Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

An applied statistical method to identify desertification indicators in northeastern Iran

An applied statistical method to identify desertification indicators in northeastern Iran Background: Desertification could be considered ultimate consequence of land degradation in an ecosystem. Iran with more than 75% arid and semi-arid areas involves fragile and susceptible ecosystems to desertification. We applied a statistical algorithm including regression trees and random forest techniques for determining main factors affecting desertification based on ESAs in Taybad-Bakharz region at northeastern Iran. Results: The results indicated a significant correlation between the desertification hazard value with variables of wind erosion, precipitation, aridity index, technology development, slope index, vegetation state and land use changes. Conclusions: Regression trees and random forest techniques in desertification hazard provide an absolute estimation of the relationship between dependent and independent variables. We can use a robust base for further investigations and refined with findings from in-depth studies carried out at the local scale. Keywords: Desertification hazard, Regression trees, Random forest, Data mining Background and measurements have suggested assessing desertifica- In recent decades, the challenges of environment regard- tion hazard in different regions of the world (Sepehr et ing drying lake, groundwater depletion, land fissures, al. 2007). In relation to desertification risk mapping, as- drought, migration, poverty etc. are highlighted in heart sessment, and forecasting, many studies can be found of Middle East, Iran, with more than 3000 years that mainly are based on empirical and regional models. civilization. All of mentioned threatens indicated that European project of ESAs (Environmental Sensitive civilization is collapsing due to ecosystem degradation Areas) team that called MEDALUS (Mediterranean particularly in arid areas. These consequences can be Desertification and Land Use) is one of the regional briefed in a word “desertification”. frameworks to detect sensitive areas regarding desertifi- If no remedial action is taken, desertification rate will cation risk in Mediterranean region (Kosmas et al. be increasing significantly and threaten sustainable liveli- 1999). Ladisa et al. (2012), applied this method for asses- hoods at least for people of arid and semi-arid regions of sing desertification risk in Apulia region, southeastern Iran, areas with more than 75% expansion in Iran. Italy and reported that this method shows efficient out- Detecting, distinguishing and mitigation of the outcomes put regarding desertification vulnerability and detecting of desertification and finding main parameters affecting sensitive environments. Leman et al. 2016 used the GIS- desertification rate in an area is most effective step in ac- based integrated evaluation model base on ESAs on two tion plans of combating desertification (UNCCD, 1994). assessment sets in Langkawi, Malaysia. “Set A included Many methods such as mathematical models, para- indicators chosen from Malaysian integrated ESA tool metric equations, remote sensing, direct observation, and set B involved indicators derived from five eco- environments in China. The results showed Set A in * Correspondence: adelsepehr@um.ac.ir order to reveal environmental sensitivity is more appro- Department of Desert and Arid Zones Management, Ferdowsi University of priate and more efficient than Set B”. There is miscel- Mashhad, Mashhad, Iran laneous reports regarding application of the ESAs Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 2 of 10 method to detect sensitive areas to desertification, such participating in desertification risk were identified by as Wijitkosum (2016), De Pina Tavares et al. (2015), and regression tree and random forest techniques. Sepehr et al. (2007). The common point of all these arti- cles is the high efficiency of ESAs method for recognizing Methods desertification-prone areas. Heretofore, many studies have Study area been done regarding desertification risk assessment and fore- This study was applied in Taybad-Bakharz located in an casting, and in the majority of those emphasized on regional arid and semi-arid environment of Khorasan Razavi prov- and empirical indicators and methods (Martínez-Valderrama ince, northeastern Iran with an area of 4800 km (Fig. 1). et al. 2016; Ferrara et al. 2012; Karamesouti et al. 2015; The livelihood of rural communities of the studied region Patriche and Bandoc 2017; Patriche et al. 2017; Salvati et al. is depends on livestock and farming. Precipitation varies 2016; Zambon et al. 2017). between 100 and 250 mm based on topographic condi- Data mining knowledge, as a logical process for find- tions with a wide temporal and spatial distribution. Mean ing useful data through large amount of data, and par- annual temperature is about 16 °C, during the day in sum- ticularly regression technique can be used for mer, temperature goes up 42 °C, and actual evapotranspir- modeling the relationship between one or more inde- ation changes between 800 and 2100 mm per year. The pendent variables and dependent variables (Ramageri wind velocity ranges from 5.3 to 6.8 m/s, with the max- 2010). For example, the regression tree and random imum occurring in May approximately 6 m/s. Moreover, forest techniques as two relatively new tree-based over 120 days per year have wind. Additional file 1 shows models optimized predictive performance by combin- the digital elevation model (DEM) of the studied area. ing a large number of simple trees into a powerful model rather than a single tree model based on trad- itional regression trees (Skurichina and Duin 2002). Calculating desertification hazard index Yang et al. (2016) used regression tree and random Interaction between main driving force factors and forest model to map topsoil organic carbon concentra- spatial-temporal changes of main factors have been tion in an alpine ecosystem. Their results showed that led to desertification in the studied area. To select the two methods can be used as strong and effective main criteria expert’sopinionsweretaken by Delphi modeling approaches in the mapping of soil organic decision-making framework. We applied the ESAs matter concentration. This article aims to investigate method to detect desertification-prone areas based on correlation between desertification indicators and modified main criteria and indicators. Choosing and choosing main indicators affecting desertification by providing the indicators were according to the avail- regression statistical methods based on ESAs frame- able information and maximum influence on desertifi- work in northeastern Iran. The most important indices cation process in the region. To map layers ArcGIS; Fig. 1 Position map of study area in northeastern Iran Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 3 of 10 version 10.2 and R software; packages of raster, sp., Table 2 Quantitative and qualitative classes of criteria and desertification hazard rgdal, maptools, and lulcc were used. According to the ESAs, the quality of each criteria was class Interval score Sensitive degree to Desertification desertification Hazard provided by geometric mean of considered effective indi- 10–1.5 Low Low cators on criteria quality. Eq. 1 shows the relation used to calculate criteria quality index by geometric mean of 2 1.6–2.5 Medium Moderate indicators. 3 2.6–3.5 High High 4 3.6–4 Very high Severe X pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi QI ¼ ¼ X  X  …; X ð1Þ 1 2 n i¼1 A land unit map (LUM) was considered for investiga- Where QI is criteria quality index and X showing indi- tion of quality of each criterion and desertification status cators affecting the quality of each criterion. based on geomorphological facies. For providing LUM, Nine (9) main criteria affecting desertification suscep- imagery data of LANDSAT (TM and ETM ) were used tibility were chosen with indicators regarding the quality (Landsat satellites). The reason for choosing geomorph- of each criterion. The criteria and indicators illustrated ology facies for LUM and calculating the status of each in Table 1. The quality layers were classified in four criterion and desertification risk in each of study unit is quantitative and qualitative classes as shown in Table 2. the relation to the slow changes of surface morphology An interval weighing was considered between 0 and 4 during short time vs rapid changes of land-cover and for the quantitative value of qualitative status of each land-use due to human activities particularly in recent criterion. decade in the studied region. After providing the quality layer for each criterion in- volving nine quality layers, desertification hazard index Regression trees was calculated by geometric mean of quality layers as To identify the most effective criteria and indicators on shown eq. 2. desertification hazard, we applied CART (classification and regression tree) analysis. Classification and regression tree 1=9 DH ¼ðÞ CQI SQI GQI AQI VQI S−EQI EQI TQI GWQI model is a nonparametric method introduced by Breiman ð2Þ et al. (1984). This method is able to predict the quantitative variables (regression tree) and classification variables (classi- where DH is desertification hazard value, and CQI, SQI, fication tree) based on a set of qualitative and quantitative GQI, AQI, VQI, S-EQI, EQI, TQI, GWQI respectively variables (Yeh 1991). A classification or regression tree refer to quality of climate, soil, geology, agriculture, model has been formed from the several branches and vegetation, socio-economic, erosion, technology develop- some nodes. The first node that includes all the sam- ment, and ground water. ples is called the parent node. Other nodes are called child nodes. Then, based on one of the predictor var- iables, two branches take place and this situation con- Table 1 Main criteria affecting desertification and considered tinues to the end node (Frisman 2008). Another indicators for quality degree of criteria parameter is pruning the tree structure and selecting Criterion Indicators representing quality the appropriate size of the tree. The CART analysis was climate precipitation, aridity index done using R software (apart and rpart.plot packages). geology geology, slope index, land use change Random Forest soil electrical conductivity (EC), sodium adsorption ratio (SAR), soil depth Random forest is a non-parametric method and belongs to the collection methods that were obtained from ma- vegetation vegetation state, vegetation utilization, vegetation restoration chine learning methods in the late nineteenth century (Catani et al, 2013; Pourghasemi and Kerle, 2016). This agriculture cropping pattern, crop yield, use of machinery algorithm is a set of classification and regression trees erosion wind erosion, water erosion developed by Breiman (2001). Breiman proposed ran- dom forests, which add an additional layer of randomness groundwater EC, SAR, water declination to bagging (Liaw and Wiener 2002). To perform this pro- social-economic poverty status cedure, several parameters must be determined. The first technology development technology development parameter is the number of predicting trees. In this study, Aridity index was calculated by ratio of annual rainfall to potential of 500 trees were created. The second parameter is the num- evapotranspiration proposed by FAO/UNEP for global desertification map AI=P/ETP (UNEP 1992) ber of the predictor variable with no need to prune trees Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 4 of 10 in classification. The proximity matrix was used for identi- area regarding vegetation, where it is 63% for agriculture fying structure in the data (Breiman 2002). The Random indicators. In terms of the technology development, 80% Forest package provides an R interface to the Fortran pro- of the study area gained moderate class of desertification grams by Breiman and Cutler (available at https:// hazard. Erosion criterion including wind and water ero- www.stat.berkeley.edu/~breiman/). sion indicated that 57% of study area are classified in After calculating desertification hazard index, regression high-risk desertification class for wind erosion and 63% trees and random forest technique was applied for identify- areas shows high level for water erosion. Moderate ing preference of main criteria affecting desertification. The conditions classified regarding groundwater and soil sample population involved 25 land units provided based criteria. Therefore, based on impact degree of indicators on geomorphology facies with 21 variables including indica- regarding desertification susceptibility, more than 20% of tors considered for criteria quality as independent variables Taybad-Bakharz region shows susceptible and prone area and desertification hazard as dependent variable. The flow- to desertification. Ultimately, a desertification suscepti- chart of methodology applied in this study is shown in Fig. 2. bility classified in two sensitivity levels involving moder- ate degree with 37% of the area, and high degree with Results and discussion 63% of the areas. The classification categories of the Desertification hazard areas indices and desertification hazard in the form of the The desertification status was calculated on each of sep- map are shown in Fig. 4. It should be noted that all of arated land units. The Kavir areas (salty-clay pans and layers and maps were classified into four classes based Sabkha) are smallest zones and low-level pediment fan on Table 3. The output maps have been presented based and valley terrace deposits cover largest areas in the on gained classes, for example for the geology and cli- studies region as shown in Table 3 and Fig. 3. Further- mate criteria, the desertification does not show severe more, most geomorphology facies involve badlands and class, so this class was ignored in the legend while for pediment deposits surfaces and concentrated on the east erosion criterion, the legend includes all of the classes as and northwest region, where land-use is mainly agricul- it gained all of the desertification risk classes. ture and cultivated areas. The geology map of the stud- ied region shown in the Additional file 2. Regression tree outcomes Results indicated that regarding climate criterion the The results of CART for the 21 targets outcome variables- studied area classified in high vulnerable to desertifica- desertification hazard are presented in Figs. 5, 6,and 7, tion hazard, in othe word shows a high-risk class of respectively. Figures 6 and 7 includes the predictor vari- desertification as 65% of the studied area are susceptible ables and the value that split each subgroup. Within each in relation to climate factors. The results showed that in- node, the mean score or proportion of participants in each dicators considered for geology criterion are main fac- response category are presented. Figure 5 shows the com- tors affecting desertification, so that about 88% areas are plexity parameter (cp) option of the summary function prone-area to desertification. In addition, the criteria of that instructs it to prune the printout, but it does not vegetation and agriculture show undesirable conditions prune the tree. The cp shows minimum error so that the as well as 85% areas considered desertification-prone regression tree is pruned by cp wherever the error shows Fig. 2 General outline method used in the study Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 5 of 10 Table 3 Frequency and distribution of geomorphological facies (lands unit) in Taybad-Bakharz area Land Geomorphological facies Area (ha) % Frequency unit 1 Rocky outcrop (Sandstone and 1387 0.3 1 sandy limestone) 2 Rocky outcrop (Conglomerate 1050 0.22 1 and sandstone) 3 Rocky outcrop (Pale red marl 16,057 3.0 1 and gypsiferous marl) 4 Rocky outcrop (Andesitic to 1165 0.25 1 basaltic volcanic tuff) 5 Rocky outcrop (Andesitic 3238 0.7 1 volcanosediment) 6 Rocky outcrop (Andesitic 15,710 3.3 1 volcanic tuff) 7 Rocky outcrop (Dacitic to 2638 0.55 1 Andesitic tuff) 8 Rocky outcrop (Rhyolitic to 3295 0.7 1 Fig. 3 lands unit map (LUM) in the studied area. The land unit map rhyolitic tuff) provided based on geomorphological characteristics including 25 9 Rocky outcrop (Rhyolitic to 9273 2.0 2 geomorphic facies rhyolitic volcanic) 10 Rocky outcrop (Granite) 4864 1.0 2 geomorphology map (geology and morphology maps). 11 Rocky outcrop (Red marl, 5363 1.2 1 gypsiferous marl) All of indicators were evaluated in each land units (poly- gone) separetely. 12 Hills with gully erosion 10,000 2 4 Given to the Fig. 6, the first split for desertification 13 Badland 41,000 8.7 9 hazard status is based on the score of the wind erosion 14 Rocky outcrop (debris) 12,900 2.7 3 index. The subgroup that represented higher hazard tak- 15 Eroded Inselberg 9980 2.0 3 ing scores (< 2.26, > = 2.26) and was further split based 16 Hills 10,000 2.0 4 on water erosion and then land use change status on the 17 Rocky outcrop (Gneiss, granite, 7600 1.6 2 left (This subgroup does not split again). Also, crop pat- amphibolite) tern status and the next split that was related to the 18 Clay flat 21,700 4.5 3 aridity index and vegetation utilization indices (on the right). Vegetation utilization and land-use change cri- 19 High level pediment fan and 52,950 11.2 12 valley terrace deposits teria showed relatively lower correlation with desertifica- (semi-Cultivated) tion hazard value. 20 Low-level pediment fan and 222,900 47 9 In Fig. 7, Parent group shows the desertification risk valley terrace deposits with mean 1.92, the children group presents the erosion (Cultivated) status including water and wind erosion. The water ero- 21 Sand dune (semi-active) 7220 1.5 1 sion is a sub-tree or a branch. The highest correlation 22 Irregular aspect 4325 1 1 was obtained between wind erosion and desertification. 23 Inselberg 4847 1 1 The water erosion only involves one branch and its 24 Low-level piedmont fan and 5480 1.2 1 division does not continue in regression tree. The crop playa pattern shows highest correlation with wind erosion with 25 Kavir (Salty-clay pan, Sabkha) 230 .1 1 continuous division and after that, aridity index indicates correlation with crop pattern. Therefore, desertification risk is affective by wind erosion with a powerful correl- minimum amount, the regression tree is pruning. The ation with crop patterns and aridity index. In addition, pruning based on cp is a validation test in regression tree, as water erosion shows a weakness correlation, we can indeed before regression; a training was used by 30% of have a judgement that wind erosion has most powerful total data. For more information the raw data used in stat- correlation with desertification in all of land units. istical methods has been presented in Additional file 3. The regression tree used when we have numerical var- In the Fig. 6, n shows the land units and means the iables. This method tries to recognize highest correlation amount of the polygons which was determined based on between variables by using Gini coefficient, Chi-square, Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 6 of 10 Climate Geology Soil Vegetation Cover Erosion Low Low Low Low Low Moderate Moderate Moderate Moderate Moderate High High Severe Severe High High High 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 Agriculture Technology Development Soci−Economic Groundwater Desertification Susceptibility Low Low Low Low Low High Moderate Moderate Moderate Moderate High High 0 20000 meters 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 Fig. 4 desertification susceptibility areas for main indicators. Desertification susceptibility map provided based on preference degree of each indicators MSE, and entropy. The Gini coefficient find homogeny impact on desertification hazard, respectively. These in- variable and regression tree will be found correlations dicators are highly associated with the value of desertifi- based on mean of variables. The regression presents a cation hazard. Wind erosion and cropping pattern classification by Gini coefficient by number of successes indices in 18 land units are identified as the most im- and failures. portant factor affecting desertification hazard. Also, arid- The cross-validation shows only a small difference in ity index is listed as the most important factors of desertification hazard status. Given to the regression desertification in 13 land units. Although, water erosion trees outcomes, among 21 independent variables, indica- index have less impact on the hazard of desertification tors such as wind erosion, precipitation, soil EC, soil tex- (7 land units). Finally, wind erosion, cropping pattern, ture, groundwater (SAR) and slope had the greatest aridity index and precipitation are identified as the most Fig. 5 cp Plot. A validation test for regression tree method based on minimum error 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 7 of 10 Fig. 6 regression Tree with Package rpart important indicators that affected desertification hazard development, aridity index, slope index, soil EC, land- in this region. use changes, vegetation state, precipitation, geology, water declination, and soil texture. Moreover, based on variable importance for the RF-model, we observed that Random Forest outcome wind erosion, technology development, aridity index, The results from the variable selection random forest slope index, vegetation state and land-use change variables are shown in Figs. 8 and 9. Independent variables in- are relatively most important on desertification hazard of volving 21 indicators ordered based on mean decrease Taybad-Bakharz region, respectively. The important per- accuracy. The accuracy measure determined main cent values of these variables measured 10.48%, 7.5%, effective criteria including wind erosion, technology 6.8%, 5.9% 5.3%, and 5.2% respectively. Fig. 7 selected regression tree Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 8 of 10 Fig. 8 plot showing the decrease of (in terms of out-of-bag classification errors, OOBE) OOBE with increasing number of trees T# in the RF structure. A working value of T# = 300 was chosen for the RFtb model structure used in the tests and experiments As shown in Fig. 8, in forest random method, at the and migration, and ultimately ecosystem collapse in Iran first will be created classification trees that called tree of with more than 75% drylands require a carefully under- vote. The forest will be decided and classified base on standing of the desertification process and recognizing the most vote, then average of tree’s outcome presents driving forces. This study examines the performance of a the regression. In Fig. 9, the preference of each variable statistical method to identify the most important criteria was calculated by mean square error (MSE) and purity affecting desertification process and risk. Studied region or entropy degree. The less MSE indicates an indicator showed desertification-prone areas with 63% high-risk with more preference. level of desertification. Application of regression trees and random forest techniques identified the most important criteria affecting desertification and recog- Conclusions nized that indicators such as wind erosion, technology de- Land degradation and desertification consequences such velopment, aridity index, slope index, precipitation, as dust storms, drying of lakes, water scarcity, poverty Fig. 9 variable importance calculated by mean square error (MSE) and purity or entropy degree Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 9 of 10 vegetation state and land use change are major indicators Received: 27 May 2017 Accepted: 23 February 2018 affecting the quality of criteria and desertification in the Taybad-Bakharz. The results of this research indicated References that erosion factors including water and wind erosion are Breiman, L., J. Friedman, C.J. Stone, and R. Olshen. 1984. Classification and the most important desertification factors in the studied regression trees. CRC press, Taylor and Francis Group. pp. 246–280. area. Over-grazing and vegetation degradation in the re- Breiman, Leo. 2001. Manual on Random Forests. University of California. p. 33. https://www.stat.berkeley.edu/~breiman/randomforest2001.pdf. gion particularly in recent decade led to degraded soils Breiman, Leo. 2002. Manual on setting up, using, and understanding random and decreasing fertility, so the erosion rate is raising in the forests. Vol. v3.1 https://www.stat.berkeley.edu/~breiman/. study area. Data mining process and particularly regres- Catani, F., D. Lagomarsino, S. Segoni, and V. Tofani. 2013. Landslide susceptibility estimation by random forests technique: Sensitivity and scaling issues. sion trees and random forest technique in desertification Natural Hazards and Earth System Sciences 13 (11): 2815–2831. hazard can be recommended as a robust base for further De Pina Tavares, J., I. Baptista, A.J.D. Ferreira, P. Amiotte-Suchet, C. Coelho, S. investigation of desertified lands for best management of Gomes, R. Amoros, E.A. Dos Reis, A.F. Mendes, L. Costa, J. Bentub, and L. Varela. 2015. Assessment and mapping the sensitive areas to desertification processes in these areas. in an insular Sahelian mountain region case study of the Ribeira Seca watershed, Santiago Island, Cabo Verde. Catena 128: 214–223. https://doi. org/10.1016/j.catena.2014.10.005. Endnotes 1 Ferrara, A., L. Salvati, A. Sateriano, and A. Nolè. 2012. Performance evaluation and - Tool in this research used for indicators system cost assessment of a key indicator system to monitor desertification vulnerability. Ecological Indicators 23: 123–129 https://doi.org/10.1016/j. ecolind.2012.03.015. Additional files Frisman L. (2008) App lying classification and regression tree analysis to identify Priso Ners with high lllV risk Behaviorst. 40 (December). Additional file 1: Digital Elevation Model (DEM) of studied area. (JPEG Karamesouti, M., V. Detsis, A. Kounalaki, P. Vasiliou, L. Salvati, and C. Kosmas. 2015. 125 kb) Catena land-use and land degradation processes affecting soil resources : Evidence from a traditional Mediterranean cropland (Greece). Catena 132: Additional file 2: Geology map of Taybad-Bakharz in northeastern Iran. 45–55. https://doi.org/10.1016/j.catena.2015.04.010. (JPEG 134 kb) Kosmas, C., M. Kirkby, and N. Geeson. 1999. The MEDALUS project Mediterranean Additional file 3: The raw data used in statistical algorithms. (CSV 3 kb) desertification and land use; manual on key indicators of desertification and mapping environmentally sensitive areas to desertification. Brussels: European Commission. Abbreviations Ladisa, G., M. Todorovic, and L. Trisorio. 2012. A GIS-based approach for CART: Classification And Regression Tree; cp: Complexity Parameter; desertification risk assessment in Apulia region, SE Italy. Physics and Chemistry ESAs: Environmental Sensitive Areas; LUM: Land Unit Map; of the Earth 49: 103–113. https://doi.org/10.1016/j.pce.2011.05.007. MEDALUS: Mediterranean Desertification and Land Use; MSE: Mean Square Leman, N., M.F. Ramli, and R.P. Khairani Khirotdin. 2016. GIS-based integrated Error; OOBE: Out-Of-Bag classification Errors evaluation of environmentally sensitive areas (ESAs) for land use planning in Langkawi, Malaysia. Ecological Indicators 61: 293–308. Acknowledgements https://doi.org/10.1016/j.ecolind.2015.09.029. The authors are thanks from administration of Natural Resources and Liaw, A., and M. Wiener. 2002. Classification and regression by randomForest. R Environment College, Ferdowsi University of Mashhad for supporting and news 2 (December): 18–22. providing the facilities, particularly Dr. Naseri and Dr. Mosaedi who are head Martínez-Valderrama, J., J. Ibáñez, G. Del Barrio, M.E. Sanjuán, F.J. Alcalá, S. of college and vice-head separately. Martínez-Vicente, A. Ruiz, and J. Puigdefábregas. 2016. Present and future of desertification in Spain: Implementation of a surveillance system to prevent Funding land degradation. Sci Total Environ. 563–564: 169–178. https://doi.org/10. The Gorgan University of Agricultural Sciences and Natural Resources and 1016/j.scitotenv.2016.04.065. Iran science and technology ministry provided the funding of this research. Patriche, C., and G. Bandoc. 2017. Quantification of land degradation sensitivity areas in southern and central southeastern Europe. New Availability of data and materials results based on improving DISMED methodology with new climate Not applicable. data. Catena 158: 309–320. https://doi.org/10.1016/j.catena.2017.07.006. Patriche, C., M. Dumitra, and G. Bandoc. 2017. Catena spatial assessment of land Authors’ contributions degradation sensitive areas in southwestern Romania using modi fi ed MS conducted the fieldwork and contributed to the analysis of the data as MEDALUS method. 153: 114–130. https://doi.org/10.1016/j.catena.2017.02.011. well as in writing the first draft. MO is a supervisor for the research, and AN Pourghasemi, H.R., and N. Kerle. 2016. Random forests and evidential belief function- is an adviser. AS supervised data analysis and contributed in revising the based landslide susceptibility assessment in western Mazandaran Province , Iran. manuscript and providing the final draft. He is corresponding the research. Environmental Earth Sciences. https://doi.org/10.1007/s12665-015-4950-1. All authors read and approved the final manuscript. Ramageri, M. 2010. Data mining techniques and applications. Indian J Comput Sci Eng 1 (4): 301–305. Salvati, L, Kosmas C, Kairis O, Karavitis C, Acikalin S, Belgacem A, Chaker M, Competing interests Fassouli V, Gokceoglu C, Gungor H, Hessel R, Sol A, Khatteli H, Kounalaki A, The authors declare that they have no competing interests. Laouina A, Ocakoglu F, Ouessar M, Ritsema C, Colantoni A, Carlucci M (2016) Assessing the effectiveness of sustainable land management policies for Publisher’sNote combating deserti fi cation : A data mining approach 183, 754–762. https:// Springer Nature remains neutral with regard to jurisdictional claims in doi.org/10.1016/j.jenvman.2016.09.017. published maps and institutional affiliations. Sepehr, A., A.M. Hassanli, M.R. Ekhtesasi, and J.B. Jamali. 2007. Quantitative assessment of desertification in south of Iran using MEDALUS method. Author details Environmental Monitoring and Assessment 134 (1–3): 243–254. Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, Skurichina, M., and R.P.W. Duin. 2002. Bagging, boosting and the random Golestan, Iran. Department of Desert and Arid Zones Management, subspace method for linear classifiers. Pattern Analysis and Applications Ferdowsi University of Mashhad, Mashhad, Iran. 5(2): 121–135. Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 10 of 10 UNCCD. 1994. Elaboration of an international convention to combat desertification in countries experiencing serious drought and/or desertification, particularly in Africa,1–58 (June). http://www2.unccd.int/. UNEP. 1992. World atlas of desertification. London: Edward Arnold. Wijitkosum, S. 2016. The impact of land use and spatial changes on desertification risk in degraded areas in Thailand. Sustainable Environ Res 26 (2): 84–92. https://doi.org/10.1016/j.serj.2015.11.004. Yang, R.M., G. Zhang, F. Liu, Y. Lu, F. Yang, F. Yang, M. Yang, Y.G. Zhao, and D.C. Li. 2016. Comparison of boosted regression tree and random forest models for mapping topsoil organic carbon concentration in an alpine ecosystem. Ecological Indicators 60: 870–878. https://doi.org/10.1016/j.ecolind.2015.08.036. Yeh, Chyon-Hwa. 1991. Classification and regression trees (CART). Chemometrics and Intelligent Laboratory Systems 12 (1): 95–96. https://doi.org/10.1016/0169- 7439(91)80113-5. Zambon, I., A. Colantoni, M. Carlucci, N. Morrow, A. Sateriano, and L. Salvati. 2017. Land quality , sustainable development and environmental degradation in agricultural districts : A computational approach based on entropy indexes. Environmental Impact Assessment Review 64: 37–46. https://doi.org/10.1016/j. eiar.2017.01.003. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Geoenvironmental Disasters Springer Journals

An applied statistical method to identify desertification indicators in northeastern Iran

Loading next page...
 
/lp/springer-journals/an-applied-statistical-method-to-identify-desertification-indicators-XhLdJidIpq
Publisher
Springer Journals
Copyright
2018 The Author(s).
eISSN
2197-8670
DOI
10.1186/s40677-018-0095-3
Publisher site
See Article on Publisher Site

Abstract

Background: Desertification could be considered ultimate consequence of land degradation in an ecosystem. Iran with more than 75% arid and semi-arid areas involves fragile and susceptible ecosystems to desertification. We applied a statistical algorithm including regression trees and random forest techniques for determining main factors affecting desertification based on ESAs in Taybad-Bakharz region at northeastern Iran. Results: The results indicated a significant correlation between the desertification hazard value with variables of wind erosion, precipitation, aridity index, technology development, slope index, vegetation state and land use changes. Conclusions: Regression trees and random forest techniques in desertification hazard provide an absolute estimation of the relationship between dependent and independent variables. We can use a robust base for further investigations and refined with findings from in-depth studies carried out at the local scale. Keywords: Desertification hazard, Regression trees, Random forest, Data mining Background and measurements have suggested assessing desertifica- In recent decades, the challenges of environment regard- tion hazard in different regions of the world (Sepehr et ing drying lake, groundwater depletion, land fissures, al. 2007). In relation to desertification risk mapping, as- drought, migration, poverty etc. are highlighted in heart sessment, and forecasting, many studies can be found of Middle East, Iran, with more than 3000 years that mainly are based on empirical and regional models. civilization. All of mentioned threatens indicated that European project of ESAs (Environmental Sensitive civilization is collapsing due to ecosystem degradation Areas) team that called MEDALUS (Mediterranean particularly in arid areas. These consequences can be Desertification and Land Use) is one of the regional briefed in a word “desertification”. frameworks to detect sensitive areas regarding desertifi- If no remedial action is taken, desertification rate will cation risk in Mediterranean region (Kosmas et al. be increasing significantly and threaten sustainable liveli- 1999). Ladisa et al. (2012), applied this method for asses- hoods at least for people of arid and semi-arid regions of sing desertification risk in Apulia region, southeastern Iran, areas with more than 75% expansion in Iran. Italy and reported that this method shows efficient out- Detecting, distinguishing and mitigation of the outcomes put regarding desertification vulnerability and detecting of desertification and finding main parameters affecting sensitive environments. Leman et al. 2016 used the GIS- desertification rate in an area is most effective step in ac- based integrated evaluation model base on ESAs on two tion plans of combating desertification (UNCCD, 1994). assessment sets in Langkawi, Malaysia. “Set A included Many methods such as mathematical models, para- indicators chosen from Malaysian integrated ESA tool metric equations, remote sensing, direct observation, and set B involved indicators derived from five eco- environments in China. The results showed Set A in * Correspondence: adelsepehr@um.ac.ir order to reveal environmental sensitivity is more appro- Department of Desert and Arid Zones Management, Ferdowsi University of priate and more efficient than Set B”. There is miscel- Mashhad, Mashhad, Iran laneous reports regarding application of the ESAs Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 2 of 10 method to detect sensitive areas to desertification, such participating in desertification risk were identified by as Wijitkosum (2016), De Pina Tavares et al. (2015), and regression tree and random forest techniques. Sepehr et al. (2007). The common point of all these arti- cles is the high efficiency of ESAs method for recognizing Methods desertification-prone areas. Heretofore, many studies have Study area been done regarding desertification risk assessment and fore- This study was applied in Taybad-Bakharz located in an casting, and in the majority of those emphasized on regional arid and semi-arid environment of Khorasan Razavi prov- and empirical indicators and methods (Martínez-Valderrama ince, northeastern Iran with an area of 4800 km (Fig. 1). et al. 2016; Ferrara et al. 2012; Karamesouti et al. 2015; The livelihood of rural communities of the studied region Patriche and Bandoc 2017; Patriche et al. 2017; Salvati et al. is depends on livestock and farming. Precipitation varies 2016; Zambon et al. 2017). between 100 and 250 mm based on topographic condi- Data mining knowledge, as a logical process for find- tions with a wide temporal and spatial distribution. Mean ing useful data through large amount of data, and par- annual temperature is about 16 °C, during the day in sum- ticularly regression technique can be used for mer, temperature goes up 42 °C, and actual evapotranspir- modeling the relationship between one or more inde- ation changes between 800 and 2100 mm per year. The pendent variables and dependent variables (Ramageri wind velocity ranges from 5.3 to 6.8 m/s, with the max- 2010). For example, the regression tree and random imum occurring in May approximately 6 m/s. Moreover, forest techniques as two relatively new tree-based over 120 days per year have wind. Additional file 1 shows models optimized predictive performance by combin- the digital elevation model (DEM) of the studied area. ing a large number of simple trees into a powerful model rather than a single tree model based on trad- itional regression trees (Skurichina and Duin 2002). Calculating desertification hazard index Yang et al. (2016) used regression tree and random Interaction between main driving force factors and forest model to map topsoil organic carbon concentra- spatial-temporal changes of main factors have been tion in an alpine ecosystem. Their results showed that led to desertification in the studied area. To select the two methods can be used as strong and effective main criteria expert’sopinionsweretaken by Delphi modeling approaches in the mapping of soil organic decision-making framework. We applied the ESAs matter concentration. This article aims to investigate method to detect desertification-prone areas based on correlation between desertification indicators and modified main criteria and indicators. Choosing and choosing main indicators affecting desertification by providing the indicators were according to the avail- regression statistical methods based on ESAs frame- able information and maximum influence on desertifi- work in northeastern Iran. The most important indices cation process in the region. To map layers ArcGIS; Fig. 1 Position map of study area in northeastern Iran Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 3 of 10 version 10.2 and R software; packages of raster, sp., Table 2 Quantitative and qualitative classes of criteria and desertification hazard rgdal, maptools, and lulcc were used. According to the ESAs, the quality of each criteria was class Interval score Sensitive degree to Desertification desertification Hazard provided by geometric mean of considered effective indi- 10–1.5 Low Low cators on criteria quality. Eq. 1 shows the relation used to calculate criteria quality index by geometric mean of 2 1.6–2.5 Medium Moderate indicators. 3 2.6–3.5 High High 4 3.6–4 Very high Severe X pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi QI ¼ ¼ X  X  …; X ð1Þ 1 2 n i¼1 A land unit map (LUM) was considered for investiga- Where QI is criteria quality index and X showing indi- tion of quality of each criterion and desertification status cators affecting the quality of each criterion. based on geomorphological facies. For providing LUM, Nine (9) main criteria affecting desertification suscep- imagery data of LANDSAT (TM and ETM ) were used tibility were chosen with indicators regarding the quality (Landsat satellites). The reason for choosing geomorph- of each criterion. The criteria and indicators illustrated ology facies for LUM and calculating the status of each in Table 1. The quality layers were classified in four criterion and desertification risk in each of study unit is quantitative and qualitative classes as shown in Table 2. the relation to the slow changes of surface morphology An interval weighing was considered between 0 and 4 during short time vs rapid changes of land-cover and for the quantitative value of qualitative status of each land-use due to human activities particularly in recent criterion. decade in the studied region. After providing the quality layer for each criterion in- volving nine quality layers, desertification hazard index Regression trees was calculated by geometric mean of quality layers as To identify the most effective criteria and indicators on shown eq. 2. desertification hazard, we applied CART (classification and regression tree) analysis. Classification and regression tree 1=9 DH ¼ðÞ CQI SQI GQI AQI VQI S−EQI EQI TQI GWQI model is a nonparametric method introduced by Breiman ð2Þ et al. (1984). This method is able to predict the quantitative variables (regression tree) and classification variables (classi- where DH is desertification hazard value, and CQI, SQI, fication tree) based on a set of qualitative and quantitative GQI, AQI, VQI, S-EQI, EQI, TQI, GWQI respectively variables (Yeh 1991). A classification or regression tree refer to quality of climate, soil, geology, agriculture, model has been formed from the several branches and vegetation, socio-economic, erosion, technology develop- some nodes. The first node that includes all the sam- ment, and ground water. ples is called the parent node. Other nodes are called child nodes. Then, based on one of the predictor var- iables, two branches take place and this situation con- Table 1 Main criteria affecting desertification and considered tinues to the end node (Frisman 2008). Another indicators for quality degree of criteria parameter is pruning the tree structure and selecting Criterion Indicators representing quality the appropriate size of the tree. The CART analysis was climate precipitation, aridity index done using R software (apart and rpart.plot packages). geology geology, slope index, land use change Random Forest soil electrical conductivity (EC), sodium adsorption ratio (SAR), soil depth Random forest is a non-parametric method and belongs to the collection methods that were obtained from ma- vegetation vegetation state, vegetation utilization, vegetation restoration chine learning methods in the late nineteenth century (Catani et al, 2013; Pourghasemi and Kerle, 2016). This agriculture cropping pattern, crop yield, use of machinery algorithm is a set of classification and regression trees erosion wind erosion, water erosion developed by Breiman (2001). Breiman proposed ran- dom forests, which add an additional layer of randomness groundwater EC, SAR, water declination to bagging (Liaw and Wiener 2002). To perform this pro- social-economic poverty status cedure, several parameters must be determined. The first technology development technology development parameter is the number of predicting trees. In this study, Aridity index was calculated by ratio of annual rainfall to potential of 500 trees were created. The second parameter is the num- evapotranspiration proposed by FAO/UNEP for global desertification map AI=P/ETP (UNEP 1992) ber of the predictor variable with no need to prune trees Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 4 of 10 in classification. The proximity matrix was used for identi- area regarding vegetation, where it is 63% for agriculture fying structure in the data (Breiman 2002). The Random indicators. In terms of the technology development, 80% Forest package provides an R interface to the Fortran pro- of the study area gained moderate class of desertification grams by Breiman and Cutler (available at https:// hazard. Erosion criterion including wind and water ero- www.stat.berkeley.edu/~breiman/). sion indicated that 57% of study area are classified in After calculating desertification hazard index, regression high-risk desertification class for wind erosion and 63% trees and random forest technique was applied for identify- areas shows high level for water erosion. Moderate ing preference of main criteria affecting desertification. The conditions classified regarding groundwater and soil sample population involved 25 land units provided based criteria. Therefore, based on impact degree of indicators on geomorphology facies with 21 variables including indica- regarding desertification susceptibility, more than 20% of tors considered for criteria quality as independent variables Taybad-Bakharz region shows susceptible and prone area and desertification hazard as dependent variable. The flow- to desertification. Ultimately, a desertification suscepti- chart of methodology applied in this study is shown in Fig. 2. bility classified in two sensitivity levels involving moder- ate degree with 37% of the area, and high degree with Results and discussion 63% of the areas. The classification categories of the Desertification hazard areas indices and desertification hazard in the form of the The desertification status was calculated on each of sep- map are shown in Fig. 4. It should be noted that all of arated land units. The Kavir areas (salty-clay pans and layers and maps were classified into four classes based Sabkha) are smallest zones and low-level pediment fan on Table 3. The output maps have been presented based and valley terrace deposits cover largest areas in the on gained classes, for example for the geology and cli- studies region as shown in Table 3 and Fig. 3. Further- mate criteria, the desertification does not show severe more, most geomorphology facies involve badlands and class, so this class was ignored in the legend while for pediment deposits surfaces and concentrated on the east erosion criterion, the legend includes all of the classes as and northwest region, where land-use is mainly agricul- it gained all of the desertification risk classes. ture and cultivated areas. The geology map of the stud- ied region shown in the Additional file 2. Regression tree outcomes Results indicated that regarding climate criterion the The results of CART for the 21 targets outcome variables- studied area classified in high vulnerable to desertifica- desertification hazard are presented in Figs. 5, 6,and 7, tion hazard, in othe word shows a high-risk class of respectively. Figures 6 and 7 includes the predictor vari- desertification as 65% of the studied area are susceptible ables and the value that split each subgroup. Within each in relation to climate factors. The results showed that in- node, the mean score or proportion of participants in each dicators considered for geology criterion are main fac- response category are presented. Figure 5 shows the com- tors affecting desertification, so that about 88% areas are plexity parameter (cp) option of the summary function prone-area to desertification. In addition, the criteria of that instructs it to prune the printout, but it does not vegetation and agriculture show undesirable conditions prune the tree. The cp shows minimum error so that the as well as 85% areas considered desertification-prone regression tree is pruned by cp wherever the error shows Fig. 2 General outline method used in the study Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 5 of 10 Table 3 Frequency and distribution of geomorphological facies (lands unit) in Taybad-Bakharz area Land Geomorphological facies Area (ha) % Frequency unit 1 Rocky outcrop (Sandstone and 1387 0.3 1 sandy limestone) 2 Rocky outcrop (Conglomerate 1050 0.22 1 and sandstone) 3 Rocky outcrop (Pale red marl 16,057 3.0 1 and gypsiferous marl) 4 Rocky outcrop (Andesitic to 1165 0.25 1 basaltic volcanic tuff) 5 Rocky outcrop (Andesitic 3238 0.7 1 volcanosediment) 6 Rocky outcrop (Andesitic 15,710 3.3 1 volcanic tuff) 7 Rocky outcrop (Dacitic to 2638 0.55 1 Andesitic tuff) 8 Rocky outcrop (Rhyolitic to 3295 0.7 1 Fig. 3 lands unit map (LUM) in the studied area. The land unit map rhyolitic tuff) provided based on geomorphological characteristics including 25 9 Rocky outcrop (Rhyolitic to 9273 2.0 2 geomorphic facies rhyolitic volcanic) 10 Rocky outcrop (Granite) 4864 1.0 2 geomorphology map (geology and morphology maps). 11 Rocky outcrop (Red marl, 5363 1.2 1 gypsiferous marl) All of indicators were evaluated in each land units (poly- gone) separetely. 12 Hills with gully erosion 10,000 2 4 Given to the Fig. 6, the first split for desertification 13 Badland 41,000 8.7 9 hazard status is based on the score of the wind erosion 14 Rocky outcrop (debris) 12,900 2.7 3 index. The subgroup that represented higher hazard tak- 15 Eroded Inselberg 9980 2.0 3 ing scores (< 2.26, > = 2.26) and was further split based 16 Hills 10,000 2.0 4 on water erosion and then land use change status on the 17 Rocky outcrop (Gneiss, granite, 7600 1.6 2 left (This subgroup does not split again). Also, crop pat- amphibolite) tern status and the next split that was related to the 18 Clay flat 21,700 4.5 3 aridity index and vegetation utilization indices (on the right). Vegetation utilization and land-use change cri- 19 High level pediment fan and 52,950 11.2 12 valley terrace deposits teria showed relatively lower correlation with desertifica- (semi-Cultivated) tion hazard value. 20 Low-level pediment fan and 222,900 47 9 In Fig. 7, Parent group shows the desertification risk valley terrace deposits with mean 1.92, the children group presents the erosion (Cultivated) status including water and wind erosion. The water ero- 21 Sand dune (semi-active) 7220 1.5 1 sion is a sub-tree or a branch. The highest correlation 22 Irregular aspect 4325 1 1 was obtained between wind erosion and desertification. 23 Inselberg 4847 1 1 The water erosion only involves one branch and its 24 Low-level piedmont fan and 5480 1.2 1 division does not continue in regression tree. The crop playa pattern shows highest correlation with wind erosion with 25 Kavir (Salty-clay pan, Sabkha) 230 .1 1 continuous division and after that, aridity index indicates correlation with crop pattern. Therefore, desertification risk is affective by wind erosion with a powerful correl- minimum amount, the regression tree is pruning. The ation with crop patterns and aridity index. In addition, pruning based on cp is a validation test in regression tree, as water erosion shows a weakness correlation, we can indeed before regression; a training was used by 30% of have a judgement that wind erosion has most powerful total data. For more information the raw data used in stat- correlation with desertification in all of land units. istical methods has been presented in Additional file 3. The regression tree used when we have numerical var- In the Fig. 6, n shows the land units and means the iables. This method tries to recognize highest correlation amount of the polygons which was determined based on between variables by using Gini coefficient, Chi-square, Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 6 of 10 Climate Geology Soil Vegetation Cover Erosion Low Low Low Low Low Moderate Moderate Moderate Moderate Moderate High High Severe Severe High High High 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 Agriculture Technology Development Soci−Economic Groundwater Desertification Susceptibility Low Low Low Low Low High Moderate Moderate Moderate Moderate High High 0 20000 meters 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 780000 820000 860000 Fig. 4 desertification susceptibility areas for main indicators. Desertification susceptibility map provided based on preference degree of each indicators MSE, and entropy. The Gini coefficient find homogeny impact on desertification hazard, respectively. These in- variable and regression tree will be found correlations dicators are highly associated with the value of desertifi- based on mean of variables. The regression presents a cation hazard. Wind erosion and cropping pattern classification by Gini coefficient by number of successes indices in 18 land units are identified as the most im- and failures. portant factor affecting desertification hazard. Also, arid- The cross-validation shows only a small difference in ity index is listed as the most important factors of desertification hazard status. Given to the regression desertification in 13 land units. Although, water erosion trees outcomes, among 21 independent variables, indica- index have less impact on the hazard of desertification tors such as wind erosion, precipitation, soil EC, soil tex- (7 land units). Finally, wind erosion, cropping pattern, ture, groundwater (SAR) and slope had the greatest aridity index and precipitation are identified as the most Fig. 5 cp Plot. A validation test for regression tree method based on minimum error 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 3800000 3850000 3900000 Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 7 of 10 Fig. 6 regression Tree with Package rpart important indicators that affected desertification hazard development, aridity index, slope index, soil EC, land- in this region. use changes, vegetation state, precipitation, geology, water declination, and soil texture. Moreover, based on variable importance for the RF-model, we observed that Random Forest outcome wind erosion, technology development, aridity index, The results from the variable selection random forest slope index, vegetation state and land-use change variables are shown in Figs. 8 and 9. Independent variables in- are relatively most important on desertification hazard of volving 21 indicators ordered based on mean decrease Taybad-Bakharz region, respectively. The important per- accuracy. The accuracy measure determined main cent values of these variables measured 10.48%, 7.5%, effective criteria including wind erosion, technology 6.8%, 5.9% 5.3%, and 5.2% respectively. Fig. 7 selected regression tree Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 8 of 10 Fig. 8 plot showing the decrease of (in terms of out-of-bag classification errors, OOBE) OOBE with increasing number of trees T# in the RF structure. A working value of T# = 300 was chosen for the RFtb model structure used in the tests and experiments As shown in Fig. 8, in forest random method, at the and migration, and ultimately ecosystem collapse in Iran first will be created classification trees that called tree of with more than 75% drylands require a carefully under- vote. The forest will be decided and classified base on standing of the desertification process and recognizing the most vote, then average of tree’s outcome presents driving forces. This study examines the performance of a the regression. In Fig. 9, the preference of each variable statistical method to identify the most important criteria was calculated by mean square error (MSE) and purity affecting desertification process and risk. Studied region or entropy degree. The less MSE indicates an indicator showed desertification-prone areas with 63% high-risk with more preference. level of desertification. Application of regression trees and random forest techniques identified the most important criteria affecting desertification and recog- Conclusions nized that indicators such as wind erosion, technology de- Land degradation and desertification consequences such velopment, aridity index, slope index, precipitation, as dust storms, drying of lakes, water scarcity, poverty Fig. 9 variable importance calculated by mean square error (MSE) and purity or entropy degree Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 9 of 10 vegetation state and land use change are major indicators Received: 27 May 2017 Accepted: 23 February 2018 affecting the quality of criteria and desertification in the Taybad-Bakharz. The results of this research indicated References that erosion factors including water and wind erosion are Breiman, L., J. Friedman, C.J. Stone, and R. Olshen. 1984. Classification and the most important desertification factors in the studied regression trees. CRC press, Taylor and Francis Group. pp. 246–280. area. Over-grazing and vegetation degradation in the re- Breiman, Leo. 2001. Manual on Random Forests. University of California. p. 33. https://www.stat.berkeley.edu/~breiman/randomforest2001.pdf. gion particularly in recent decade led to degraded soils Breiman, Leo. 2002. Manual on setting up, using, and understanding random and decreasing fertility, so the erosion rate is raising in the forests. Vol. v3.1 https://www.stat.berkeley.edu/~breiman/. study area. Data mining process and particularly regres- Catani, F., D. Lagomarsino, S. Segoni, and V. Tofani. 2013. Landslide susceptibility estimation by random forests technique: Sensitivity and scaling issues. sion trees and random forest technique in desertification Natural Hazards and Earth System Sciences 13 (11): 2815–2831. hazard can be recommended as a robust base for further De Pina Tavares, J., I. Baptista, A.J.D. Ferreira, P. Amiotte-Suchet, C. Coelho, S. investigation of desertified lands for best management of Gomes, R. Amoros, E.A. Dos Reis, A.F. Mendes, L. Costa, J. Bentub, and L. Varela. 2015. Assessment and mapping the sensitive areas to desertification processes in these areas. in an insular Sahelian mountain region case study of the Ribeira Seca watershed, Santiago Island, Cabo Verde. Catena 128: 214–223. https://doi. org/10.1016/j.catena.2014.10.005. Endnotes 1 Ferrara, A., L. Salvati, A. Sateriano, and A. Nolè. 2012. Performance evaluation and - Tool in this research used for indicators system cost assessment of a key indicator system to monitor desertification vulnerability. Ecological Indicators 23: 123–129 https://doi.org/10.1016/j. ecolind.2012.03.015. Additional files Frisman L. (2008) App lying classification and regression tree analysis to identify Priso Ners with high lllV risk Behaviorst. 40 (December). Additional file 1: Digital Elevation Model (DEM) of studied area. (JPEG Karamesouti, M., V. Detsis, A. Kounalaki, P. Vasiliou, L. Salvati, and C. Kosmas. 2015. 125 kb) Catena land-use and land degradation processes affecting soil resources : Evidence from a traditional Mediterranean cropland (Greece). Catena 132: Additional file 2: Geology map of Taybad-Bakharz in northeastern Iran. 45–55. https://doi.org/10.1016/j.catena.2015.04.010. (JPEG 134 kb) Kosmas, C., M. Kirkby, and N. Geeson. 1999. The MEDALUS project Mediterranean Additional file 3: The raw data used in statistical algorithms. (CSV 3 kb) desertification and land use; manual on key indicators of desertification and mapping environmentally sensitive areas to desertification. Brussels: European Commission. Abbreviations Ladisa, G., M. Todorovic, and L. Trisorio. 2012. A GIS-based approach for CART: Classification And Regression Tree; cp: Complexity Parameter; desertification risk assessment in Apulia region, SE Italy. Physics and Chemistry ESAs: Environmental Sensitive Areas; LUM: Land Unit Map; of the Earth 49: 103–113. https://doi.org/10.1016/j.pce.2011.05.007. MEDALUS: Mediterranean Desertification and Land Use; MSE: Mean Square Leman, N., M.F. Ramli, and R.P. Khairani Khirotdin. 2016. GIS-based integrated Error; OOBE: Out-Of-Bag classification Errors evaluation of environmentally sensitive areas (ESAs) for land use planning in Langkawi, Malaysia. Ecological Indicators 61: 293–308. Acknowledgements https://doi.org/10.1016/j.ecolind.2015.09.029. The authors are thanks from administration of Natural Resources and Liaw, A., and M. Wiener. 2002. Classification and regression by randomForest. R Environment College, Ferdowsi University of Mashhad for supporting and news 2 (December): 18–22. providing the facilities, particularly Dr. Naseri and Dr. Mosaedi who are head Martínez-Valderrama, J., J. Ibáñez, G. Del Barrio, M.E. Sanjuán, F.J. Alcalá, S. of college and vice-head separately. Martínez-Vicente, A. Ruiz, and J. Puigdefábregas. 2016. Present and future of desertification in Spain: Implementation of a surveillance system to prevent Funding land degradation. Sci Total Environ. 563–564: 169–178. https://doi.org/10. The Gorgan University of Agricultural Sciences and Natural Resources and 1016/j.scitotenv.2016.04.065. Iran science and technology ministry provided the funding of this research. Patriche, C., and G. Bandoc. 2017. Quantification of land degradation sensitivity areas in southern and central southeastern Europe. New Availability of data and materials results based on improving DISMED methodology with new climate Not applicable. data. Catena 158: 309–320. https://doi.org/10.1016/j.catena.2017.07.006. Patriche, C., M. Dumitra, and G. Bandoc. 2017. Catena spatial assessment of land Authors’ contributions degradation sensitive areas in southwestern Romania using modi fi ed MS conducted the fieldwork and contributed to the analysis of the data as MEDALUS method. 153: 114–130. https://doi.org/10.1016/j.catena.2017.02.011. well as in writing the first draft. MO is a supervisor for the research, and AN Pourghasemi, H.R., and N. Kerle. 2016. Random forests and evidential belief function- is an adviser. AS supervised data analysis and contributed in revising the based landslide susceptibility assessment in western Mazandaran Province , Iran. manuscript and providing the final draft. He is corresponding the research. Environmental Earth Sciences. https://doi.org/10.1007/s12665-015-4950-1. All authors read and approved the final manuscript. Ramageri, M. 2010. Data mining techniques and applications. Indian J Comput Sci Eng 1 (4): 301–305. Salvati, L, Kosmas C, Kairis O, Karavitis C, Acikalin S, Belgacem A, Chaker M, Competing interests Fassouli V, Gokceoglu C, Gungor H, Hessel R, Sol A, Khatteli H, Kounalaki A, The authors declare that they have no competing interests. Laouina A, Ocakoglu F, Ouessar M, Ritsema C, Colantoni A, Carlucci M (2016) Assessing the effectiveness of sustainable land management policies for Publisher’sNote combating deserti fi cation : A data mining approach 183, 754–762. https:// Springer Nature remains neutral with regard to jurisdictional claims in doi.org/10.1016/j.jenvman.2016.09.017. published maps and institutional affiliations. Sepehr, A., A.M. Hassanli, M.R. Ekhtesasi, and J.B. Jamali. 2007. Quantitative assessment of desertification in south of Iran using MEDALUS method. Author details Environmental Monitoring and Assessment 134 (1–3): 243–254. Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, Skurichina, M., and R.P.W. Duin. 2002. Bagging, boosting and the random Golestan, Iran. Department of Desert and Arid Zones Management, subspace method for linear classifiers. Pattern Analysis and Applications Ferdowsi University of Mashhad, Mashhad, Iran. 5(2): 121–135. Sarparast et al. Geoenvironmental Disasters (2018) 5:3 Page 10 of 10 UNCCD. 1994. Elaboration of an international convention to combat desertification in countries experiencing serious drought and/or desertification, particularly in Africa,1–58 (June). http://www2.unccd.int/. UNEP. 1992. World atlas of desertification. London: Edward Arnold. Wijitkosum, S. 2016. The impact of land use and spatial changes on desertification risk in degraded areas in Thailand. Sustainable Environ Res 26 (2): 84–92. https://doi.org/10.1016/j.serj.2015.11.004. Yang, R.M., G. Zhang, F. Liu, Y. Lu, F. Yang, F. Yang, M. Yang, Y.G. Zhao, and D.C. Li. 2016. Comparison of boosted regression tree and random forest models for mapping topsoil organic carbon concentration in an alpine ecosystem. Ecological Indicators 60: 870–878. https://doi.org/10.1016/j.ecolind.2015.08.036. Yeh, Chyon-Hwa. 1991. Classification and regression trees (CART). Chemometrics and Intelligent Laboratory Systems 12 (1): 95–96. https://doi.org/10.1016/0169- 7439(91)80113-5. Zambon, I., A. Colantoni, M. Carlucci, N. Morrow, A. Sateriano, and L. Salvati. 2017. Land quality , sustainable development and environmental degradation in agricultural districts : A computational approach based on entropy indexes. Environmental Impact Assessment Review 64: 37–46. https://doi.org/10.1016/j. eiar.2017.01.003.

Journal

Geoenvironmental DisastersSpringer Journals

Published: Dec 1, 2018

Keywords: Environment, general; Earth Sciences, general; Geography, general; Geoecology/Natural Processes; Natural Hazards; Environmental Science and Engineering

References