-acquire environmental data in order to provide an all-inclusive application of the proposed methodology. Hence, the main contribution of this paper is the probabilistic site selection scheme which would serve the decision-makers with much more flexible decision-making maps—a theme that is beyond a binary presence-absence classification.

Study area

The Ardak watershed extends for an area of about 479 km2 and is located in the north of Khorasan Razavi province, Iran. It lies between latitudes 4,066,633 to 4,096,281 N and longitudes 690,339 to 725,597 E, UTM zone 40 N. Elevation ranges between 1235 and 2950 m.a.s.l (Fig. 1). A total of 15 residential centers are evident which are sporadically distributed across the study area. Roads at 55.5 km—paved an unpaved—are extended across the region. A 146-km stream network is sprawled over a diverse land cover types. Land covers were disintegrated in terms of type and the density of the vegetation cover and usage-wise to attain more realistic and precise results. Based on the latter, poor pastures cover the largest area (74.5% of the entire region), followed by moderate-condition pastures (8.5%), and very-low-density forests (8.1%) while the remaining area is shared between orchards, rainfed farming, low-density forests, and woodlands. The appearance of 238 springs across the region intuitively implies the high groundwater potential in the area. Moreover, there is an important dam reservoir in the southernmost part of the study area. Mashhad, one of the largest metropolitan cities of Iran, is located at the reservoir’s downstream end and accommodates above 3 million residents. The latter indicates the pivotal role of this watershed in providing the water supply of the downstream regions. Many garbage dumps were evident mostly nearby the streams, roads, and residential centers during the filed surveys where an instant change in natural color and quality of surface water were discernible right after the contaminated area, endangering the natural habitats and human beings (Fig. 2).

Fig. 1
figure 1

The geographical location of the study area in Khorasan Razavi province in Iran. Right, an oblique view of the region

Full size image
Fig. 2
figure 2

Some photographs of open garbage dumps nearby streams (a, c, e) causing appearance of harmful green algae in water (b) and fish die-offs (d) due to the contaminations

Full size image

Materials and methods

The workflow of this study is presented in Fig. 3. What follows is thorough explanations of the adopted methodological stages.

Fig. 3
figure 3

Methodological flowchart adopted in this study

Full size image

Data compilation

Although most of the spatial modeling efforts require inventory data for the modeling or model validation phases, the methodology proposed in this work has an inventory-irrespective scheme. The process consists of two main phases, namely, areal constraining and probabilistic modeling. In fact, the final product of areal constraining signifies the most suitable sites for landfills which, itself, provides the inventory for probabilistic modeling. This procedure follows the selection and classification of landfill-representative factors. The adopted factors were selected based on different criteria such as literature review, data availability, spatial heterogeneity of the factors’ pattern across the study area (i.e., excluding homogenous factors with less unique information), and, most importantly, the multicollinearity issue. The latter can cause bias in models’ results due to the existing strong correlation between the used thematic maps which should be obviated prior to the modeling process. In this work, we used the variance inflation factor (VIF) and its reciprocal, tolerance index, to assess the strength of correlation between independent variables (i.e., thematic maps) and the multicollinearity issue. The VIF value higher than 5 and accordingly the tolerance index lower than 0.2 indicates critical multicollinearity (O’brien 2007). In this regard, Kutner et al. (2004) also suggested that the VIF values higher than 10 can represent critical multicollinearity (Kutner et al. 2004); however, the VIF value of 5 was considered the basis of factor exclusion procedure to ensure the latter issue (Table 1). The correlation matrix is another way to address the strong correlation between the factors in a pairwise manner which can be calculated with different methods such as principle comparison analysis (PCA), yet the VIF test is comparatively more elaborative and distinctly points out the responsible factor for the multicollinearity issue, although both methods revolve around the same concept and selecting between them is the matter of choice. Additionally, the VIF method has been increasingly used and getting more common in the literature specifically for multicollinearity issue (Kornejady et al. 2018; Chen et al. 2017; Haghizadeh et al. 2017; Rahmati et al. 2019; Moghaddam et al. 2020a, b).

Table 1 The VIF values representing the multicollinearity between different opted factors
Full size table

In light of these sieving procedures, eight predictors (i.e., drivers) were selected including slope degree, proximity to faults/reservoir/residential areas/roads/springs/streams, and land cover types. The thematic maps of the conditioning factors were prepared in ArcGIS 10.3 as presented in Fig. 4. More detailed information regarding the factors’ scale, their role, and the areal exclusion thresholds are provided in Tables 2 and 3. Constraining threshold values were adopted from similar works such as Şener et al. (2010) and Abd-El Monsef and Smith (2019). The final product of areal constraining would give us the proper place to extract training and validation points for probabilistic modeling. By doing so, two sets of 1000 random points were selected each respectively per training the Mahalanobis distance model and validating the results. The number of samples was arbitrarily defined just to ensure the data scarcity issue and to be certain of the adequacy of samples both for training the model and validation of its results. The latter, in fact, guarantees that the balanced samples would satisfy the training and validation tasks since the number of samples are abundant which ease the data noise issue derived from inadequate sample number.

Fig. 4
figure 4

Causative environmental factors used for areal constraining, a slope degree, b proximity to faults (m), c proximity to the reservoir (m), d proximity to residential areas (m), e land cover types, f proximity to roads (m), g proximity to springs (m), and h proximity to streams (m)

Full size image
Table 2 Role of the opted factors in landfill site selection, their spatial scale, and established exclusion zones
Full size table
Table 3 Assigned rates to different land covers based on their compatibility with a landfill site (adopted from Șener et al. 2010)
Full size table

Most of the landfill site selection studies wind up in a mere spatial constraining; however, the product of such procedure, despite its substantial merits, is only a binary classification (i.e., suitable or unsuitable) at a rather rudimentary level since areal constraining follows an arbitrary factor classification scheme. Hence, probabilistic models can enable the decision-makers to avoid such determinism in identifying the most suitable sites and not only define the latter in probability terms but also give a chance to other areas with somewhat similar suitability degrees. Such areas can be classified as the transient zone which can locate in a flexible zone somewhere in between the most and the least suitable zones. Such a view also gives the authorities the opportunity to prioritize different options rather than relying only on two distinct and restricted zones. As opposed to deterministic models that only consider a single outcome, probabilistic models consider the geo-environmental random variables (here predictors) and their probability distributions to consider different possibilities and are very useful in predicting future events and patterns.

Given that landfill sites are subjected to different natural factors that act as random variables, probabilistic modeling pays the way to identify different locations with similar conditions as well as avoids determinism in site selection which, itself, can be problematic when dealing with such environmental agents. Such probabilistic modeling was carried out by incorporating the areal constraining scheme with a robust pattern recognition algorithm as explained in the following section. What follows is the detailed explanation of the proposed landfill suitability mapping framework.

Landfill site selection using the coupled MCDM-Mahalanobis distance

Determination of landfill sites merely based on areal constraining as a multi-criteria decision-making method would provide binary presence/absence locations at a rather rudimentary level which may manifest shortcomings in terms of the uncertainty emanated from the arbitrary nature of factor classification and combination techniques. Instead, feeding such localities into a powerful modeling scheme with the capability of pattern recognition would provide the final suitable sites in probabilistic terms which is in more favor of decision-makers (i.e., avoiding absolute certainty and determinism in their reports) and also has prediction merits. Such an information retrieval process is attainable by the Mahalanobis distance approach. Mahalanobis distance was first expounded by Mahalanobis (Mahalanobis 1936) and ever since some application of this method has been reported such as data enrichment in landslide susceptibility modeling (e.g., Tsangaratos and Benardos 2014; Kornejady et al. 2018), outlier detection (Todeschini et al. 2013; Leys et al. 2018), and proposing novel validation techniques (Zhao et al. 2017). In a nutshell, it measures the similarity of different locations with different physical conditions in the study domain (i.e., different pixels with different environmental, geological, or anthropogenic factor values) to the target conditions (i.e., the conditions of the rudimental sites for landfill). In other words, locations (i.e., pixels) with the similar values of conditioning factors to those of landfill suitable sites would be equivalently considered suitable. The notion of suitability in Mahalanobis distance is translated into the average condition. For instance, if the average value of slope degrees at the rudimental landfill sites is equal to 5°, the latter would then signify the suitable slope for landfill sites. The same applies to the other environmental factors with the difference that qualitative factors such as land cover should be first quantified in order to become mathematically analyzable. Moreover, the notion of similarity to the ideal locations is Mahalanobis distance is rendered into a distance in which the closer distances to the ideal location indicates a higher probability of the overall environmental suitability. Such distance follows the expression (Fukunaga 1990):

$$ Dm=\sqrt{{\left(x-m\right)}^T{\sum}^{-1}\left(x-m\right)} $$
(1)

where x is the vector of conditioning factors, m is the factors’ mean value, T is the transposed matrix of the factors’ values, and Ʃ−1 is the inverse covariance matrix. In order to transform the distance values into probabilities, chi-square values of Mahalanobis distance values will then be transformed into P values based on the chi-Square distribution with N−1 degrees of freedom (number of variables 1). The latter was carried out in Land Facet Corridor Tools extension in ArcGIS 10.3.

Performance metrics

The area under the success and prediction rate curves (AUPRC and AUSRC), as probability cutoff-independent metrics, will be the judge of the model’s performance in terms of goodness of fit and prediction power, respectively. The success rate curve plots the cumulative percentage of suitable areas (from the highest values to the lowest) on the X-axis against the cumulative percentage of the training set on the Y-axis (Chung and Fabbri 1999; Kornejady et al. 2018). The same applies to the prediction rate curve with the difference that, instead of including the training set, the validation dataset is used. The AUSRC and AUPRC values close to 100 (i.e., 1 in the range of 0–1) indicate a perfect model (i.e., highly trained and highly predictive), while values close to 50 (i.e., 0.5 in the range of 0–1) signify a neutrally operating model with an unsatisfying performance. Plotting the SRC and PRC was executed in the recently developed performance analyses toolbox in ArcGIS, named PMT (Rahmati et al. 2019). It is noteworthy that the presence-absence points were balanced (i.e., an equal number of inventory and noninventory locations) in order to calculate the performance metrics. The balanced sample partitioning technique was selected due mainly to avoid the variance inflation in the results (Petschko et al. 2014).

Additionally, three holdout-dependent performance assessment metrics including accuracy (efficiency), precision, and sensitivity was calculated both in the training and validation stages following Eqs. 2–4 (Rahmati et al. 2019).

$$ \mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{T} $$
(2)
$$ \mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}} $$
(3)
$$ \mathrm{Sensitivity}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}} $$
(4)

where TP, TN, FP, and FN, respectively, denote true positives, true negatives, false positives, and false negatives.

Results and discussion

Table 1 provides in detail the multi-collinearity values between the controlling factors of landfilling sites based on which the highest VIF values correspond to proximity to residential areas, proximity to springs, and proximity to the reservoir, while the lowest values are mainly attributed to land cover and slope angle. The results indicated that there is no critical multi-collinearity between the factors since the VIF values are in an acceptable range (i.e., less than 5). The latter is due to the fact that the authors avoided involving different DEM-derivative indices that have been commonly used in spatial modeling, but rather factors that can decisively delineate the landfilling sites are engaged. Among these, gentle slopes are the first opted criterion for most landfill sites and other criteria are sieved afterward. Faults are also critical gateways in such a way that contaminated water from landfills can percolate into the deep strata and contaminate the groundwater resources. Hence, landfill sites should be simply far from the faults. Dam reservoir is also considered a critical source of water supply for the residential centers, particularly Mashhad, and agricultural uses for arable lands. Moreover, the inhabitants of residential areas in the region would be bothered by the unpleasant odor of landfills as well as other imminent diseases. Regarding the land covers, most national environmental laws are strictly against siting landfills accordingly in dense forests, woodlands, good-condition pastures, and orchards, while nonarable lands and poor pastures are more desirable for such a purpose (Department of Environment 2018) (Table 3). Although roads are necessary for transporting waste to landfilling sites, considering setback margins is necessary particularly for the local residents commuting from adjacent areas. The presence of springs indicates high groundwater potential in the study area. Hence, the farther the landfills from the springs, the better. Such areal filtering resulted in narrowing down that study area to the final extent which is suitable for siting landfills. As presented in Fig. 5, suitable sites identified at the rudimentary level are mostly concentrated in westernmost parts of the area while some sporadic spots are evident in the north, center, and south which altogether account for about 2% of the study area (i.e., 915.12 ha).

Fig. 5
figure 5

Final delimited area derived from the stepwise areal constraining procedure

Full size image

These sites, however, resulted from arbitrary classifications, and one cannot decisively introduce the final areas as the most suitable sites. Additionally, such a constraining process would signify the remaining area as unsuitable due to the binary format of the final map, while areas with lower suitability yet somewhat similar to the ideal sites would be left undiscovered. Moreover, even the sites identified as ideal by areal constraining need further investigation and cannot be determinately selected for pragmatic actions. Such an issue was obviated by the context of Mahalanobis distance. Extracting training and validation points (Fig. 6) and feeding them, respectively, into the modeling process and model validation created a probabilistic map for siting landfilling areas. The mean values, correlation matrix, and covariance matrix are, respectively, presented in Tables 4, 5, and 6 which together made the main components of the Mahalanobis distance. The mean values revealed that preselected suitable landfilling sites are located at an average distance of 3.7 km from the faults, 4.3 km from residential areas, 16 km from the reservoir, 4.4 km from roads, 2.8 km from springs, 1.7 km from streams, and a slope degree of 7.7°. Covariance and correlation matrices give valuable information regarding the dependency between the factors. What sets them apart is that in addition to the covariance matrix that indicates the direction of the linear correlation between factors, the correlation signifies the strength and direction of such a relationship in a limited range of + 1 and − 1. The correlation value of 0.71 between the proximity to springs and residential centers is also in accordance with the VIF results and the positive sign of the values also signifies the positive direction of their correlation. Mahalanobis distance uses such spatial dependence and the interconnection between the factors instead of considering them completely independent.

Fig. 6
figure 6

Extracted training and validation (presence) samples from the constrained domain as well as the noninventory (absence) samples from the remaining area

Full size image
Table 4 Factor’s mean values at the training samples
Full size table
Table 5 Correlation matrix representing the strength and direction of the normalized dependence between the factors
Full size table
Table 6 Covariance matrix representing the direction of the un-normalized dependence between the factors
Full size table

Landfill suitability map derived from Mahalanobis distance is presented in Fig. 7. The excerpt area shows that based on the probabilistic pattern recognition algorithm of Mahalanobis distance, the target regions derived from areal constraining do not necessarily pinpoint the high suitability but spatially varying suitability pattern may exist within the final constraints. More interestingly, a suitable patch is also evident in the bottom right corner of the excerpt which has been previously missed by areal constraining. The latter indicates that the rigidity of areal constraining may ignore spatial similarities due to lacking a perfect condition (i.e., being entitled to all the constraining rules), while a flexible algorithm acknowledges the random variation. In other words, if fully deterministic analysis (i.e., areal constraining) is taken into account, considering the same constraining rules on environmental factors in different areas would always give the same results. However, factors’ randomness is highly appraised through many simulations which are mostly based on the distribution of each environmental factor. Since environmental factors often exhibit random variation, considering such matter would have a major influence on final results.

Fig. 7
figure 7

Landfill suitability probability map obtained from the ensemble of stepwise areal constraining and Mahalanobis distance models

Full size image

The coupled Mahalanobis distance and areal constraining techniques were found to have an AUSRC value of 89.36 (Fig. 8a). According to a classification range proposed by Hosmer Jr et al. (2008), the resulted value indicates that our ensemble model has almost shown an outstanding training performance (i.e., close to 0.9). Interestingly, the predictive power of the model is also very close to an excellent performance where the AUPRC value amounted to 89.25 (Fig. 8b). The latter highlights the merits of probabilistic spatial modeling especially when spatial dependencies between the covariates are taken into account. The highly identical values of AUSRC and AUPRC suggests two important features: (1) a model that has obviated the overfitting issue since the higher value of goodness of fit has not backfired into yielding a lower predictive power, (2) a robust model where changes in the inputs do not much influence the model’s performance, and (3) an inadequate segregation between the training and validation datasets due mainly to selection of the input data from a very small domain (i.e., 2% of the entire study area). The latter can be further assessed by spatially differentiating the training and validation dataset so that the former is selected from a separated area and fed into the model while another region is kept apart for validating the previous results.

Fig. 8
figure 8

Success rate curve (a) and prediction rate curve (b) generated in the PMT toolbox

Full size image

In addition to the ROC curve, the holdout-dependent performance metrics were calculated (Table 7) based on which, in line with the results of ROC curve, the goodness of fit and prediction performances of the proposed suitability assessment framework found to be good which remains almost intact in both training and validation stages as a criterion of the robustness and reliability of model’s results through the modeling and prediction stages.

Table 7 Accuracy, precision, and sensitivity values calculated for the Mahalanobis distance model at the training and validation stages
Full size table

The literature review showed a discernable pattern of adopting areal constraining as the basis of landfill suitability zonation. As such, Simsek et al. (2014a, b), Manzo et al. (2016), Mahmood et al. (2017), Rahmat et al. (2017a, b), Chabuk et al. (2017), Goulart Coelho et al. (2017), Yousefi et al. (2018), Spigolon et al. (2018), Khan et al. (2018), Șener et al. (2010), Kamdar et al. (2019), Chabuk et al. (2019), Özkan et al. (2019), Richter et al. (2019), and Akintorinwa and Okoro (2019) although attempted to incorporate the most novel methods to map suitable landfill sites, ultimately ended up in preparing a binary classification. As opposed to the traditional binary classification (i.e., suitable or unsuitable) derived from areal constraining methods, the probabilistic maps are highly differentiable in such way different transient zones can be located in between the suitable and unsuitable classes. Here, we chose the natural break classification scheme and classified the Mahalanobis distance-based landfill suitability map into two common classes of five-point and three-point scales (Fig. 9). A higher number of classes will, in turn, result in a reduced areal extent of each class. Selecting either of these two classifications would also result in different areal delineation, and hence different budget and time allocated to the area. Highly differentiated maps may be more in favor of on-site pragmatic actions, while less differentiated ones are of more general use in less detailed projects and less urgent circumstances. The areal extent of the suitability classes obtained from the five-point scale and three-point scale classification of the Mahalanobis distance map are also provided in Tables 8 and 9.

Fig. 9
figure 9

Five-point scale (a) and three-point scale (b) classification of landfill suitability

Full size image
Table 8 The areal extent of the suitability classes obtained from a five-point scale classification of Mahalanobis distance map
Full size table
Table 9 The areal extent of the suitability classes obtained from a three-point scale classification of Mahalanobis distance map
Full size table

Limitations and future works

The difference in spatial resolution of the thematic maps is a basic yet critical limitation of any spatial modeling work. This issue roots mainly from data scarcity especially where high-resolution maps are not provided due to being an exorbitant task for organizations. In such a predicament, the best getaway to govern the available data is in favor of the detailed data (i.e., here DEM-derivatives as the limiting factors) rather than calculation space. Hence, all the spatial resolutions were transformed into the resolution of DEM-derived factors which, in this way, guarantees that no data have been lost. The other issue regards the sieving process and selection of the most influential factors. In this regard, soil permeability and wind velocity (i.e., speed and direction) are also among the most determinant factors. However, the former would require a detailed map consisting of soil physical properties including permeability yet such detailed data were not available for our study area. Although the Harmonized World Soil Database (FAO I, ISRIC I 2012) provides valuable information on soil properties on a global scale, using such data at a watershed scale would not be technically applicable. Hence, we avoided engaging less detailed data in the modeling process alongside other factors. Concerning wind velocity, it is noteworthy that, despite its invaluable information in identifying the exposed area to odor nuisance, generating the wind velocity map would require the synoptic data, while there is no synoptic station in the study area. Additionally, we strived to fill this gap by engaging another complementary factor, distance to residential areas, as potential sites were ultimately proposed to be located in the farthest areas possible from the residents. Nonetheless, the main objective of this study is to propose a globally generalizable modeling framework for landfill suitability assessment and certainly engaging more factors with unique information would improve the final performance of the proposed model. As stated in different literature (e.g., Allen et al. 1997; Allen et al. 2003; Simsek et al. 2014a, b; Barakat et al. 2017; Rahmat et al. 2017; Alkhuzaie and Janna 2018; Deswal and Laura 2018), abovementioned maps are pivotal for landfill suitability assessment and should be used in future analyses.

The last issue corresponds to the validation stage of the model’s results which stems from the fact that there might not be a precise and generalizable concept of a suitable landfill site in the first place. In other words, validation of the landfill suitability map can be carried out by the actual landfill sites that have been proved to be perfectly selected in terms of geo-topological criteria and residents’ qualification metrics. However, we presented a pseudo-validation test in which validation points are selected from the sites that are the most ideal sites for landfills. For future studies, it is suggested to compare the proposed technique with other machine learning and deep learning techniques with flexible probabilistic themes in terms of their performance and robustness. Although almost intact values of performance metrics through the training and validation stages somewhat indicate the stability of the model, further analyses need to be conducted to assess the overall robustness which can be ascertained by k-fold cross-validation techniques.

Conclusion

Uncontrolled landfills and open garbage dumps have been significantly contributing to surface and groundwater resource pollution and become a major concern for municipalities of the Ardak watershed and adjacent cities. The watershed exhibits high environmental and socio-economic values. The take-home messages of this work are as follows. Despite the merits of areal constraining techniques in narrowing down the domain to delineate the optimal landfill sites, it is underpinned by determinism while a probabilistic model is required to consider both the inherent randomness within the factors’ nature and spatial dependence between them. Mahalanobis distance supports such objectives and based on our results, gives a satisfying performance. The latter is apparent from the high goodness of fit and prediction power of the adopted model. Coupling Mahalanobis distance to areal constraining builds a powerful inventory-irrespective ensemble model. Areal constraining showed that almost 98% of the study area is unsuitable for landfilling, while the suitability map derived from the coupled areal constraining and Mahalanobis distance avoids such a rigid and rather inflexible response and appoints the end user for such a deduction. In other words, depending on the differentiation degree of interest, different suitability percentages can be obtained and used for budget allocation. Higher differentiation (i.e., a five-point scale classification) would normally result in more strict areal concentration based on which 2.6% of the study area has very high suitability for landfilling, while 2.7% of the entire area is highly suitable. Additionally, a less strict classification scheme (i.e., three-point scale) suggests that up to 4.6% of the entire region can be highly suitable for landfilling, while the remaining area has low-to-moderate suitability. By using the coupled Mahalanobis distance and stepwise areal constraining method, we have provided a framework of landfill management for generating advice to managers and policymakers concerned with downstream water quality. The conventional suitability assessment techniques provide the authorities with a binary classification which is limited by a mere presence-absence scheme and accordingly distorts the reality of the area. Conversely, the proposed technique provides a flexible map that can be classified into different suitability ranges according to the priority of decision-makers. In such a way, different areas will be treated relying on their true potentials while no place is left alone or forced to be categorized into only two specific classes.