Soil Salinity Mapping by Multiscale Remote Sensing in Mesopotamia, Iraq

Soil salinity has become one of the major problems affecting crop production and food security in Mesopotamia, Iraq. There is a pressing need to quantify and map the spatial extent and distribution of salinity in the country in order to provide relevant references for the central and local governments to plan sustainable land use and agricultural development. The aim of this study was to conduct such quantification and mapping in Mesopotamia using an integrated, multiscale modeling approach that relies on remote sensing. A multiyear, multiresolution, and multisensor dataset composed of mainly Landsat ETM+ and MODIS data of the period 2009-2012 was used. Results show that the local-scale salinity models developed from pilot sites with vegetated and nonvegetated areas can reliably predict salinity. Salinity maps produced by these models have a high accuracy of about 82.5-83.3% against the ground measurements. Regional salinity models developed using integrated samples from all pilot sites could predict soil salinity with an accuracy of 80% based on comparison to regional measurements along two transects. It is hence concluded that the multiscale models are reasonably reliable for assessment of soil salinity at local and regional scales. The methodology proposed in this paper can minimize problems induced by crop rotation, fallowing, and soil moisture content, and has clear advantages over other mapping approaches. Further testing is needed while extending the mapping approaches and models to other salinity-affected environments.

the severity and distribution of soil salinity varies with space 37 and time [2]- [4]. In order to prioritize any remediation effort 38 and better plan for agricultural improvements and food security, 39 it is of prime importance for Iraqi central and local govern-40 ments to understand the distribution and severity of salinity in 41 Mesopotamia. 42 Soil salinity is a common form of land degradation in 43 irrigated areas located in dryland environments [5]- [8]. The 44 physical appearance of salinity is strongly influenced by soil 45 properties (e.g., moisture, texture, mineral composition, and 46 surface roughness) as well as type of vegetation cover (e.g., 47 halophyte and nonhalophyte, salt-tolerant and nonsalt-tolerant) 48 [5]- [8]. Remote sensing has been widely applied for mapping 49 and assessment of soil salinity in recent decades using veg- 50 etation indices (VIs) and combined spectral response index 51 (COSRI) [9]- [16], best band combination [17], [18], maximum 52 likelihood and fuzzy logic-based classifications [19]- [23], prin-53 cipal component analysis (PCA), surface feature unmixing, 54 and data fusion [6], [7], [24]. Predictive models have been 55 developed for soil salinity using different regression analysis, 56 artificial neural network (ANN), and Kriging/CoKriging tech-57 niques [9]- [16], [18], [24]- [26]. Very recently, along with 58 vegetation indices and reflectance of certain spectral bands, 59 evapotranspiration (ET) and land surface temperature (LST) 60 have been used to predict salinity in salt-affected areas 61 [16], [27]- [29]. 62 While these and other studies demonstrate the feasibility, 63 advantages, and potential of remote sensing to assess soil salin-64 ity, there remain certain challenges. First, although in strongly 65 salinized areas, salt tends to concentrate on terrain surfaces 66 and can be easily detected by conventional remote sensing 67 tools; however, for low-to-moderate salinity (salt <10−15%), 68 spectral confusions with other different surface features may 69 arise leading to identification failure (either overestimation or 70 underestimation) [6], [7]; especially, when salt concentrates in 71 subsoil, optical remote sensing is restricted [8]. Second, soil 72 moisture, halophyte vegetation, and salt-tolerant crops such 73 as barley, cotton, and alfalfa can modify the overall spectral 74 response pattern of salt-affected soils, especially in the green 75 and red bands [6], [7], [30]. Third, lands in the states of fal-76 low, noncrop interval in-between rotations, and crop rotations 77 tend to be interpreted as salinized areas if only soil bareness or 78 vegetation greenness of a single image is investigated. To avoid 79 these problems, some authors have suggested: 1) to use images 80 acquired at the end of dry or hot season or of multiple cropping 81 periods [7], [8], 2) to conduct regression analysis against VIs 82 [9]- [16] and geophysical measurement [8] in combination with 83 soil sampling and analysis. These are, no doubt, useful sugges- 84 tions to minimize the mentioned problems and accomplish a 85 better mapping work. However, most of the available studies 86 have employed single or multidate single images to assess salin-87 ity at local scale, and their approaches are not fully repeatable 88 or extendable for regional-scale assessment due to spatial vari-89 ability and diversity in climate conditions, soil properties, and 90 land use/management. It is, therefore, essential to develop new 91 processing methods and approaches technically operational for 92 regional-scale salinity mapping. 93 The main objectives of this study are, hence, to develop an 94 integrated methodology operational for regional salinity quan-95 tification and assessment based on the available approaches 96 considering the above-mentioned problematic issues, to pro-97 vide relevant multiscale salinity maps for Iraqi governments, 98 and finally, to lay a foundation for the successive regional-scale 99 tracking of salinity change trends in space and time that may 100 provide spatial reference for the governments to understand 101 the impacts of land management on salinization processes in 102 Mesopotamia. 103 As well as for salinity assessment, remote sensing technol-104 ogy has also been widely applied in other dryland research. Mesopotamia, "the land between rivers" in ancient Greek and 144 encompassing a surface area of about 135 000 km 2 , is a typ-145 ical alluvial plain between the two famous rivers, Euphrates 146 and Tigris (Fig. 1) and the home of multiple ancient civiliza-147 tions namley Sumerian, Akkadian, Babylonian, and Assyrian 148 [4]. As an arid subtropical region, the climate is characterized 149 by dry hot summers and cooler winters [2], [3], [29], where 150 annual rainfall is mostly below 200 mm, of which the average 151 is 110 mm in Baghdad and 149 mm in Basrah in the past three 152 decades. The mean maximum and minimum temperatures are 153 44 • C and 25.6 • C, respectively, in Baghdad, 46 • C and 29.15 • C 154 in Basrah in July-August, whereas they are 16.5 • C and 4.8 • C 155 in Baghdad, 19 • C and 8.4 • C in Basrah in December-January. 156 As a fluviatile plain, soils are extremely calcareous (20-30% 157 lime) alluvial silty loam or loamy silts [2], [3], typical 158 Fluvisols in terms of WRB (the World Reference Base for 159 Soil Resources), and mostly saline as a result of cumula-160 tive salinization in the past 6000 years [2]- [4]. Archeological 161 evidence revealed that crop cultivation (e.g., wheat and bar-162 ley) was started as early as 4000 BC in Mesopotamia [2], 163 [4]. Due to aridity, farming is impossible if not irrigated. 164 Irrigation increases soil moisture and crop production, nonethe-165 less, leads to elevation of water-table or water-logging in the 166 area where there is no drainage or draining is slow [2]- [4]. 167 Consequently, salts accumulate in soils after evaporation and 168 transpiration year by year. According to Jacobsen and Adams 169 [4], salinity had already become a serious hazard in south-170 ern Mesopotamia in the late Sumerian or early Akkadian 171 periods, e.g., around 2400-2300 BC, and led to a decline 172 in wheat production. The proportions of wheat and barley 173 were nearly equal in about 3500 BC but became 1 to 6 174 in 2400 BC in Girsu (nowadays Thi-Qar); wheat cultivation 175 was completely abandoned after 1700 BC and land produc-176 tivity declined from 2537 l/ha before 2400 BC to 897 l/ha in 177 1700 BC in Larsa (also in Thi-Qar) as a consequence of salin-178 ization. Salinity is hence an old problem that contributed to 179 the breakup of ancient civilization [4]. Unfortunately, saliniza-180 tion has never stopped but progressively extended to the whole 181 Mesopotamian plain to the state as described in the beginning 182 of the paper. 183 As Buringh investigated [2], the most common salt in 184 saline soils is sodium chloride (NaCl) followed by other 185 chlorides (e.g., CaCl 2 , MgCl 2 , and KCl), and sulfates (e.g., 186 CaSO 4 ·2H 2 O, Na 2 SO 4 .10H 2 O, and MgSO 4 ). Saline-alkaline 187 soils may exist locally but real alkali soils (in black) are very 188 scarce in Mesopotamia.

190
To achieve our objectives, comprehensive observations and 191 measurements at different scales are required. The experi-192 ment was hence designed to be conducted at three levels, i.e., 193 plot, local (pilot site), and regional scales, corresponding to 194 the proposed multiscale approach. Both local (pilot site)-and 195  However, the apparent salinity has to be calibrated by labora-208 tory measured soil salinity. The false salinity caused by metal 209 and/or soil moisture should be avoided while measurement is 210 conducted. 211 In order to be comparable with the pixels of high-resolution 212 satellite images such as Landsat and SPOT (e.g., 10-30 m), the 213 survey was planned to be conducted in three plots distributed 214 at three corners of a triangle, respectively, with a distance of 215 about 15-20 m from each other in the same patch of land. The 216 averaged values of the EM38 readings including both EM V and 217 EM H of the three corner plots would be taken to represent the 218 salinity of the center of the observed triangle. Soil samples for 219 laboratory chemical analysis were to be taken from soil profiles 220 at the depth of 0-30, 50-70, 90-110 and 120-150 cm, and from 221 surface (0-30 cm in depth) using auger tools in the plots where 222 EM38 was also measured. 223 Pilot site level survey was to serve for integrated pilot 224 study, e.g., salinity model development and mapping at local 225 scale. As recommended by the Iraqi government, five sites 226 namely Musaib, Dujaila, West Gharraf, Shat-Al-Arab, and Abu 227  Table I. 248 In order to extend plot level measurements to pilot site, and 249 then to regional-scale salinity mapping, a multiyear dataset As indicated in Section I, apart from the geophysical survey 260 by EM38 meter to understand salinity in surface and subsoil, 261 different remote sensing indicators that can characterize the 262 multiaspect surface biophysical features, e.g., VIs, LST, soil 263 brightness (albedo), and principal components (PCs), need to 264 be derived. 265 Instead of using one single image, a 4-year imagery dataset 266 registered both spring and summer acquisitions, which was 267 used to derive the multiyear maximal values of a set of VIs and 268 nonvegetation indices (NonVIs) for each pixel. This would help 269 avoiding some false alarm of salinity arising from fallowing, 270 crop rotation, and variation in soil moisture. This processing 271 can also largely remove the problem caused by the image gaps 272 left by the Scan-Line Corrector failure (SLC-Off) in the Landsat 273 ETM+ imagery since 2003. We assumed that it is always possi-274 ble for a given piece of cropland to be cultivated in either spring 275 or summer with normal performance in the observed period 276 because fallow state lasts, in general, 2-3 years in Central and 277 Southern Iraq. 278 Image processing in combination with field survey would 279 allow the identification of the salt-tolerant areas, and the con-280 centration of salt in subsoil, for example, areas with high 281 vegetation greenness but moderate salinity as revealed by the 282 readings of EM 38 or as measured by soil laboratory analy-283 sis. Such areas have to be defined for a specific analysis since 284 salinity cannot be reflected by vegetation indices. 285 Furthermore, it is essential to separate vegetated and non-286 vegetated areas, as the expression of salinity in remote sensing 287 images is different in these two types of areas. For exam-288 ple, the low values of VIs in nonvegetated areas (e.g., bare 289 soil and desert) do not mean that they are all strongly salin-290 ized (high salinity). As a matter of fact, salinity is negatively 291 correlated with VIs such as NDVI [11], [13], [28], [29], and 292 it tends to be overestimated in the nonvegetated areas just 293 based on VI-related models. We have to consider the inte-294 grated information from multiple spectral and thermal bands, 295 e.g., spectral reflectance, LST, PCs, and the brightness of 296 the Tasseled Cap transformation (TCB) [47]- [49], for salin-297 ity assessment in these areas. The rationale behind is that 298 the spectral reflectance and its multiband linear combination 299 (e.g., TCB and PCs) together with LST might be able to 300 highlight the subtle difference in soil brightness (or albedo) 301 corresponding to the difference in salinity in the nonvegetated 302 areas. 303 The procedure for local-scale study in the pilot sites is 304 presented as follows. 305 305 1) Atmospheric correction using FLAASH model [50] for all 306 Landsat ETM+, SPOT, and RapidEye images. 307 2) Multispectral transformation: A set of most frequently 308 applied VIs such as NDVI [31], SAVI (soil-adjusted 309 vegetation index) [51], SARVI (soil-adjusted and atmo-310 spherically resistant vegetation index) [52], and EVI 311 (enhanced vegetation index) [53] were produced from 312 the atmospherically corrected and reflectance-based satel-313 lite imagery. We also introduced a new vegetation index 314 in this work, the generalized difference vegetation index 315 (GDVI) developed by Wu [54] and in the form of 1) Regional-scale modeling: Models obtained from any 390 pilot site cannot be directly applied to regional-scale 391 salinity mapping due to lack of spatial representative-392 ness. That is why we proposed here a "multiscale 393 modeling" approach to upscale plot-level measurements 394 and high-resolution-derived models to regional-scale 395 assessment. To do so, the data from different pilot sites, 396 which are situated in different locations in Mesopotamia 397 (Fig. 2), were integrated together for regional-scale mod-398 eling using the same multiple regression model. 399 2) Upscaling test and regional salinity mapping: Since we 400 will use MODIS data (VIs and LST) for regional salinity  (2) and (3) 3) Validation: The regional salinity map derived from the 434 MODIS data was evaluated against the field samples from 435 two regional transects (blue points in Fig. 2 After the above processing, both local-and regional-scale 439 salinity models obtained are listed in Table II, and local-scale 440 and regional-scale salinity maps are presented in Figs. 3 and 4 441 for discussion.

443
As our test revealed in the Dujaila site [29], specific models 444 for vegetated and nonvegetated areas were not recommended 445 for salinity mapping due to their low reliability (e.g., < 37%). 446 Thus, what are presented in Table II are the integrated mod-447 els taking all the samples into account, whereas vegetated and 448 nonvegetated areas were separated during the multiple linear 449 regression analysis in each pilot site. We see that among all 450 the VIs, GDVI or its variant such as ln(GDVI) is the most rep-451 resentative indicator for vegetated areas, and LST (and NDII) 452 for nonvegetated areas in all pilot sites. By the way, for sites 453 Shat-Al-Arab and Abu Khaseeb, independent models were 454 not developed due to limited soil sample number (8 and 5, 455 respectively). 456 It is also noted that the salinity models obtained are different 457 from each other in all pilot sites; none of them can be directly 458 extended to regional-scale mapping due to spatial variability. 459 However, these models can reliably predict soil salinity with 460 an accuracy of about 82.57% in Dujaila and 83.01% in Musaib 461 against the field measured data. Hence, they were considered 462 operational for their respective pilot sites.   For the regional-scale models, the multiple correlation coef-464 ficients R 2 are relatively lower than those in pilot sites due to 465 homogenization of samples from different pilot sites after inte-466 gration; nonetheless, they have higher applicability in regional-467 scale mapping. 468 It is worth mentioning that most of the EM38 measurements 469 in spring (March-April) 2012 did not show any promising cor-470 relation with VIs except for the Dujaila site perhaps due to 471 the problem of soil moisture after rainfall or irrigation while 472 measurements were undertaken in the field. For this reason, a 473 . We consider that these maps are reliable. 485 As for the regional salinity map (Fig. 4), the accuracy evalu-486 ation revealed that 23 of the 121 regional samples taken along 487 two transects and the surface EC of 27 soil profiles in pilot sites 488 that were not used for modeling were abnormal due to inter-489 nal problem of samples, most probably, derived from laboratory 490 analysis (because the correlation among Cl − , Na + , and EC is 491 very low, e.g., R 2 = 0.047); however, the remaining 98 samples 492 show a good accordance with remote sensing predicted salinity. 493 The observation accuracy is 80.9%, and the statistical accuracy 494 of the regional salinity map obtained by linear regression analy-495 sis at the confidence level of 95% is 80.02% (Fig. 5). Therefore, 496 the regional map presented in Fig. 4 was considered reliable. 497 The agreement between the measured and remote sensing 498 predicted salinity as shown in Fig. 5 is higher in the high salin-499 ity part than low salinity one. This is probably due to the fact 500 that coarse-resolution LST has lower sensitivity to low salin- 501 ity. An overestimation of about 2-10 dS/m may occur in some 502 places in the weakly salinized areas. However, the sensitivity 503 to low salinity can be improved if high resolution LST data are 504 available. 505 One may have concern about the reasonability to use soil 506 surface temperature, LST, as salinity indicator which was 507 finally retained in the models for the nonvegetated areas. As 508 Wu et al. [29] argued, it is commonly known that thermal 509 conductivity of materials is temperature (T )-dependent, and 510 the former is associated with electrical conductivity (EC). 511 However, the interrelationship between the thermal and 512 electrical conductivities is complex and may change signifi-513 cantly depending on materials, e.g., soil types. Some authors 514 [5]-[7] have explored the possibility to use the thermal band 515 to identify the salt-affected soils but they have not discussed 516 the mechanism behind. Abu-Hamdeh and Reeder [57] ascer-517 tained the relationship between thermal conductivity and salin-518 ity, and found that thermal conductivity decreases with the 519 increase in the amount of added salts at given moisture content. 520 Sepaskhah and Boersma [58] found that the apparent thermal 521 conductivity is independent of water content at very low water 522 contents. Consequently, in driest condition (at lowest moisture 523 or water content), thermal conductivity is associated with the 524 salt amount-salinity. We believe, therefore, that LST-based 525 models are relevant for mapping salinity in nonvegetated areas. 526 Concern may also be addressed on the applicability of the 527 models. It is clear that the models obtained from pilot sites 528 are not recommended for direct application to similar areas for 529 salinity mapping without relevant adaptation. Of higher repre-530 sentativeness, the regional-scale models can be disseminated to 531 the similar environment for this purpose.

533
Different from the other authors (e.g., [10], [17], and [18]), 534 we used multiyear imagery dataset to derive the multiyear 535 maxima of VIs and NonVIs for multiscale salinity model-536 ing followed with an upscaling analysis. The above-mentioned 537 problematic issues that are commonly faced in salinity mapping 538 by remote sensing were successfully minimized, and salinity 539 maps with high reliability were produced. 540 Despite a number of authors [10], [17] have conducted salin-541 ity mapping and best band combination studies, but they used 542 single or multiple single images and did not differently treat the 543 vegetated and nonvegetated areas. Especially, authors [17] did 544 not take into account the nonvegetated area. Their approaches 545 cannot avoid the influences from crop rotation/fallow, and 546 moisture, which are often problematic in large area (or scale) 547 mapping. Hence, our approach has evident advantages over and 548 its uniqueness different from others. 549 However, some imperfection was also noted. As a matter 550 of fact, salinity has strong spatial variability; even in a small 551 1 × 1 m 2 plot, salinity may change after each 20-30 cm inter-552 val, not to mention in the 250 m pixels of MODIS data which 553 were used for regional-scale mapping in this study. That is to 554 say, it is unlikely to produce a regional salinity map with an 555 accuracy of 2-3 dS/m based on the proposed methodology. 556 What can be done is to approach the reality as much as possible 557 by increasing the sampling number and density with a relevant 558 spatial distribution if both time and fund are available.

896
He is an Agricultural Hydrologist with the 897 International Center for Agricultural Research in the 898 Dry Areas (ICARDA), Amman, Jordan. During a 899 9-year Postdoctoral research career, he has served as 900 a PI of co-PI on research projects worth about 5.75 901 million, and authored or coauthored 59 technical pub-902 lications that include 22 refereed journal articles in national or international 903 journals. He is an internationally recognized authority in hydrologic and water 904 quality modeling and GIS applications in water resources management. He has 905 offered more than 20 trainings (covering a total of 400 participants) on hydro-906 logic modeling in 10 countries. He has served as Research Advisor/Committee 907 Member to M.S. and Ph.D. students and was a Visiting Assistant Professor 908 (2007)(2008)(2009)(2010)(2011)

933
He has published 11 books and book-chapters, 53 refereed journal papers, 934 and more than 100 nonrefereed conference abstracts and proceedings papers.

935
Dr. Payne has been named Fellow of five international scientific societies 936 and has held numerous leadership roles at the state, national, and international 937 levels.