Rothamsted Repository Download

Study focus: The main objective of this study was to apportion sub-basin spatial source relative contributions to target channel bed sediment samples using a composite ﬁ ngerprinting procedure including a Bayesian un-mixing model. In total, thirty-four geochemical tracers, eleven elemental ratios and di ﬀ erent weathering indices were measured or estimated for 43 tributary sediment samples collected to characterise three sub-basin spatial sediment sources and eleven target bed sediment samples collected at the outlet of the main basin. Statistical analysis was used to select three di ﬀ erent composite signatures. New hydrological insights for the region: Using a composite signature based on KW-H and DFA, the respective relative contributions (with uncertainty ranges) from tributary sub-basins 1, 2 and 3 were estimated as 54.3% (47.8 – 62.0), 11.4% (4.2 – 18.7) and 34.3% (27.6 – 39.9), compared to 72.0% (61.6 – 82.7), 13.6% (9.0 – 18.5) and 14.2% (3.1 – 25.4) using a combination of KW-H and data mining, and 50.8% (42.8 – 59.9), 28.7% (20.2 – 37.3) and 20.3% (12.7 – 27.2) using a ﬁ ngerprint selected by KW-H and PCCA. The root mean square di ﬀ erence between these source estimates highlighted sensitivity to the composite signatures. Evaluation of the un-mixing model predictions using virtual mixture tests con ﬁ rmed agreement between modelled and known source proportions.

monthly means of 11 m 3 s −1 and 5.81 m 3 s −1 , respectively. On the basis of 10 field surveys and associated observations over a 12month period, it was established that the three sub-basins ( Fig. 1) comprising sub-basin 1 (168.3 km 2 ), sub-basin 2 (63.7 km 2 ) and sub-basin 3 (43 km 2 ) in the study area dominate tributary sediment inputs to the main river. Land use in the three sub-basins is summarized in Table S1 in Supplementary Information.
From a stratigraphical point of view ( Fig. 1), ancient geological formations belonging to the Precambrian period and more recent formations belonging to the Quaternary period are observed in the study area (see Table S2 in Supplementary Information). Based on field observations, erosion is very evident across the study area. After any rainfall event, a large volume of soil undergoes erosion and subsequent transport and is deposited on the main stem channel bed at the basin outlet. The average daily suspended sediment export has been estimated at 3814 tons (Iran Water Resources Management Company). Runoff and erosion has important off-site impacts in this drainage basin with, for example, muddy floods affecting crop fields and gardens (Fig. 2). As a result, considerable areas of the gardens and agricultural land surrounding the river have been affected by sediment mobilization and delivery, leading to social and economic problems in the locality. Measures for controlling runoff such as terracing, strip cropping and improved vegetation cover are still lacking.
The Kamish River catchment has at its center Harsin City and it is also home to a large number of villages mostly located alongside the banks of the Kamish River. Most villagers in this area are heavily dependent on agriculture, gardening and animal husbandry as their main sources of income. Hence, soil protection in this area is critical for the sustainability of the villages and the wellbeing of villagers. At the present time, the Kamish dam (4.5 km to the south of Harsin City and about 1.5 km to the north of Baba Zeyd village) with a storage capacity of 16.6 million m 3 and an average annual input flow of 20.5 million m 3 is being constructed. Current high levels of erosion and sediment production in the region have the potential to impact on the water storage capacity behind the dam. Accordingly, appraisal of the spatial origins of sediment are very important for controlling and managing this risk to new critical infrastructure for supporting livelihoods and well-being.

Field sampling
Although most sediment fingerprinting studies have relied on sets of suspended sediment samples from various flood events or different times across flood hydrographs (Haddadchi et al., 2014;Hughes et al., 2009), fine-grained sediment deposited on the river bed has also been successfully used as an alternative type of target sediment in various environmental situations (Dirszowsky, 2004;Evrard et al., 2011Evrard et al., , 2013Haddadchi et al., 2014Haddadchi et al., , 2013Hughes et al., 2009;Lamba et al., 2015;Nosrati et al., 2018a;Olley and Caitcheon, 2000;Smith and Blake, 2014). This type of target fine-grained sediment, which has been referred to in some previous studies as "drape" sediment (Nosrati et al., 2018a), can be assumed to provide a surrogate for suspended sediment-associated geochemistry (Horowitz et al., 2012) and to be more temporally representative than snapshot instantaneous suspended sediment samples (Zhou et al., 2016), whilst negating the need for water sample collection during actual storm periods (Miller et al., 2015). On this basis, samples of sediment deposited on the river bed were collected at the outlet of the main drainage basin and from the outlets of the three sub-basins (Fig. 1). Here, the samples collected from the main basin are used as the target sediment for apportioning the relative contributions from the upstream sub-basins (Nosrati et al., 2018a). Each bed sediment sample was a combination of 5 subsamples each of which was taken at an interval of 200 m along the selected reach. All samples were collected using a non-metallic shovel to prevent sample contamination (Franz et al., 2013) and using a portable stilling well to avoid the winnowing of fines during sample extraction through the water overlying the channel bed. The inclusion of sub-samples was aimed at taking some account of the local scale spatial variation in bed sediment tracer properties. In total, 10 composite samples were retrieved from the main basin outlet and a further 43 composite samples from the three sub-basin outlet sampling locations. The mass of each composite bed Fig. 2. Photos taken at the outlets of (A) the main catchment (target sediment sampling location for apportioning sub-basin sources) and three upstream sub-basins (representing potential spatial sediment sources): (B) sub-basin 1, (C) sub-basin 2, and (D) sub-basin 3. Photos showing the offsite impacts of sediment mobilisation and delivery on (E) gardens and (F) agricultural land in the Kamish River catchment. sediment sample was˜1 kg. Due to possible problems associated with the erosion and collapse of channel banks proximal to the bed sediment sampling points, and the risk of bias in the bed-sediment associated geochemistry (Zhou et al., 2016), the downstream reaches on either the sub-basins or on the main stem were pinpointed to channel sections where local bank erosion was not evident.

Laboratory measurements of tracers and estimation of weathering indices
All sediment samples retrieved from the river bed were slowly air-dried and then manually disaggregated using a mortar and pestle. The samples were dry sieved using a < 63 μm sieve prior to laboratory analysis in order to compare a similar particle size fraction of all samples (Palazón and Navas, 2017;Pulley et al., 2015;Tiecher et al., 2018). Selection of this size fraction is commonly applied to help reduce the potential bias related to the effects of grain size (Collins et al., 1997;D'Haen et al., 2012;Dirszowsky, 2004;Fu et al., 2008;Manjoro et al., 2017;Nosrati et al., 2011Nosrati et al., , 2014Peart and Walling, 1986;Tiecher et al., 2018;Walling et al., 1999), which could directly distort comparison of the sub-basin source and target sediment samples (Nosrati et al., 2018a). Sieving and weighing of the resulting fractions of the sediment samples collected from both the tributaries and the downstream main outlet confirmed it was appropriate to use the < 63 μm fraction.
In order to measure the concentration of geochemical tracers including Ag, Ar, As, Ba, Be, Ca, Cd, Ce, Co, Cr, Cu, Fe, K, La, Li, Mg, Mn, Mo, Na, Ni, P, Pb, S, Sb, Sc, Sr, Th, Ti, U, V, Y, Ya, Zn and Zr using an Agilent (Varian) 735 Inductively Coupled Plasma Optical Emission Spectrometer (ICP-OES), 0.25 g of each sediment sample (< 63 μm) was digested with hydrofluoric acid (HF) and aqua regia (HCl-HNO 3 ; 3:1) at 220°C for 4 h in a hot box. The concentrations were expressed in milligrams per kilogram. The results showed that the analytical error for each element was < 5%. In addition to the 34 elements, several elemental ratios (Al/Ti, Cu/Pb, Fe/Mn, Fe/P, K/Na, Pb/Ni, Al/(K + Na + Ca + Mg), the Chemical Index of Alteration (CIA), the Weathering Index of Parker (WIP), the Chemical Index of Weathering (CIW), the Plagioclase Index of Alteration (PIA), and a recycling ratio (the ratio between the CIA and the WIP) was also computed for potential use in the composite signatures. Table 1 presents a summary of the laboratory results, elemental ratios and weathering indices.
The weathering indices were computed based on the molecular proportion of associated elements. Here, the concentration of elements (in ppm) was converted to elemental oxides based on each element compound ratio (Guo et al., 2018;Motha et al., 2003;Olley and Caitcheon, 2000). The Chemical Index of Alteration (CIA), originally proposed by Nesbitt and Young (1982) and applied by numerous studies (e.g., Guo et al., 2018;Shao et al., 2012), is potentially useful as a tracer since it provides a basis for identifying chemical changes caused by weathering of alumino-silicate minerals (Haddadchi et al., 2013;Motha et al., 2003;Owens et al., 2016). It is defined as Guo et al., 2018;Olley and Caitcheon, 2000;Shao et al., 2012). A CIA value of approximately 45-55 shows a lack of weathering, whereas a value of 100 indicates intense weathering with the complete removal of alkali and alkaline earth elements from the parent material (Motha et al., 2003;Shao et al., 2012). In general, higher CIA values suggest more weathering of silicates; however, it is difficult to make a simple linear relationship between weathering and CIA values (Shao et al., 2012). The WIP was proposed by Parker (1970) to evaluate the weathering intensity of silicate rocks based on the ratios of alkali and alkaline earth elements (Price and Velbel, 2003), and it can be defined as MgO/0.9 2 O/0.25 CaO/0.7) 100 2 2 (Guo et al., 2018;Shao et al., 2012). According to the definition of WIP, and contrary to CIA values, smaller WIP values are indicative of stronger chemical weathering (Guo et al., 2017). This index is most suitable for weathering profiles on heterogeneous parent rocks but potentially less suitable for highly weathering coatings, because its formulation is based on only highly mobile alkali and alkaline elements (Hamdan and Bumham, 1996;Price and Velbel, 2003). In the case of the mobility of aluminum during chemical washing, some studies have shown that WIP may be more suitable for characterizing weathering intensity than CIA, although both indicators are well correlated. However, to date, WIP has been rarely used to measure the weathering intensity in different stream sediment samples from around the world, whereas CIA has been widely documented for drainage basins (Shao et al., 2012).
CIA-WIP plots can be used to discriminate the modification of sediment compositions due to weathering induced by recycling over longer timescales. Thus, with an increase in sediment recycling, the value of WIP is linearly reduced, but CIA is largely independent of this (Guo et al., 2017). The effects of chemical recycling on river sediment chemistry should be carefully considered in order to provide stronger constraints on the interpretation of sediment-associated geochemical profiles (Guo et al., 2018).
The Chemical Index of Weathering (CIW) developed by Harnois (1988) is defined as . This index is equivalent to the CIA, except that the potassium was eliminated from the CIA equation. Like the CIA, the CIW is also essentially a measure of the extent of conversion of feldspars to clays. Because the CIW does not account for the aluminum associated with K-feldspar, it may yield very high values for K-feldspar rich rocks, whether they are chemically weathered or not. Fedo et al. (1995) proposed the Plagioclase Index of Alteration (PIA) as an alternative to the CIW. Because plagioclase is abundant in silicate rocks and dissolves relatively rapidly, the PIA may be used when plagioclase weathering needs to be taken into consideration (Price and Velbel, 2003).

Tracer conservation tests
Sediment tracers should exhibit conservative behavior during sediment detachment and transport through a catchment (Haddadchi et al., 2015;Lamba et al., 2015). Accordingly, the first step in addressing this requirement is removal of non-conservative tracers from further analysis. Non-conservative tracers were identified and eliminated using three methods (Haddadchi et al., 2015;Nosrati et al., 2018b). Initially, the standard bracket or range test was performed to ensure that the tracer concentrations of the individual main stem target sediment samples (which were collected from the outlet of the main basin) were within the concentration range of the sub-basin source groups (Mukundan et al., 2010;Navratil et al., 2012;Nosrati et al., 2018b;Tiecher et al., 2018). At the next stage, the mean concentration of each tracer in each main stem target sediment sample was compared with the corresponding mean concentration range in the sub-basin samples (Haddadchi et al., 2015;Wilkinson et al., 2013). Again, the tracers that were not in the mean concentration range of the sub-basin spatial source bed sediment samples were removed from further analysis. These range tests do not confirm the complete absence of changes to tracer properties, but instead provide an initial screening for the removal of tracers that show significant alteration between upstream sampled sources and downstream sites for the collection of target sediment samples. Lastly, biplots of tracer pairings selected in the final composite signatures were used to assess the degree to which the relationships between tracers judged to be generally conservative, by the first two tests, were consistent between upstream and downstream bed sediment sampling sites.

Discrimination of the tributary sub-basin spatial sediment sources
Applying statistical tests is one widely used means of identifying sets of tracers that can discriminate robustly between potential sediment sources (Haddadchi et al., 2014;Nosrati et al., 2011). Three statistical methods were used to provide composite fingerprints for discriminating the three sub-basin spatial sources: (1) a combination of KW-H and discriminant function analysis (DFA), (2) a combination of KW-H and data mining analysis, and (3) a combination of KW-H and principal components & classification analysis (PCCA).
KW-H or one-way ANOVA is a nonparametric test that compares more than two groups, and tests the null hypothesis that different groups in comparison comprise a similar distribution or distributions with equal means (Collins andWalling, 2002, 2007;Smith and Blake, 2014). Therefore, the KWeH test helps eliminate any tracer that does not show significant differences in the concentrations between at least two sediment sources (Pulley et al., 2015).
Discriminant function analysis (DFA) was performed based on minimizing the Wilks' lambda of those tracer properties that passed the KW-H test. DFA has been widely used to determine the final composite fingerprints that have the highest discrimination between sediment sources but with a minimum number of variables (Collins et al., 1998;Evrard et al., 2011;Lamba et al., 2015;Smith and Blake, 2014). Wilks' lambda is a multivariate statistical test that measures the significance of the discriminatory power of the model. Wilks' lambda values are between 0-1. Values close to 0 indicate high discriminatory power, while values close to 1 indicate poor discrimination (Hughes et al., 2009). To determine whether discriminant functions are statistically significant, various tests including eigenvalue, canonical correlation and Mahalanobis distance were used (Nosrati, 2017;Nosrati et al., 2018b).
Tracers exhibiting statistically significant differences between the potential sub-basin spatial sediment sources, using KWeH, were also entered into classification tree analysis. Classification tree analysis is one of the main techniques used in data mining (Lemon et al., 2003). Classification trees are used to predict membership of cases in the classes of a categorical dependent variable from their measurements on one or more predictor variables (i.e. tracers) and as such, this method has much in common with DFA. Classification tree analysis determines a set of logical if-then conditions (instead of the linear equations estimated in DFA) for classifying cases. Classification trees readily lend themselves to being displayed graphically, helping to make them easier to interpret. There are two key tools to interpret the results of classification tree analysis: the final trees and the predictor importance. The final graphical sequence of (boosted) trees can be a useful way to examine the significance of the final classification model. The predictor importance is computed as the relative (scaled) average value of the predictor statistic over all trees. The bar plot of the predictor importance provides insight on the variables that make the major contributions to the prediction of group membership. The fourth and final set of tracer properties was identified by applying a PCCA test to those tracers that passed the KWeH test. A comprehensive review of the principal component method is presented by Nosrati et al. (2018a). All statistical analyses were performed using STATISTICA V.8.0 (StatSoft, 2008).

Source apportionment using a sediment un-mixing model
The Modified MixSIR Bayesian un-mixing model (Nosrati et al., , 2018b quantifies the relative contributions of sediment from different sources by calculating probability distributions for the proportional contribution (f i ) of each source i to the downstream target sediment samples in three stages: 1) prior probability distributions for model parameters, 2) construction of a likelihood function for the statistical model, and 3) derivation of the posterior probability distributions for the parameters using the Bayes rule to adjust the prior distributions based on the observed data. The Bayes rule states that the posterior probability distribution for all f i is proportional to the prior probability distributions multiplied by the likelihood, and then dividing by their sum, viz.: where L(data|fq) is the likelihood of the data given f q , p(f q ) representing the prior probability being true, based on prior information, and f q is the proportional source contributions of q proposed vectors. The relative contributions of sediment are factored into the model by defining mean and variance parameters for each sediment source i and the final sets of tracers (composite fingerprints; j).

Results and discussion
3.1. Confirmation of conservative tracers and different composite fingerprints for discriminating the spatial sediment sources Table 1 compares the tracer concentrations in the sub-basin outlet (spatial sediment sources) and downstream target sediment samples. The results of the standard bracket test showed that Ag, Ce, Cu, La, Mn, Pb, Sb, Sc, U, V, Yb, Zn, Zr, Cu/Pb and Fe/Mn were non-conservative. In addition to the standard test, the results of comparing the sediment means with the corresponding sub-basin spatial source means showed that all remaining tracers, except Y and Fe/P, were conservative (Table 1). On the basis of these two tests, the 29 remaining tracers were retained and tested using the KW-H test. Table 1 also shows the results of applying the KW-H test which indicated that all tracers exhibited a statistically significant difference between the tributary sub-basin spatial sediment sources using the bed sediment samples collected from the field.
The 29 tracers (Al, As, Ba, Be, Ca, Cd, Co, Cr, Fe, K, Li, Mg, Mo, Na, Ni, P, S, Sr, Th, Ti, Al/Ti, K/Na, Pb/Ni, (K + Na + Ca + Mg)/ Al, CIA, WIP, Recycling, CIW and PIA) passing the KWeH test were entered into the stepwise DFA ( Table 2). The largest eigenvalue of the first function (61.1) corresponds to the eigenvector in the direction of the maximum spread of the groups' means. The Wilk's lambda value of the first function (0.001) indicated that 83% of the total variance among the sub-basin spatial source bed sediment samples was explained by these tracers. The canonical correlation value was 0.99 and indicated a strong correlation between the discriminant scores and the individual source groups (Table 3). The squared Mahalanobis distance showed that the spatial sediment sources were well separated by the shortlisted tracers (Table 3). The backward stepwise DFA yielded classification matrices assigning 100% of the cases (i.e., spatial source bed sediment samples) to the correct groups (Table 3).
Stepwise selection using Wilks' lambda indicated that a composite signature comprising seven tracers (Li, Al, Ba, P, Al/Ti, Na and Fe) provided significant discriminatory power on the basis of the DFA model ( Table 2). The results of different tests within DFA indicated that the discriminatory power of all of these tracers was perfect (Table 2). Partial Wilks' lambda is the Wilks' lambda for the unique contribution of the respective tracer to the discrimination between individual source groups. The smaller the Partial Wilks' lambda, the greater the contribution to the overall discrimination. The Partial Wilks' lambda values suggested that Al contributed the most, Li second most, P third most, Al/Ti fourth most, Fe fifth most, Ba sixth most and Na the least to the overall discrimination (Table 2). A scatterplot (Fig. 3) using the first and second discriminant functions calculated using backward DFA confirmed that the sediment samples collected from the outlets of the different sub-basin spatial sediment sources were well separated.
During the data mining stage of the tracer analysis, the properties passing KW-H were entered into general classification and regression tree models (GC&RT). Fig. 4a displays the final tree selected on the basis of the sub-basin properties (tracers) considering v-fold cross-validation (CV cost = .03) analysis. The final tree classified the sub-basin spatial sediment sources into three terminal nodes (i.e., the sampled sub-basins) using the sediment samples collected to characterise these potential sources. The histograms of cases in each class at the nodes showed that 100% of the cases were classified correctly (Fig. 4a).
The bar plot of predictor importance based on tributary sub-basin spatial source as the dependent variable and the tracer properties as the predictors is presented in Fig. 4b. For each tracer, predictor importance shows the relative average value of the sums-of-squares prediction over all trees and nodes, where the maximum value sums to 1. Hence, these values reflect the strength of the relationship between the predictors and the dependent variable of interest over the successive boosting steps. In this case, the tracers Ba, Co, Cr, Mg, Ni, and Pb/Ni (importance value = 1.0) stand out as the most important predictors within this particular composite signature (Fig. 4b). These results provide evidence that general classification and regression tree models can be used as a tool for identifying important dimensions in a set of tracers included within a composite signature and to identify those sediment sources with similar or dissimilar characteristics. Tracers passing the KWeH test (Al, As, Ba, Be, Ca, Cd, Co, Cr, Fe, K, Li, Mg, Mo, Na, Ni, P, S, Sr, Th, Ti, Al/Ti, K/Na, Pb/Ni, (K + Na + Ca + Mg)/Al, CIA, WIP, Recycling, CIW and PIA) were also tested using PCCA. Here, all the tracers in question were further explored as an alternative means of reducing the number of tracers and problems of multicollinearity. The results of PCCA showed that the first five principal components (PCs) with eigenvalues > 1 accounted for > 90% of the variability among the tracer values for the three spatial source groups (Table 4). The PC corresponding to the largest eigenvalue (16.0) accounted for approximately 55% of the total variance. The second PC corresponding to the second eigenvalue (5.9) accounted for approximately 20% of the total variance. The third PC corresponding to the third eigenvalue (1.9) accounted for approximately 6% of the total variance. The fourth PC corresponding to the fourth eigenvalue (1.1) accounted for approximately 4% of the total variance. The PC corresponding to the smallest selected eigenvalue (1.0) accounted for approximately 3.6% of the total variance (Table 4).
The highly-weighted tracers under PC1 with absolute values within 10% of the highest tracer (0.93 value for Be) loading (the loading of selected tracers should be larger than 0.84) were Al, K, Mg, Ni, Pb/Ni, Al/Ti, K/Na, (K + Na + Ca + Mg)/Al, CIA and CIW.  Only Be was retained for the final composite signature because these eleven tracers were strongly inter-correlated. Under PC2, the highly-weighted tracers with absolute values within 10% of the highest tracer (0.77 value for WIP) loading (the loading of selected tracers should be greater than 0.70) were Sr and Ti. Only WIP and Ti were retained for the final composite signature because WIP was strongly correlated with Sr (r = 0.92). Under PC3, the highly-weighted tracer (0.66 value for Li) with absolute values within 10% of the highest tracer loading (the loading of selected tracers should exceed 0.60) was Li. Under PC4, the highly-weighted tracers with absolute values within 10% of the highest tracer (0.47 value for P) loading (the loading of selected tracers should exceed 0.42) were P and Cr. Both were retained for the final composite signature because they were not strongly correlated. Under PC5, the highlyweighted tracer (0.63 value for S) with absolute values within 10% of the highest tracer loading (the loading of selected tracers should exceed 0.57) was S. PCs scores were calculated using the resulting component score coefficient matrix and tested for significant differences between the sediment sources using one-way ANOVA (F-test) and Tukey HSD post-hoc tests (P < 0.05) ( Table 4). The results showed that the PC scores for the first four PCs varied significantly with sediment sources. So the tracer associated to the PC5 (S) was excluded from the composite signature. These results thereby selected six tracers (Be, Fe, WIP, Li, P and Cr) as an alternative composite fingerprint on the basis of the PCCA model (Table 4). The plot of factor coordinates of variables for the first two PCs associated with the six tracers selected by PCCA is presented in Fig. 5a. This composite fingerprint clearly provided strong discrimination between the three potential sub-basin sediment sources (Fig. 5b).
For the tracers selected in the final composite signatures, the biplots of some tracer pairings for source and sediment samples were compared. The results confirmed that there was no major tracer transformations during sediment mobilization and delivery in the study area (Fig. 6).  (Table 5), the relative contributions (with corresponding uncertainty ranges) from the sub-basin spatial sources 1, 2 and 3 were estimated as 54.3% (47.8-62.0), 11.4% (4.2-18.7), and 34.3% (27.6-39.9), respectively. Using the alternative composite signature selected by a combination of KW-H and data mining (Table 5), the corresponding respective contributions and associated uncertainty ranges were estimated as 72.0% (61.6-82.7), 13.6% (9.0-18.5) and 14.2% (3.1-25.4). Finally, on the basis of the composite signature selected using a combination of KW-H and PCCA (Table 5), the relative contributions from the sub-basin spatial sources 1, 2 and 3 were estimated as 50.8% (42.8-59.9), 28.7% (20.2-37.3) and 20.3% (12.7-27.2), respectively. The root mean square difference (Table 5) between the estimated sediment contributions from the sub-basin spatial sources using the three different composite signatures ranged from 22.3% (sub-basin 3) to 47.4% (sub-basin 2). The predicted spatial source contributions were therefore highly sensitive to the composite fingerprint used in the Bayesian un-mixing model. Comparisons of the predicted and known relative contributions from the tributary sub-basin spatial sources 1, 2 and 3 using the three different composite signatures and the virtual sediment mixtures showed that the RMSE ranged between 0.6% and 35.8% and MAE between 0.2% and 11.9% (Table 6). The overall average RMSE and MAE for the modelled source predictions using the virtual mixtures were 10.6% and 9.6%, respectively. The RMSE results showed that 15 of 30 possible predicted values were < 10%, whilst in the case of MAE, 29 of the 30 values were < 10%. These error levels were judged to be acceptable in the context of those reported elsewhere in the international literature (e.g., Haddadchi et al., 2013;Nosrati and Collins, 2019;Palazón et al., 2015).

Sediment source contributions
The weathering indices included in the initial potential tracer datasets reflect the intensity of chemical weathering processes and the degree of elemental mobility resulting from depletion of sediment samples in mobile as compared with immobile components (Nadlonek and Bojakowska, 2018;Price and Velbel, 2003). Only WIP was selected in one of the final composite signatures for discriminating the three tributary sub-basin spatial sources (Tables 5 and 6), although the results of the KW-H in Table 1 show that the indices were individually able to discriminate between the three sub-basin spatial sources. The WIP values in Table 1 suggest that the degree of chemical weathering reflected in the bed sediment samples collected to characterise sub-basins 2 and 3 is marginally greater than that for the equivalent samples collected from sub-basin 1. The estimates of WIP shown in Table 1 are considerably higher than those reported by some recent studies in the international literature (e.g., Nadlonek and Bojakowska, 2018) thereby suggesting comparatively lower chemical weathering intensity in the study area. The low values for CIA in Table 1 are also suggestive of a lack of intensive weathering and chemical alteration in the study area and reflect the local arid climate (Nadlonek and Bojakowska, 2018). Weathering index values reflect the complex interplay between climate, lithology, tectonism, topography, vegetation cover and human activities (Gibbs, 1970;Meybeck, 1987;Oliva et al., 2003;Shao et al., 2012). On this basis, they potentially offer a physically-grounded basis for inclusion in tracer datasets for identifying statistically robust composite signatures for source discrimination. Here, however, scale is likely to be an important factor driving the magnitude of differences between weathering index values for soil or sediment samples collected to characterise different river basin sources. The source apportionment results based on the three composite signatures consistently suggested that tributary sub-basin 1 is the dominant source of the target sediment samples collected further downstream from the bed of the main stem. This sub-basin is the largest (168.3 km 2 ) of the spatial zones apportioned in this study and comprises larger areas of both dry-land farming (72 km 2 ) and rangelands (64 km 2 ) compared to the other two spatial sources. Locally, dry-land farming is characterized by cultivation using conventional tillage practices on steep slopes, which renders bare tilled soils susceptible to rain splash detachment and subsequent mobilization in surface runoff. Recent expansion of such farming in sub-basin 1 on the Quaternary deposits has exacerbated erosion problems in this zone of the study area. At present, water scarcity problems and limited availability of new land for conversion to agriculture, mean that land use change from rangeland to arable cropping under dry-land farming is common, especially in sub-basin 1. Conventional dry-land farming practices are widespread, including intensive tillage with ploughing and heavy harrowing plus the routine removal of crop residues by burning to reduce biomass volume and to facilitate mechanical operations which are the primary factor leading to increased soil erosion and sediment production on the steep slopes (Jin et al., 2007). It should be noted however, that flood water irrigation in riparian areas in the lower and middle portions of sub-basin 1, which is assisted by siltation encouraging more overbank inundation, is used to support crop production on low lying riparian plots.
The source apportionment results presented herein should, however, be interpreted as a tier 1 assessment of the sources of downstream sediment issues. In this regard, they provide a basis for targeting follow-up work for confirming the relative losses from dry-land and rangeland farming practices, especially in sub-basin 1, to help engage local farmers and select and target appropriate mitigation measures for soil tillage or grazing management. Here, further work could also usefully assess the relative age of sampled sediment (e.g. using the ratios of fallout radionuclides with shorter and longer half-lives) to elucidate the likely temporal duration of sediment storage and remobilization in the study catchment and to help link the sediment source tracing results to oral cropping histories developed with local farmers.

Conclusions
A spatial sediment source fingerprinting exercise has been used to provide a tier 1 screening of sediment provenance in a mountainous agricultural catchment in western Iran. Three composite signatures were selected using different statistical tests, all of which suggested that sub-basin 1, draining the upper portions of the study area, dominates source contributions to the target sediment samples collected on the riverbed at the catchment outlet. The consistency of the signatures in all identifying sub-basin 1 as the dominant spatial sediment source lends weight to the results, since it is established that different composite signatures have the potential to generate contrasting sediment source estimates using mixing models (Collins et al., 2017b). The inclusion of weathering indices in initial tracer sets from which composite signatures are selected warrants further evaluation. Regardless of their inclusion or otherwise in final composite signatures, weathering indices potentially offer useful information for helping pre-selection of tracers (e.g. mineralogical, elemental) and for interpreting tracer conservation test results in the context of climate, lithology and hydrogeomorphological processes in river basins. With regards tracer conservation, longitudinal sediment sampling in large drainage basins would permit the weathering indices to be used to explore any potential evolution of tracer transformation risks across scales. Weathering indices offer potential for ensuring composite signatures have a physical, rather than a purely statistical basis, in the context of study area lithology and the requirement to discriminate individual spatial sources across landscapes. Follow-up investigations are required to confirm the relative contributions of the dry-land and rangeland farming, and of channel banks, in the study area to downstream sediment loads. Nonetheless, the evidence generated by this study will be used to help engage land owners in the need for regenerative farming practices.

Acknowledgements
This article is an output of a Geomorphology Ph.D. thesis from Shahid Beheshti University in Tehran, entitled 'Spatial and temporal analysis of sediment fingerprinting using statistical methods and an un-mixing model in the Kamish catchment, Kermanshah Province, Iran'. The thesis was approved and funded by the specialized committee of the Iran National Science Foundation (INSF) (grant number 96015282). The contribution of ALC to this manuscript was funded by the UK Biotechnology and Biological Sciences Research Council (BBSRC) via the Soil to Nutrition institute strategic programme (grant award BBS/E/C/000I0330). a The values in parentheses show the uncertainty ranges (90% confidence limits: 5%-95%).

Table 6
Comparison of the predicted and known relative contributions from the sediment sources to the virtual sediment mixtures using the composite signatures selected by different statistical approaches and the corresponding root mean square error (RMSE) and mean absolute error (MAE).