Our study area within the Dohuk governorate is located within the northwestern segment of the Zagros-fold thrust belt (ZFTB), a mountain belt that extends from southern Iran NW-ward to the Kurdistan Region of Iraq and SE Turkey. The ZFTB resulted from the ongoing convergence between the Arabian and Eurasian plates . The convergence started in the late Cretaceous with the subduction of the Neotethys oceanic crust beneath the Eurasia Plate and the obduction of the ophiolite sequences on Arabia's margin, followed by the subsequent continent-continent collision between the Arabian and Eurasian plates during the Oligocene-Early Miocene . Since the onset of continental deformation on the northeastern margin of the Arabian Plate (including the study area), it has propagated for 250–350 km . Within the external part of the ZFTB, these zones include the Imbricated Zone, the High Folded Zone, and the Foothill (Low Folded) Zone .

The study area covers parts of the High Folded and Foothill zones (Fig. ), where structures are mainly trending in a nearly E–W direction . The Bekhair/Ǧebel Bihair Anticline ( Bixêr) is the main structural and morphological feature in the area and plunges at the western end of our study area. It separates the Selevani Plain ( Selevani), which stretches from the northern bank of the Mosul Dam Lake to the anticline, in the south, from the Zakho-Cizre-Silopi Plain – later shorten into Zakho Plain – and Little Khabur Valley to the north .

The exposed rocks in the area include sedimentary units ranging in age from the Upper Cretaceous to the Pliocene . The Upper Cretaceous units consist of platform carbonates and siliciclastic rocks . The Paleocene-Eocene units consist mainly of marginal marine marls and shales that interfinger with rigid carbonate units, followed by red Eocene clays and carbonates. These Upper Cretaceous-Eocene units are exposed within the anticlinal structures in the area. The Oligocene units are missing in the area; thus, the Eocene carbonates underlie the Middle Miocene clays, evaporites, and limestones. The Upper Miocene–Pliocene units consist of fluvial sandy succession, clay, and conglomerate deposited in the Zagros foreland basin . The Miocene-Pliocene units are exposed within the synclines and low-elevation area to the north and south of Bekhair Anticline.

The Quaternary deposits cover three different environments of the study area (Fig. ). First, the flat area of the Selevani Plain and north of the Zakho Plain (Türkiye) is covered by residual clayey soil material, coming from the erosion of the Bekhair and Zagros anticlines. Second, along the riverbanks of the Tigris and Little Khabur rivers, sand and gravel-sized terrace deposits, as well as floodplain sediments of fine sand and clay, can be observed. Finally, Quaternary formations from alluvial fan sediments of clayey soil, combined with rock fragments coming from colluvial deposits, are visible in the foothills of the Bekhair and the shallow Little Khabur Valley. Sometimes, calcrete is also developed within the these Quaternary deposits .

The central part of the study area falls within a Csa (Hot-summer Mediterranean) agro-climatic zone, according to the Köppen Geiger classification . Annual precipitation ranges from 200 to 500 mm, with an average yearly temperature exceeding 16 °C . Only the Little Khabur Valley, located north of the Bekhair anticline, experiences slightly cooler winters and receives higher rainfall, typically between 500 and 800 mm per year . South of the Bekhair anticline, the Selevani Plain belongs to the Mesopotamian steppe floral complex, which supports a limited number of xerophytic shrubs and herbs, primarily Artemisia herba-alba mesopotamica often associated with Aristida plumosa (pp. 78–80; p. 183). In contrast, the northern region, encompassing the Zakho Plain and Little Khabur Valley, falls within the Kurdo-Zagrosian climate zone, characterised by a denser xerophilous deciduous steppe forest, driven by its higher elevation and more favourable climatic conditions (Fig. ; p. 68). Dominant shrubs include Anagyris foetida or Pistacia khinjuk are associated with trees as Quercus brantti, or Quercus boissieri, which grows between 800 and 1700 m of altitude. Historical records mention the presence of pine forests pp. 183–190, though they are likely no longer extant. In both the foothills of the Mesopotamian Plain and the Kurdo-Zagrosian space, cultivated Olivae europanis can be sporadically observed.

In the southern part of our study area, the Tigris floodplain and its Quaternary alluvial deposits have largely disappeared due to the construction of the Mosul Dam Lake . What remains are sporadic surface exposures of conglomerates and marls (Fig. ; ) and three to four terraces levels . North of the Tigris river, in the Selevani Plain, combined action of wind and irregular water action of wa‾dı‾s, have led to the formation of gullies on this depositional glacis , shaping a badland landscape (Fig. B). The Bekhair anticline and its imbricated zone form a structurally homogenous ridge dominated by exposed limestone and sandstone formations (Fig. , ), which are subject to lift-up process and tectonic action. The foothills on both sides of the ridge, however, are subject to wind, water erosion, and gravitational processes, resulting in extensive colluvial deposits (Fig. C; ). In the area of the Tswoq anticline and the Little Khabur Valley, the landscape is dominated by sandstone and conglomerate. Soil surface erosional process are less pronounced in the Little Khabur Valley region due to the protective effect of denser vegetation cover. The Zakho Plain, located within a synclinal structure (Fig. ), is a flat alluvial area, also less affected by erosion.

Soil mapping in the region was initially carried out in the 1950s and 1960s as part of the Iraq soil mapping project . We adapted Buringh's classification to the WRB system , improved spatial detail using modern satellite imagery (Sentinel 2 ; DEM ; and Bing Maps https://www.bing.com/maps, last access: 6 September 2024), and completed the map with unrecorded Regosols and Fluvisols. This resulted in a new regional soil map built on an expert-based knowledge method (Fig. ). The semi-arid climate, marked by sharp temperature variations and high subsurface CaCO₃ concentrations, has favoured the development of vertic and calcic features in many soil profiles . However, significant local variability exists. Soils adjacent to the Tigris River are typically calcic Vertisols, likely due to subsurface marl and conglomerate permeability. The Selevani Plain's glacis deposits are dominated by cambic and gypsic Calcisols (Figs.  and ), with mediumly developed soil horizons, and a soil depth of 100 to 200 cm. North of the Selevani Plain, on the structural ridge and the steep slopes of the Bekhair anticline, soil development is minimal, due to active erosion, resulting in nudilithic Leptosols. On the northern side of the ridge, the Little Khabur Valley and its surroundings are dominated by poorly developed soil such as calcic Cambisols, Regosols and Leptosols, shaped by steep slopes and more erodible parent materials (conglomerate and sandstone), compared to the Simiele Plain. In contrast, the flat, irrigated Zakho alluvial plain, with higher precipitation, supports more developed soils, such as calcic isomeric Kastanozems (Fig. ). Finally, Fluvisols occur sporadically along the Tigris and Little Khabur floodplains and major wa‾dı‾s channels riverbanks (Figs.  and ), identifiable by their ochric and/or umbric horizons.

Sampling design plays a critical role in ensuring that selected locations reflect the spatial and environmental variability of the study area . We adopted a conditioned Latin hypercube sampling (cLHS) approach, a method particularly suited to digital soil mapping applications . The cLHS method ensures that sampling points are distributed across the full range of values in selected environmental covariates by stratifying feature layers. For each covariate, a set of marginal strata is defined (intervals). The number of marginal strata for each covariate equals the sample size, resulting in pn a total of marginal strata, where p is the number of covariates and n is the sample . Sampling was performed with R 4.4.0 using the clhs package .

We selected six covariates (Table ) which represent a broad range of parameters influencing soil variability. These included physical characteristics, underlying geomorphological formations , potential soil properties and erosion process.

The potential soil properties layer was constructed using spectral indexes (clay minerals, ferrous minerals, rock outcrop, carbonate) derived from climatic and satellite datasets . Erosion risk was modelled using the Revised Universal Soil Loss Equation (RUSLE; ), incorporating five key factor: soil erodibility (K), soil coverage (C), topographic effect (LS), rainfall-runoff (R) and erosion control practices (P; ). In our RUSLE model (Table ), we set K as the soil strength factor based on texture and organic carbon values , and C was set with values from and . Slope length and steepness factor LS were based on Desmet and Govers' method (), while the R factor from was used. Finally, the conservation factor P, which could not be observed, has been set to 1 artificially, as suggested by .

At each site, samples were collected for the top 50 cm using a 3.5 cm ø auger and from depths up to 100 cm with a 2 cm ø auger. The different soil horizons' depths were measured, and colour was determined according to the Munsell soil colour chart. Samples were collected at five depth increments: 0–10, 10–30, 30–50, 50–70 and 70–100 cm. Bulk density was calculated for the topsoil using a 5.3 cm ø ring . After removing the surface litter and loose sand, the sampling ring was used on the 0–10 cm soil layer. All samples were air-dried at 40 °C for 48 h before sieving at 2 mm for subsequent analysis.

Mid-infrared spectroscopy to measure physical and chemical soil properties has significantly evolved over the past decades and offers reliable results while saving time and resources . All the 561 soils samples were ground under 1 µm with a Pulverisette 5/4, classic line (Fritsh, Idar-Oberstein, Germany) before being pressed into a tablet, mixing 1–1.3 mg of soil and 250 mg of potassium bromide (KBr). The spectra were analysed with a Vertex 80v (Bruker OPTIK GmbH, Germany), with a 4 cm⁻¹ resolution, on the 375–4500 cm⁻¹ interval.

The spectra were imported into R 4.4.0 and analysed using the prospectr and simplerspec packages. To reduce noise interference, we decided to remove the measurements between 375–499 and 2451–2500 cm⁻¹ intervals and spectra value higher than 2 and lower than -2 . One sample was therefore removed due to its low values (< -2). The remaining 560 soil spectra were enhanced by applying three spectral transformations , Savitzky-Golay with a polynomial order of 2 and a window size of 11 (SG 2.11), a moving average of 11 and standard normal variate transformation on the SG transformed spectra (SNV-SG). A total of 108 samples were selected for laboratory measurements (Table ) using Kennard-Stone sampling , ensuring a high diversity and variability of individuals, based on their spectral data .

On these 108 selected samples, seven properties were measured: pH, CaCO₃, N_t, C_t, OC, EC and texture. The pH was measured using a potassium chloride (KCl) solution, with a ProfiLine pH 3310 and a WTW SenTix 81 pH electrode (Fisher Scientific, Strasbourg, France). Carbonate calcium (CaCO₃) content was determined as a percentage using a calcimeter 08.33 (Royal Eijkelkamp, Giesbeek, Netherlands). Total nitrogen (N_t), total carbon (C_t) and total organic carbon (OC) were quantified as percentages with a CNS analyser, Vario EL III (Elementar, Hanau, Germany). The electro-conductivity (EC) was measured in micro-siemens per centimeter (µS cm⁻¹) using a Cond 330i/340i (WTW, Weilheim in Oberbayern, Germany). Texture property was determined as a percentage and measured through wet sieving for sand fraction and a SediGraph III for finer fractions (Micromeritics, Norcross, USA). Additionally, we estimated the mean weight diameter in mm (MWD, Eq. ) based on the texture results. 2MWD=Xi×Wi/100 where Xi is the mean diameter of i size fraction, and Wi is the weight percentage of i size fraction.

The Cubist model is a regression-based machine learning algorithm that extends the ideas of decision trees by combining rule-based predictive models with linear models at the leaves, enhancing both interpretability and predictive accuracy . This model excels at handling both continuous and categorical data, providing robust predictions even in the presence of complex interactions and non-linear relationships . Cubist’s strength lies in its ability to partition the data space and fit separate linear models to each segment, making it particularly effective for problems with distinct patterns or heteroscedasticity . This model has been applied in a variety of studies for soil property prediction from spectral predictors, such as , , and . We tested a Cubist regression model on the four spectral datasets from each of our 560 samples in a Python environment using the Cubist library . The model used a stratified 5-folds cross-validation and a tuned grid (n rules: 20, 30, 40 and n committees: 5, 10, 15).

We based our soil property model on the soil formation factors of the scorpan equation (Eq. ) developed by . We included 85 covariates (Tables  and ). The remote sensing variables were accessed through an API of Google Earth Engine (https://earthengine.google.com, last access: 15 January 2026) on Python, via the ee library , and the different indices computed in R with the terra package . The terrain variables were computed on SAGA GIS 9.3.1 based on a filled and filtered DEM from GLO-30 . All the computation was realised under R 4.4.0 environment . We included only the 2022 and 2023 samples to produce the DSM for two reasons. First, these campaigns followed a cLHS sampling strategy (cf. Sect. ), whereas the 2017–2018 campaigns did not; including them would have reduced the consistency of the sampling design. Second, the 2017–2018 samples lacked depth information and were primarily limited to topsoil. To ensure data comparability, these earlier samples were therefore excluded, resulting in a dataset of 531 samples from 122 sites (Table ). These samples were included in the DSM as input with: 122 samples for the 0–10 cm depth, 111 for the 10–30 cm increment, 108 for the 30–50 cm depth, 98 for the 50–70 cm increment and 92 for the 70–100 cm depth. We divided the mapping of each variable for each soil depth increment, resulting in 45 models and 50 maps in total (the three texture variables only include two alr models). We performed a standardisation of the predicted textures values from the Cubist model, with TT.normalise.sum function and an additive-log ratio transformation with the alr function . This transformation preserved the spatial information of the prediction with a repartition close to a normal distribution . Digital soil mapping have adapted this additive-log ratio on the texture with success, alr_sand = ln(sand/clay) and alr_silt = ln(silt/clay) . Once the models were performed, the additive-log ratio was reversed into the three texture with the alrInv function, before being evaluated.

During the pre-processing, we performed a feature selection with the Boruta package . Using a random forest-based model, Boruta validated or rejected the selection of variables regarding their influence on the inputs (Fig. ). This method improves model accuracy and reduces overfitting results , and its efficiency has been proven for digital soil mapping . We also performed a recursive feature elimination (RFE; ) on the covariates with the caret package . The results were more conservative with the number of covariates selected (> 60 for each variables), longer in time computing capacities (800 %), and provided lower accuracy scores compared to Boruta selection, for the tested 0–10 cm depth increment.

For each soil depth, we used a 10 k-fold cross-validation repeated 3 times to tune the model and choose the final settings. This resampling strategy allowed us to avoid potential overfitting due to the small size of our training data set (< 100). We trained the models on a specific random forest algorithm called “quantile regression forest model” (QRF), which has shown good performance for digital soil mapping . The model was implemented with the caret , and quantregForest packages. We tuned the mtry hyperparameters (1:n, by = 1; n beeing the number of covariates), which corresponds to the number of covariates randomly sampled as candidates at each split, and the minimum node size hyperparameter nodesize (5:31, by = 5, ), which defines the minimum number of samples required to be at a leaf node. The number of trees was set as default at 500 . Best hyperparameters were selected based on the lowest RMSE (Table ), before beeing applied to produce a final model with the whole training dataset.

Regression trees use a tree-based structure, splitting the data into different nodes. In the end, the model evaluates the leaves and selects those with the best performance. Their specificity in regression is to predict continuous values at the terminal nodes rather than classes, unlike classification trees. Random forests build upon this principle by combining many regression trees grown on bootstrapped samples of the data, which improves prediction performance and stability. Based on this random forest framework, the quantile regression forest (QRF) , tracks each sample’s value at each node, providing a conditional response distribution. This allows the model to produce prediction intervals and to assess accuracy through quantiles .

To predict soil depth, we developed a prediction model using remote sensing data and ground-truth control points, collected during surveys to estimate soil depth. By soil depth, we defined the height of the active horizons in the pedogenesis process p. 225, the so-called solum. The measurement is taken from the top of the solum downward to the mineral layer (C horizon) or consolidated rock (R horizon). The soil depth was measured from 0 to 100 cm on the 122 sampling sites; we added 25 zero values (negative sites) from remote sensing imagery observation on bare rock points. Soil depth is mainly determined by climate, terrain, parent material, vegetation, and land uses . Consequently, we used 26 environmental covariates to predict the soil depth (Table ). Original soil depth data were first square root-transformed, similar to the DSM, we performed the training on a 10 k-fold cross-validation repeated three times. A quantile regression forest model was chosen and implemented in the R 4.4.0 environment using the caret and quantregForest packages. As described above (cf. Sect. ), the QRF model is fitted for digital soil mapping. The tuning of the hyperparameters was similar to the DSM with a tuned grid for mtry, and nodesize, while the number of trees was leave by default at 500. The optimisation of the hyperparameters (Table ), and training of the final model followed the same procedure as the DSM.

To validate the performance of our models, we evaluated prediction accuracy on for both spectroscopy predictions , and DSMs . Prediction accuracy was assessed using complementary metrics capturing systematic error, error magnitude, and the strength of the relationship between predicted and observed values. Bias, was expressed as the mean error (ME, Eq. ), used to quantify systematic over, or underestimation by the model. Values closer to zero indicate a lower systematic bias in the predictions. The root mean square error (RMSE, Eq. ) was selected as the primary error metric, as it quantifies the average magnitude of prediction errors and indicates how far model predictions deviate from the observed values. To characterize the extent to which the model explains the variability of the observed data, we reported the coefficient of determination (R2, Eq. ), which reflects the proportion of variance in the observations explained by the predictions. An R2 value of 1 indicates that the model explains 100 % of the observed variability. For spectroscopy-based predictions, we additionally report the ratio of performance to interquartile range (RPIQ, Eq. ), which normalizes the prediction error by the interquartile range of the observed values. This metric is commonly used in infrared spectroscopy to facilitate comparisons across skewed response variables, this enhanced the reliability of spectroscopic prediction models . Spatial models, such as DSMs, are prone to various sources of error, which should be quantified and represented spatially as prediction uncertainty . We evaluated the prediction uncertainty, for the quantile regression forest models of the DSM and soil depth, using the prediction interval coverage probability (PICP, Eq. ). The PICP correspond to the probability that all observed values fall within their prediction intervals (; ), here set at the 90 % level. Values of PICP under 0.9 suggest an overestimation of the uncertainty whereas a PICP above 0.9 indicates it was underestimated . However, recent research in DSM has shown that PICP alone does not account for the distribution of the predictions inside the interval and can also be bias due to the PI boundaries . Other metrics can overcome these issue by capturing the full coverage of the reliability of a regression model such as quantile coverage probability (QCP) and probability integral transform (PIT). Their implementation, while simple, does rely on an independent test set and were limited due to our workflow based on the caret package. These limitations of the PICP have to be taken into account while reading at the quantification of the uncertainty. 3ME=1n∑i=nn(y^i-yi) where n is the number of observations, yi the observed value for i, and y^i the predicted value for i. 4RMSE=∑i=1n(yi-y^i)2n where n is the number of observations, yi the observed value for i, and y^i the predicted value for i. 5R2=1-∑i=1n(yi-y^i)2∑i=1n(yi-y‾)2 where yi the observed value for i, y^i the predicted value for i, and y‾ is the mean of observed values. 6RPIQ=(Q3-Q1)/RMSE7PICP=1vcountji=PLL≤obsi≤PLiU where v is the number of observations, obsi is the observed value for i, PLiL is the lower prediction limit for i, and PLiU is the upper prediction limit for i.