Unraveling the Spatial Signature of Gully Erosion in the Arid and Semi-Arid Regions of the Northeast of Iran: Every Single Factor Matters!

doi:10.21203/rs.3.rs-2483100/v1

Download PDF

Research Article

Unraveling the Spatial Signature of Gully Erosion in the Arid and Semi-Arid Regions of the Northeast of Iran: Every Single Factor Matters!

https://doi.org/10.21203/rs.3.rs-2483100/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The present study is aimed to model the spatial pattern of gully distribution across the silt-rich soils of the northeastern loess of Golestan province, Iran. To this end, multiple data-involvement scenarios were designed, and their further contributions to the performance of a presence-only probabilistic model (Maximum Entropy) were assessed. The results revealed that the interconnection of topological and hydrological factors mainly corresponds to the spatial signature of gully distributions across the study area. In particular, valley depth, height above the nearest drainage, and drainage networks contributed the most to the model's performance. Based on the area under the receiver operating characteristic curve (AUROC) index, engaging factors collectively result in the highest goodness-of-fit and prediction power with the respective values of 0.889 and 0.867. The areal extent of the susceptibility classes indicates that approximately 15% (~ 121 km²) of the region is highly susceptible to gully occurrence, which entails further pragmatic actions and mitigation measures.

Maximum Entropy

Goodness-of-fit

Prediction power

ROC

Golestan Province

Amongst different types of soil erosion, water erosion is a widespread phenomenon responsible for tons of soil loss annually washed away towards the watershed outlets or local natural/artificial detention ponds (Berhe and Kleber 2013). Depending on the scale, the involved contributing factors, and the erosion facets, water erosion can be classified into different categories, from which splash, sheet, rill, and gully erosion are the ones globally recognized (Smith and Wischmeier 1957). Gully erosion is one of the most pervasive and morphologically emerged erosion types, which is mostly exhibited as land incisions due to gradual soil dissolution caused by drainage systems channeled across subsurface soil (Imeson et al. 1980; Poesen et al. 1998; Casalı et al. 1999). The synergistic act of climatic agents- mainly rainfall-runoff generation mechanisms, top/ subsurface soil conditions, and anthropogenic agents (i.e., mostly in the form of human interventions) such as unbridled farming operations are amongst the main drivers of gully formation (Poesen 1993; Fitzjohn et al. 1998; Douglas-Mankin et al. 2020). In some areas, hydrological drivers are key in exacerbating gully propagation across unprotected lands, especially when soils with sparse vegetation are predisposed to erosion. It consequently causes an interaction between gullies and the hydrological system, changing the watershed hydrology (Poesen et al. 2003; Poesen 2011). Gullies mainly propagate retrogressively from headcut, a vertical drop in gully channel-bed, which depending on the level of longitudinal and transverse growth, ephemeral and permanent gullies are formed. The former can be easily eliminated by tillage practices, while the latter can suddenly migrate by hundreds of meters in width and tens of meters in depth and cannot be easily eliminated nor controlled (Frankl et al. 2013; Capra and La Spada 2015).

Northeastern Iranian loess is gradually generated and developed in response to the deposition of airborne dust and experienced at least several erosion, transportation, and deposition phases. Its hexagonal structure forms a network of holes containing humidity as the result of a temperate climate prevailed in the region, and together with its silt-rich constituents, the soil is highly susceptible to being eroded by hydrological drivers (Nachtergaele et al. 2002; Hessel and van Asch, 2003). The distribution of loess formations near the drainage networks and the presence of erosional facets attests to the latter. Moreover, the intensive use of fertile loess soils for agricultural use and grazing—sporadically marked by sheep traces can accelerate the erosional processes, among which gullies seem to be the most pervasive type. Consequently, many gas pipes have conspicuously stuck out of the ground, and roads with poor foundations entirely washed from underneath due to the active role of runoff and the susceptible soil across the area. It is evident that gullies occur under the presence of specific spatial signatures, which can be explained and contextualized through spatial data-driven models.

Data-driven models are based on evidence and a set of predisposing/triggering factors, known as causative factors or predictors. Many data mining and machine/deep learning models are developed, enabling users to emerge the hidden patterns and predict the future pattern. The latter, for years, was the next promising horizon for traditional data mining models, which was fulfilled by machine/deep learning models. Application of Random Forest (Kuhnert et al. 2010; Jiang et al. 2021), Classification Trees (Svoray et al. 2012; Pouyan et al. 2021), Support Vector Machine (Pourghasemi et al. 2022; Wang et al. 2021), Artificial Neural Networks (Pourghasemi et al. 2017; Rahmati et al. 2017), and Convolutional Neural Networks (Gafurov and Yermolayev 2020; Saha et al. 2021) are only a few examples of the previous use of probabilistic data-driven models for spatial prediction of gully erosion in the literature. However, a few handfuls of research have been carried out in the northeastern gullies of Iran in the sense that spatial signatures have not been yet studied, nor the spatial pattern of susceptible areas has been investigated and mapped.

Based on these premises, this study sets out to strategize the selection of the most representative gully-controlling factors in which different spatial modeling attempts are passed through a data-involvement scenario-based sieve of performance assessment. Eventually, the optimal set of causative factors is determined, and their amalgamated spatial signature of gully occurrences is further discussed.

The methodological framework of this research is illustrated in Fig. 1, which mainly includes the selection strategy of gully-controlling factors, sampling strategy of gully samples, and construction of the data-involvement scenario-based sieve of performance assessment for the selection of best-performing modeling instance.

Study area

The study area is situated in the northeastern part of Iran, in Golestan Province (Fig. 2). The area extends for about 784 km². The maximum and minimum elevation ranges from 160 to 1490 m.a.s.l. A temperate Mediterranean climate is prevailed in the area, with the minimum and maximum precipitation of 346 and 610 mm. The most dominant land use/cover corresponds to agricultures (45%), followed by rangelands (38%), forests (16%), and residential areas (1%) (Table 1). The formation symbolized as Qsw, formed of grey to black shale and thin layers of siltstone and sandstone and Cenozoic in age, is the main geological outcrop in the area, accounting for 73% portion of the region. Table 2 details the areal extent of the geological formations in the study area.

Table 1

Land use classes in the study area
Land use	Area (he)	Area (%)
Forest	12513.04	15.95
Residential Areas	498.6	0.64
Rangelands	29858.8	38.07
Agricultural	35568.2	45.35

Table 2

Description of the geological formations in the study area
Geo unit	Description	Age	Area (ha)	Area (%)
Qm	Swamp	Cenozoic	2269.48	2.88
Qsw	Grey to black shale and thin layers of siltstone and sandstone	Cenozoic	58087.92	73.73
Ksn	Ammonite bearing shale with interaction of limestone	Mesozoic	9886.63	12.55
Ksr	Grey thick - bedded limestone and dolomite	Mesozoic	4906.5	6.23
Jmz	Olive - green shale and sandstone	Mesozoic	1857.68	2.36
Ekh	Swamp	Cenozoic	1772.73	2.25

Data construction phase

The main data architecture required for spatial modeling of any natural phenomenon revolves around two main datasets: 1) evidence samples and 2) causative factors. Gully occurrences across the study area signify the evidence samples. Among different gully positioning and sampling techniques, gully headcuts well represent the development activity of gullies, according to the consensus in the literature (e.g., Hosseinalizadeh et al. 2019; Kariminejad et al. 2019). Hence, 803 gully headcuts were considered and put into the modeling process. As a rule of thumb for large samples, a 70:30% partitioning balance was considered to split the gully samples into two sets of training and validation, respectively (Fig. 2). Training samples were put directly into the spatial modeling phase, while the validation samples were kept apart for validating the model's results.

For the causative factors, we strived to involve the most representative gully-controlling factors, among which anthropogenic (road density, distance to roads), hydrological (Height Above the Nearest Drainage, stream density, distance to streams, and Topographical Wetness Index), environmental (soil texture and land use), pedological (bulk density, lithology, clay, sand, and silt), and topographical (elevation, slope, aspect, general curvature, plan curvature, profile curvature, Relative Slope Position, LS factor, Topographical Position Index, and valley depth) drivers were selected. Factor selection strategy consisted of three phases: 1) using each causative category separately for the modeling phase, 2) using the best nominees from different categories, highly contributing to gain the best model performance, and 3) using all factors collectively. Consequently, the best contribution of the factors was identified and used, together with the training gully samples, to build the gully spatial model. The causative role of each controlling factor and their spatial scale are provided in detail in Table 3. Thematic maps of the causative factors are presented in Fig. 3.

Table 3

Causative role and spatial scale of gully-controlling factors
Conditioning Factors	Causative role	Scale
Anthropogenic	Human interventions such as unplanned road construction with poor foundations, impeding the hydrological cycle and drainage system, artificial accumulated water alongside the road and predisposing the soil	1:25,000
Hydrological	Soil dissolution and dispersion and retrogressive headcut development due to concentrated flow	1:25,000
Environmental	General soil susceptibility in terms of runoff generation (hydrological groups) and resistivity to gully propagation, land covers and their protective roles on the soil surface	1:100,000
Pedological	Soil erosion resistance and shear stress	1:100,000
Topographical	Elevation and proxy to different hydro-erosional mechanisms	1:25,000

Varian Inflation Factor (VIF) and its reciprocal index, Tolerance, were calculated to test the multicollinearity among the data and to avoid factors with strong correlations and the emanated bias in the model's result (Miles 2014). According to the VIF values reported in the literature (e.g., Malik et al. 2020), VIF values below 5 are acceptable (i.e., a Tolerance value above 0.2), allowing factors with the acceptable VIF range to be included in the modeling process.

Modeling phase

Different presence-absence and presence-only machine learning models have been used to reveal the hidden patterns of natural phenomena; each simultaneously offers different potentials and suffers from specific shortcomings. Generally, the main pitfall of presence-absence models is attributed to the selection of absence samples. In essence, absence location should identify areas with no evident activity or existence of the phenomenon. Antithetically, different types of gully erosion can inconspicuously develop in subterranean regions due to piping and soil dissolution, which may exhibit no evident symptoms unless the overlying strata are collapsed. Hence, selecting the absence locations may be biased by misrepresenting the actual absence locations and diluting to areas where the gully erosion is well developed but has not yet morphologically emerged. However, presence-absence models can better distinguish the presence pattern from absence, and their results tend to be more precise, leading to more compendious mitigation plans. On the other hand, presence-only models operate only based on the presence samples, although they consider some pseudo-absence locations called background points to obviate their deficiency on differentiating the presence: absence pattern. Overall, according to the phenomenal behavior of gully initiation and development mechanism, we used Maximum Entropy (MaxEnt hereafter). Many studies have been conducted on the application of MaxEnt in natural hazard modeling and attested to its capability and tremendous performance (e.g., Arabameri et al. 2020; Javidan et al. 2021).

MaxEnt is a generalized presence-only model that processes gully occurrence locations and different categorical and continuous thematic maps, treated as gully-controlling factors, irrespective of their diverse spatial scale. In contrast to deterministic models that operate on a phenomenon's prevalence, MaxEnt uses Bayes' theorem to circumvent any assumption on the prevalence value. Commencing with a random walk into the parameter space and assuming that all the pixels have the same probability of occurrence, MaxEnt further tries to find distinct patterns between the presence and pseudo-absence background locations. By doing so, through hundreds of modeling iterations, MaxEnt reaches to a probability distribution as similar as possible to the target probability distribution of the gully occurrence. During this process, all the thematic maps are rendered into their arithmetic replicas, known as features. More details on the mathematical background of MaxEnt can be found in Phillips et al. (2004, 2006, 2009), Elith et al. (2011), and Kornejady et al. (2017). Maximum entropy model was executed in MaxEnt software, considering 10^− 5 as convergence threshold, 500 modeling iterations, and 10,000 background points.

Performance assessment

Performance metrics can be employed to congruently test the goodness-of-fit (using training dataset) and prediction power/generalization capacity (using validation dataset). These metrics are calculated from a 2×2 prediction versus reality comparison table called confusion matrix. As for cutoff-independent metrics, the Receiver Operating Characteristic (ROC) curve is a widely used index for the performance assessment of spatial models. It plots the false positive ratio (so-called 1–Specificity, describing the actual non-gully areas that are incorrectly predicted as gully by the model) on the x-axis against the true positive ratio (so-called sensitivity, describing the correct prediction of gully-affected areas as observed in the study area) on the y-axis. The area under the ROC curve (AUROC) represents the degree to which the model has successfully trained and predicted, depending on the input sample dataset (training or validation). The latter ranges from 0.5 to 1, where 0.5 represents a model acting out of random chance while AUROC of 1 signifies a model with perfect performance. Plotting ROC curves were carried out in MaxEnt software in the form of jackknife plots.

Multicollinearity test

Table 4 summarizes the VIF values and their reciprocal index, Tolerance, as per each causative factor, which indicates that all values fall within the acceptable range (i.e., below 5). The latter allows using all the factors without encountering a multicollinearity-driven bias in the result. More specifically, high VIF values are mostly exhibited in morphometric indices as they store multilateral attributes of the study area that can be partially similar. More importantly, they are derived from the same source of data, the digital elevation model. The DEM is a proxy layer with the most hidden topo-hydrological and topo-erosional features. However, in our case, such an issue is obviated, and any bias derived from factors' multicollinearity is resolved due mainly to selecting the most representative and heterogeneous morphometric indices.

Table 4

Multicollinearity test among the gully-controlling factors
Conditioning Factors	Collinearity Statistics
Conditioning Factors	Tolerance	VIF
TPI	0.923	1.079
HAND	0.921	1.118
Valley depth	0.916	1.124
Lithology	0.915	1.127
Land use	0.888	1.279
RSP	0.823	1.483
Bulk density	0.813	1.492
Distance from road	0.778	1.532
Soil texture	0.754	1.611
Plan	0.745	1.721
Distance from stream	0.743	1.865
TWI	0.739	1.897
Slope	0.728	1.932
Curvature	0.565	1.954
Profile	0.554	2.112
Road density	0.458	2.293
Drainage density	0.425	2.364
Slope length	0.387	2.624
Elevation	0.346	2.715
Aspect	0.345	2.817
Silt	0.233	3.534
Clay	0.313	3.696
Sand	0.231	4.749

Gully susceptibility maps: Preliminary analysis

Figure 4 depicts the gully susceptibility maps derived from the MaxEnt model using different categories of causative factors. To summarize, Fig. 4a is built upon using anthropogenic factors, including road density and distance to roads. Although proximity and distance functions can have inherently different connotations, the limited number of anthropogenic factors selected for the modeling has generated a gully susceptibility map with a somewhat identical scheme to road distribution. This issue can diminish the integrity and the performance of the model's result unless gullies have occurred under the direct and distinct influence of road constructions across the entire study area. Figure 4b shows the gully susceptibility map derived from engaging hydrological factors, namely HAND, stream density, distance to streams, and TWI, which all signify the runoff generation mechanisms. Most notably, the HAND index is designed to differentiate three hydrological response zones (from the lowest to the highest values): waterlogged (saturation zone), slope (rapid subsurface zone), and plateau (deep percolation). High values in TWI also reflect areas with the highest contribution to runoff generation. As for this case, higher susceptibility values are centered around areas with high stream concentration and saturation zones. Figure 4c represents the gully susceptibility map generated by involving two environmental factors: soil texture and land use. As a result, the most susceptible areas are distributed in rangelands with a silty-loam soil texture. Figure 4d represents a gully susceptibility map engaged pedological factors such as bulk density, lithology, clay, sand, and silt. As a result, areas with lower bulk density and clay content and higher concentration of silt and sand exhibit higher gully susceptibility. Figure 4e depicts the gully susceptibility map that solitarily used topographical indices (elevation, slope, aspect, curvature, plan curvature, profile curvature, RSP, LS factor, TPI, and valley depth) as inputs. Accordingly, valleys and lower altitudes where streams are situated have the highest gully susceptibility. Figure 4f represents the gully susceptibility model obtained from engaging the most influential factors with the highest contribution to the model's performance. Lastly, Fig. 4g is the gully susceptibility model, employing all the factors collectively. A preliminary visual check on the produced susceptibility maps, especially Figs. 4_b,e and Figs. 4_f,g, interestingly attest to the fact that topo-hydrological factors can be the ones with the highest contribution to the model's significant performance since they share the same spatial pattern with the gully susceptibility map gained from important factors (Fig. 4f). Moreover, the collective use of all selected factors aligns with that of important factors (Fig. 5), based on which the highest susceptibility values of gully occurrence is located near the drainage density and the related topological factors with the same hydrological connotations (e.g., HAND, TWI, and valley depth). These preliminary judgments should be supported by an in-depth analysis which will be further discussed with due attention to performance metrics.

Gully susceptibility maps: in-depth analysis

As mentioned earlier in the methodology section, we used the performance metric as a data-involvement scenario-based sieve that can screen different scenarios purposively. As the most important cutoff-independent metric, the AUROC can be a valid judge of both goodness-of-fit and prediction performance of the MaxEnt model in various data-involvement alterations. These results are presented in the form of jackknife plots (i.e., ROC curves) for different scenarios (Fig. 6_a − g) and summarized in Table 5. The ROC curves and the details presented in Table 5 attest to the previous discussion on the most influential factors, where topological and hydrological categories have the highest contribution to the model's performance, with the respective AUROC of 0.84 and 0.79 in the training stage and 0.82 and 0.78 in the validation stage. As expected, engaging the most influential factors from each category and using them as the main inputs to the MaxEnt model would result in higher performance, in that the goodness-of-fit and prediction power are both increased to 0.86. Most interestingly, the highest performance of the model, particularly in the training stage, is gained by the collective use of all the factors; the AUROC in the training stage is increased to 0.89, although the AUROC in the validation stage stayed intact with an almost negligible increase. Yet a model with the highest training and prediction powers is always in favor and treated as the best-performing superior scenario (Fig. 5).

Table 5

AUROC values and performance-based ranking of the MaxEnt model considering different data-involvement scenarios
Row	Categories	AUC		Prioritizing
Row	Categories	Training	Validation	Priority Based on Training	Priority Based on Validation
1	Anthropogenic	0.691	0.721	6	5
2	Hydrological	0.793	0.782	4	4
3	Environmental	0.687	0.704	7	6
4	Pedological	0.698	0.694	5	7
5	Topographical	0.84	0.823	3	3
6	Important factors	0.859	0.863	2	2
7	All factors	0.889	0.867	1	1

Based on the gully susceptibility classes derived from the superior scenario (i.e., engaging all factors), approximately 15% of the study area falls within high and very high susceptibility classes (Table 6). Moreover, the decreasing trend of the areal extent of susceptibility classes aligns with the reality where most areas exhibit low to no susceptibility while highly susceptible areas are concentrated in small regions. Additionally, the produced susceptibility map seems to have high practicality since designating the pragmatic mitigation measures is feasible in limited areas with high susceptibility.

Table 6

Areal extent of the gully susceptibility classes derived from MaxEnt model using all factors
Categories	Class	Area (he)	Area (%)
All factors	Very low	33632.64	42.9
	Low	20326.6	25.93
	Moderate	12314.79	15.71
	High	7867.71	10.04
	Very high	4249.8	5.42

According to Table 7, some factors have hardly contributed to the modeling process, yet the question arises of how the collective use of factors can outperform the case where the most important factors are used. This reason roots directly in the concept of unique information, which refers to the fact that some factors may present unique information, and discarding them from the modeling process would result in a decreased performance. This notion is further discussed in the "Factor importance" section.

Table 7

Relative importance of the gully-controlling factors in the MaxEnt model based on jackknife test
Factor	Weight
Factor	Anthropogenic	Hydrological	Environmental	Pedological	Topographical	Important factors	All factors
TPI					0.23	0.21	0.21
HAND		0.42				0.42	0.42
Valley depth					0.46	0.47	0.47
Lithology				0.04			0.03
Land use			0.3			0.28	0.27
RSP					0.51	0.5	0.5
Bulk density				0.03			0.02
Distance from road	0.713					0.31	0.3
Soil texture			0.07				0.07
Plan					0.04		0.03
Distance from stream		0.33					0.32
TWI		0.08					0.08
Slope					0.011		0.011
Curvature					0.07		0.07
Profile					0.08		0.08
Road density	0.668						0.22
Drainage density		0.38				0.38	0.38
Slope length					0.01		0.01
Elevation					0.2	0.2	0.2
Aspect					0.06		0.08
Silt				0.13		0.12	0.12
Clay				0.1		0.1	0.1
Sand				0.075			0.08
Bolded values represent the factors with the highest contribution to model’s significant performance.

Factor Importance

Assessing the relative importance of each factor to the model's performance can better distinguish factors' contribution solitarily and complement the previously discussed category-based analysis. According to Table 7, the most influential factors with the highest contribution to the model's significant performance include valley depth (topological), HAND, drainage density, and distance to streams (hydrological). Hence, it is evident that the gully distribution across the area is predisposed by the interconnection of topo-hydrological factors. Valley depth, in particular, pinpoints the main developed gullies that can gradually cut through the streams and shape the new drainage network of the region. HAND and streams (rendered by both density and distance functions) directly address the active role of hydrological response zones and runoff generation mechanism in the retrogressive development of gully headcuts. Such an inconspicuous propagation can stem from the runoff sprawling over the gully-affected area due to man-made channel narrowing measures for agriculture purposes and floods which consequently trigger the gully headcuts. It is noteworthy that regarding different soil textural contents, silt concentration has better explained the gully distribution pattern.

Gully occurrence is the most pervasive form of water erosion across the loess soils of the northeastern part of Golestan Province, Iran. Despite the previous studies conducted in the region, a thorough investigation of the spatial signatures of gully distribution across the area is overlooked. Here we designed a data-involvement scenario-based strategy to answer which factors contribute the most to the model's performance and does it imply that such a case may guarantee building a perfect model. Regarding the latter, it appeared that topo-hydrological factors could outperform their counterparts in anthropogenic, environmental, and pedological categories. This mainly stems from the significant potential of morphometric indices in carrying the multilateral hydro-erosional attributes of the region. However, factors with even the least contribution to the modeling process hold specific and unique information that serves the model to reach higher performance, and discarding them may harm either model's goodness-of-fit or prediction power, if not both. Moreover, the tremendous application of MaxEnt, as a presence-only probabilistic model, in retrieving the unique information stored in causative factors should be acknowledged and potentially proposed for other gully-affected areas. Given that gullies are closely distributed to the drainage network and near farmlands, it is advised to set restrictions on unprincipled channel narrowing activities and assign biomechanical mitigation measures to gully headcuts to decelerate their development rate.

Arabameri A, Asadi Nalivan O, Saha S, Roy J, Pradhan B, Tiefenbacher JP, Ngo T, P. T (2020) Novel ensemble approaches of machine learning techniques in modeling the gully erosion susceptibility. Remote Sens 12(11):1890
Berhe AA, Kleber M (2013) Erosion, deposition, and the persistence of soil organic matter: mechanistic considerations and problems with terminology. Earth Surf Proc Land 38(8):908–912
Capra A, La Spada C (2015) Medium-term evolution of some ephemeral gullies in Sicily (Italy). Soil Tillage Res 154:34–43
Casalı J, López JJ, Giráldez JV (1999) Ephemeral gully erosion in southern Navarra (Spain). CATENA 36(1–2):65–84
Douglas-Mankin KR, Roy SK, Sheshukov AY, Biswas A, Gharabaghi B, Binns A, …, Daggupati P (2020) A comprehensive review of ephemeral gully erosion models. CATENA 195:104901
Elith J, Phillips SJ, Hastie T, Dudík M, Chee YE, Yates CJ (2011) A statistical explanation of MaxEnt for ecologists. Divers Distrib 17(1):43–57
Fitzjohn C, Ternan JL, Williams AG (1998) Soil moisture variability in a semi-arid gully catchment: implications for runoff and erosion control. CATENA 32(1):55–70
Frankl A, Poesen J, Scholiers N, Jacob M, Haile M, Deckers J, Nyssen J (2013) Factors controlling the morphology and volume (V)–length (L) relations of permanent gullies in the northern Ethiopian Highlands. Earth Surf Proc Land 38(14):1672–1684
Gafurov AM, Yermolayev OP (2020) Automatic gully detection: Neural networks and computer vision. Remote Sens 12(11):1743
Hessel R, van Asch T (2003) Modelling gully erosion for a small catchment on the Chinese Loess Plateau. CATENA 54(1–2):131–146
Hosseinalizadeh M, Kariminejad N, Chen W, Pourghasemi HR, Alinejad M, Behbahani AM, Tiefenbacher JP (2019) Gully headcut susceptibility modeling using functional trees, naïve Bayes tree, and random forest models. Geoderma 342:1–11
Imeson AC, Kwaad FJPM (1980) Gully types and gully prediction. Geografisch Tijdschrift 14(5):430–441
Javidan N, Kavian A, Pourghasemi HR, Conoscenti C, Jafarian Z, Rodrigo-Comino J (2021) Evaluation of multi-hazard map produced using MaxEnt machine learning technique. Sci Rep 11(1):1–20
Jiang C, Fan W, Yu N, Liu E (2021) Spatial modeling of gully head erosion on the Loess Plateau using a certainty factor and random forest model. Sci Total Environ 783:147040
Kariminejad N, Hosseinalizadeh M, Pourghasemi HR, Bernatek-Jakiel A, Campetella G, Ownegh M (2019) Evaluation of factors affecting gully headcut location using summary statistics and the maximum entropy model: Golestan Province, NE Iran. Sci Total Environ 677:281–298
Kornejady A, Ownegh M, Bahremand A (2017) Landslide susceptibility assessment using maximum entropy model with two different data sampling methods. CATENA 152:144–162
Kuhnert PM, Henderson AK, Bartley R, Herr A (2010) Incorporating uncertainty in gully erosion calculations using the random forests modelling approach. Environmetrics 21(5):493–509
Malik S, Pal SC, Chowdhuri I, Chakrabortty R, Roy P, Das B (2020) Prediction of highly flood prone areas by GIS based heuristic and statistical model in a monsoon dominated region of Bengal Basin, vol 19. Society and Environment, Remote Sensing Applications, p 100343
Miles J (2014) Tolerance and variance inflation factor. Wiley statsref: statistics reference online
Nachtergaele J, Poesen J, Wijdenes DO, Vandekerckhove L (2002) Medium-term evolution of a gully developed in a loess-derived soil. Geomorphology 46(3–4):223–239
Phillips SJ, Anderson RP, Schapire RE (2006) Maximum entropy modeling of species geographic distributions. Ecol Model 190(3–4):231–259
Phillips SJ, Dudík M, Schapire RE (2004) A maximum entropy approach to species distribution modeling. In Proceedings of the twenty-first international conference on Machine learning (p. 83)
Phillips SJ, Dudík M, Elith J, Graham CH, Lehmann A, Leathwick J, Ferrier S (2009) Sample selection bias and presence-only distribution models: implications for background and pseudo‐absence data. Ecol Appl 19(1):181–197
Poesen J (1993) Gully typology and gully control measures in the European loess belt. Farm land erosion in temperate plains environments and hills. Proceedings, 221–239
Poesen J (2011) Challenges in gully erosion research. Landf Anal 17:5–9
Poesen J, Nachtergaele J, Verstraeten G, Valentin C (2003) Gully erosion and environmental change: importance and research needs. CATENA 50(2–4):91–133
Poesen J, Vandaele K, Wesemael BV (1998) Gully erosion: importance and model implications. Modelling soil erosion by water. Springer, Berlin, Heidelberg, pp 285–311
Pourghasemi HR, Gayen A, Haque S, Bai S (2020) Gully erosion susceptibility assessment through the SVM machine learning algorithm (SVM-MLA). Gully Erosion Studies from India and Surrounding Regions. Springer, Cham, pp 415–425
Pourghasemi HR, Yousefi S, Kornejady A, Cerdà A (2017) Performance assessment of individual and ensemble data-mining techniques for gully erosion modeling. Sci Total Environ 609:764–775
Pouyan S, Pourghasemi HR, Bordbar M, Rahmanian S, Clague JJ (2021) A multi-hazard map-based flooding, gully erosion, forest fires, and earthquakes in Iran. Sci Rep 11(1):1–19
Rahmati O, Tahmasebipour N, Haghizadeh A, Pourghasemi HR, Feizizadeh B (2017) Evaluation of different machine learning models for predicting and mapping the susceptibility of gully erosion. Geomorphology 298:118–137
Saha S, Sarkar R, Thapa G, Roy J (2021) Modeling gully erosion susceptibility in Phuentsholing, Bhutan using deep learning and basic machine learning algorithms. Environ Earth Sci 80(8):1–21
Smith DD, Wischmeier WH (1957) Factors affecting sheet and rill erosion. Eos, Transactions American Geophysical Union, 38(6), 889–896
Svoray T, Michailov E, Cohen A, Rokah L, Sturm A (2012) Predicting gully initiation: comparing data mining techniques, analytical hierarchy processes and the topographic threshold. Earth Surf Proc Land 37(6):607–619
Wang F, Sahana M, Pahlevanzadeh B, Pal SC, Shit PK, Piran MJ, …, Mosavi A (2021) Applying different resampling strategies in machine learning models to predict head-cut gully erosion susceptibility. Alexandria Eng J 60(6):5813–5829

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Unraveling the Spatial Signature of Gully Erosion in the Arid and Semi-Arid Regions of the Northeast of Iran: Every Single Factor Matters!

Status:

Version 1

Abstract

Figures

Introduction

Materials And Methods

Study area

Data construction phase

Modeling phase

Performance assessment

Result And Discussion

Multicollinearity test

Gully susceptibility maps: Preliminary analysis

Gully susceptibility maps: in-depth analysis

Factor Importance

Conclusion

References

Additional Declarations

Status:

Version 1