Estimation of soil properties with mid-infrared soil spectroscopy across yam production landscapes in West Africa

across yam production landscapes in West Africa Philipp Baumann1, Juhwan Lee2, Emmanuel Frossard3, Laurie Paule Schönholzer3, Lucien Diby4, Valérie Kouamé Hgaza5, 6, Delwende Innocent Kiba3, 7, Andrew Sila8, Keith Sheperd8, and Johan Six1 1Group of Sustainable Agroecosystems, Institute of Agricultural Sciences, ETH Zurich, 8092 Zürich, Switzerland 2Department of Smart Agro-industry, Gyeongsang National University, Jinju, 52725, Republic of Korea 3Group of Plant Nutrition, Institute of Agricultural Sciences, ETH Zurich, 8315 Lindau, Switzerland 4World Agroforestry Centre (ICRAF), Côte d’Ivoire Country Programme, BP 2823 Abidjan, Ivory Coast 5Centre Suisse de Recherches Scientifiques en Côte d’ivoire, 01 BP 1303 Abidjan, Ivory Coast 6Département d’Agrophysiologie des Plantes, Université Peleforo Gon Coulibaly, BP 1328 Korhogo, Ivory Coast 7Institut de l’Environnement et Recherches Agricoles, 01 BP 476 Ouagadougou, Burkina Faso 8Land Health Decisions, World Agroforestry Centre (ICRAF), Nairobi, Kenya Correspondence: Philipp Baumann (baumann-philipp@protonmail.com)

has been a trend of shortened fallow periods in the cropping areas of West Africa over the last decades, which has further exacerbated the decline in soil fertility across the yam belt. Traditionally, yam is grown without external input in these areas.
Therefore, the production of yam and other crops grown in the region depends on soil organic matter (SOM) status (Padwick, 1983), which serves as a main pool of plant-available nutrients and provides cation exchange surfaces for soil nutrients (Syers et al., 1970;Soares and Alleoni, 2008). A particularly strong positive relationship between high organic matter stocks and yam 25 productivity is reported after fallow and when no fertilizer is added (Diby et al., 2009;Kassi et al., 2017). Thus, maintaining or increasing SOM and available nutrient levels is of utmost importance for sustainable production of yam and other crops in West Africa (Carsky et al., 2010). Furthermore, linking soil properties and yam yields (Frossard et al., 2017) and accounting for soil macro-and micronutrient status (O'Sullivan and Jenner, 2006) is fundamental to improving crop yields and soil management strategies. 30 Soil fertility is an integrative measure of soil attributes and their interactions that support the long-term agricultural production potential. Soil fertility is commonly decomposed into the physical, chemical and biological major components (Abbott and Murphy, 2007). Here, it is important to interpret soil fertility in the form of soil conditions and functions at an adequate resolution over time and space, and in relation to the crop of interest. For yam, low tuber yields are often attributed to an unbalanced ratio of essential nutrients (i.e. N, P, K) available in the soil (Enyi, 1972) and a fast mineralization and hence depletion of 35 organic matter (Carsky et al., 2010;Hgaza et al., 2011). Yet, the relationship between soil properties and tuber yield is not fully understood (Frossard et al., 2017). The reason is that the response of yam to mineral fertilization is highly variable because of confounding environmental and management variables, such as climate, soil type, inherent soil fertility, micronutrient deficiencies, tillage, seed tuber quality, planting date and density, staking and disease pressure across the yam belt (Kang and Wilson, 1981;O'Sullivan and Jenner, 2006;Cornet et al., 2016;Enesi et al., 2018). Further, there are no soil fertility recommendations 40 specific for yam under West African conditions. For this reason, establishing yam field trials designed with different organic and mineral fertilization strategies within different yam growing regions is required to optimize yam fertilization targeting regional soil and environmental conditions (Frossard et al., 2017). Despite the importance of soil fertility, it is challenging to quantify soil measures at sufficient temporal and spatial resolution to relate them to yam productivity together with other management effects. 45 In order to quickly assess key soil properties, such as soil organic carbon (SOC) and cation exchange capacity (CEC), we need more cost-and time-efficient methods in addition to the traditional wet chemistry laboratory analyses that are often cost-intensive and time consuming. Proximal sensing is a method that can provide reliable soil measurements rapidly and inexpensively (UNEP, 2012). Soil visible and near infrared (vis-NIR), and mid-infrared (mid-IR) diffuse reflectance spectroscopy has gained popularity over the past 30 years to assess soil properties in a complementary manner to conventional laboratory 50 analytical methods (Nocita et al., 2015). For model development and calibration but importantly also for validation purposes, soil IR spectroscopy requires laboratory reference analysis data. Previous studies have shown successful spectroscopic predictions of soil properties, such as organic C, texture, cation exchange capacity (CEC), and exchangeable K (Viscarra Rossel et al., 2006;Cécillon et al., 2009;Nocita et al., 2015;Sila et al., 2016). Many soil chemical and physical properties, such as soil mineralogy, the concentration, forms and distribution of SOM, are closely associated with IR spectral diversity. However, for a 55 range of extraction-based soil methods, the predictive capability seems variable. This can be because complex surface chemical processes that are not directly related to soil organic matter are involved and/or insufficient data densities are available at local scale to represent such locally complex relationships (Viscarra Rossel et al., 2006;Abdi et al., 2012;Sanderman et al., 2020).
Further, a library that includes a broad range of soil biophysical conditions found in the region in which it is used needs to be established. Depending on the study scale -field (e.g., Cambou et al., 2016), region, country (e.g., Clairotte et al., 2016), con-60 tinent (e.g., Sila et al., 2016)), world (e.g., Viscarra Rossel et al., 2016) -various statistical predictive modeling strategies are typically employed to account for regional variability in soil properties and determine empirical relationships between spectra and soil attributes. However, particular regions in spectra are characteristic for functional groups of soil components and thus, elucidating spectral features that are important for the prediction of a particular soil attribute helps to understand and validate the mechanisms based on which the empirically models predict the soil properties.

65
In this work, we aim to develop mid-IR spectroscopy as a diagnostic tool for key analytical soil variables within four climatically, ecologically, and agriculturally distinct landscapes in Burkina Faso and Ivory Coast. For yam and other cash crops, there is a lack of soil diagnostic tools to identify factors limiting yields and to derive site-specific fertilizer recommendations within and across landscapes. In these regions, yam has substantial economic importance for small-holder farmers. As land management and soil status is a key factor not only for yam but also other high-value crops in the region, quick and cost-70 effective soil status assessments should be transferable to other crops with similar nutrient demands. Thus, the main objectives of this study are to (1) develop and evaluate openly accessible and re-usable mid-IR spectroscopic models to estimate soil properties for selected landscapes representing major soil and climatic conditions in the West African yam belt, (2) to determine important spectral features for specific soil properties, and (3) to build a new soil spectral library in four landscapes of the West African yam belt for soil prediction and assessment. Finally, we make specific recommendations on whether and how specific 75 mid-infrared diagnostic measures are applicable for different soil management and screening purposes. We also discuss the spectroscopic evaluation for the soil's capacity to retain and release nutrients for sustained and improved cropping in the region.

Landscapes and soil sampling 80
Our study area covered the climatic and soil biophysical conditions representative of the West African yam belt. We selected four landscapes, two in Ivory Coast and two in Burkina Faso. Each landscape (approximately 10 km x 10 km) represents a diverse geographic ecoregion. The landscapes cover a gradient between humid forest and the northern Guinean savannah.
Specifically, the landscape Liliyo in Ivory Coast is at 5.88 • N and in the humid forest zone. The predominant soil type is Ferralsol (FAO, 2014). The landscape Tieningboué in Ivory Coast is at 8.14 • N and belongs to the forest savannah transitional 85 zone. The soils are dominated by Nitisols and Lixisols (FAO, 2014). The landscape Midebdo is at 9.97 • N and in the sub-humid savannah of Burkina Faso. Its dominant soil types include Lixisols, Gleysols, and Leptosols (FAO, 2014). The landscape Léo is at 11.07 • N and in the northern Guinean savannah of Burkina Faso and has Lixisols and Vertisols as the dominant soil type (FAO, 2014). The mean annual rainfall were approximately 1300 mm in Liliyo, and 900 mm in Tiéingboué, Midebdo, and Léo.
During July and August 2016, we sampled the soil from a total of 80 fields under yam cultivation across the four landscapes, 90 i.e. 20 yam fields in each landscape. The fields were selected in advance by taking into account visual variation in soil color and texture across the landscape. The yam fields selected contained the maximum soil variability based on soil colour and cropping history, taking into account both local farmers' knowledge on soil fertility and agronomic extension expertise. Yam is typically planted on soil mounds, ranging from 5000 to 10000 mounds per hectare with a single yam plant per mound. Within each field, we sampled the soil at four adjacent mounds in square arrangement, which were spaced between 0.5 and 2 m. At 95 each mound, 6 to 8 auger cores (25 mm in diameter) to the 0.3 m depth were taken at a radius between 0.15 and 0.3 m away from the center of a mound, depending on the size of the mounds. Then the soils from the four mounds were combined into one composite sample per field (around 500 to 1000 g of soil).
An additional set of 14 composite soil samples was collected by the International Center for Research in Agroforestry (ICRAF) at Liliyo from one sentinel site called "Petit-Bouaké" (UNEP, 2012). Sampling took place between 25 and 29 August, 100 2015 at positions that were previously selected for the Land Degradation Surveillance Framework (LDSF) in a spatially stratified manner (Vagen et al., 2010). The soil samples received from ICRAF were within the same landscape as the sampled soils in Liliyo within YAMSYS, but sampled from different positions. All soil samples were air-dried and stored in plastic bags until further analysis.

105
The air-dried soil samples were crushed and sieved at 2 mm. About 60 to 70 g of the sieved soil was oven-dried at 60 • C for 24 hours, of which 20 g were ball-milled. All chemical analyses except soil pH were conducted both on the soils sampled in yam fields (n = 80) and the LDSF soils obtained from ICRAF (n = 14).
The milled soils were analyzed for total C and macronutrient (N and S) concentrations using an elemental analyzer (vario PYRO cube, Elementar Analysensysteme GmbH, Germany). For each of the four landscapes, two soils were selected and ana-110 lyzed based on three analytical replicates for quantifying within-sample variance of the elemental analysis. For the remaining samples, the analysis was not repeated. Sulfanilamide was used as a calibration standard for the dry combustion. For pH determination 10 g of air-dried soil per sample was placed in a 50 mL Falcon tube and 20 mL of de-ionized water was added. The samples were shaken in a horizontal shaker for 1.5 hours and measured for pH using a pH electrode (Benchtop pH/ISE meter model 720A, Orion Research Inc., USA).

115
Resin-extractable P was used as an indicator of plant-available P, as it correlates with P uptake by plants (Nuernberg et al., 1998). Inorganic P was extracted using an anion exchange resin membrane (Kouno et al., 1995). The extraction method was slightly modified by using only one instead of two resin strips of 60 × 20 mm (No. 55164 2S, BDH Laboratory Supplies, Poole, England) saturated with CO −2 3 , and 2 g instead of 4 g of dried soil was weighted. No fumigation step to determine microbial P was performed, as the soils had been dried and had storage periods longer than one month between sampling 120 and analysis. In the resin eluates (a mixture of 0.1 M NaCl and 0.1 M HCl), the concentrations of inorganic P were measured colorimetrically using the malachite green method (Ohno and Zibilske, 1991). Bioavailable micronutrient (Fe, Mn, Zn, and Cu) concentrations in soils were determined with the diethylenetriaminepentaacetic acid (DTPA) extraction method, as described in Lindsay and Norvell (1978). The extracting solution consisted of 0.0005 M DTPA, 0.01 M CaCl 2 , and 0.1 M triethanolamine.
Briefly, 10 g of the sieved <2 mm) soils were extracted with 20 mL of DTPA solution. Micronutrient concentrations in the 125 filtrates were measured by inductively coupled plasma optical emission spectroscopy (ICP-OES, a Shimandzu Plasma Atomic Emission Spectrometer ICPE-9820). Final DTPA extractable concentrations of Fe, Mn, Zn, and Cu were calculated back to per kg dry soil. For each landscape, two soils were selected and analyzed in triplicates to assess analytical errors. For the remaining soils the analysis was not repeated.
For each sample, the concentrations of total element (Fe, Si, Al, K, Ca, P, Zn, Cu, and Mn) in the soil was assessed by 130 energy dispersive X-ray fluorescence spectrometry (ED-XRF) measurements on 4 g of the milled soil with a SPECTRO XEPHOS instrument (SPECTRO Analytical Instruments GmbH, Germany). The soil was mixed with equal amount of wax using a ball mill and pressed into pellets. Exchangeable cations (Ca 2+ , Mg 2+ , K + , Na + , and Al 3+ ) were determined with the BaCl 2 method (Hendershot and Duquette, 1986). About 2 g of the air-dried soil (<2 mm) were extracted by shaking for 2 hours with 30 mL of 0.1 M BaCl 2 on a horizontal shaker (120 cycles min −1 ). The suspension was filtered through no. 40 filter 135 paper (Whatman, Brentford, UK). For each landscape, two soils were analyzed in analytical triplicates. The concentrations of exchangeable cations in the BaCl 2 extract were measured by inductively coupled plasma optical emission spectroscopy (ICP-OES, Shimandzu Plasma Atomic Emission Spectrometer ICPE-9820). Different BaCl 2 extract dilutions were used in order to obtain an optimal signal intensity for the quantification of specific elements across all samples. Concentration of H + per kg dry soil was calculated based on the pH measured in the BaCl 2 extractant. The BaCl 2 extraction does only slightly modify 140 pH and is therefore an appropriate method to calculate effective CEC (CEC eff ) at native soil pH. Using the concentrations of the BaCl 2 -extractable cations (i.e. Ca 2+ , Mg 2+ , K + , Na + , Al 3+ and H + ), CEC eff was calculated as sum of exchangeable cations in cmol of cation charge per kg dry soil. Exchangeable acidity was defined by the sum of exchangeable Al 3+ and H + . Base saturation in % was calculated as ratio of the sum of basic cations (Ca 2+ , Mg 2+ , K + , and Na + ) in cmol(+) per kg soil to the CEC eff multiplied by 100.
Particle size analysis was conducted by the International Institute of Tropical Agriculture (IITA) in Cameroon, as described in Bouyoucos (1951). Briefly, 50 g of dried 2 mm sieved soil was stirred with 50 mL 4 % sodium hexametaphosphate and 100 mL of deionized water in a mixer, for breaking down the aggregates into into individual particles. Readings with a hydrometer (ASTM 152 H, Thermco, New Jersey, USA) were taken after letting it stand in the suspension for 30 minutes. The silt content was calculated by subtracting the measured proportion of sand and clay from 100%.

Spectroscopic measurements
The milled soils (n = 94) were measured on a Bruker ALPHA DRIFT spectrometer (Bruker Optics GmbH, Ettingen, Germany), which was equipped with a ZnSe optics device, a KBr beamsplitter, and a DTGS (deuterated tri-glycine sulfate) detector. Mid-IR spectra were recorded between 4000 cm −1 and 500 cm −1 with a spectral resolution of 4 cm −1 and a sampling resolution of 2 cm −1 . Reflectance (R) spectra were transformed to apparent absorbance (A) using A = log 10 (1/R) and corrected for 155 atmospheric CO 2 using macros within the OPUS spectrometer software (Bruker Corporation, US). The spectra were referenced to a IR-grade fine ground potassium bromide (KBr) powder spectrum, which was measured prior to the first soil sample and measured every hour again. All spectra were recorded by averaging 128 scans (internal measurements) to improve the signalto-noise ratio for each of the three independent replicate samples of each soil.

Processing of soil spectra
Three replicates of spectra were averaged for each sample. The spectra were transformed by using a Savitzky-Golay smoothed first derivative using a third-order polynomial and a window size of 21 points (42 cm −1 at spectrum interval of 2 cm −1 ) (Savitzky and Golay, 1964). Prior to spectral modeling, Savitzky-Golay preprocessed spectra were further mean centered and scaled (divided by standard deviation) at each wavenumber.

Model development and validation
The measured soil properties were modeled by applying partial least squares regression (PLSR) (Wold et al., 1983) with the preprocessed spectra as predictors. The models were fitted using the orthogonal scores PLSR algorithm. 5-times repeated 10fold cross-validation was performed to provide unbiased and precise assessment of PLSR model performance (Molinaro et al., 2005;Kim, 2009). For each individual soil property, the number of factors for the most accurate PLSR model was tuned 170 separately. For each soil property model, the sample set was repeatedly randomly split into k = 10 (approximately) equallysized subsets without replacement for all repeats r = 1, 2, .., 5 and all candidate values in the tuning grid with the number of PLSR factors (ncomp) = 1, 2, ..., 10. Within each of the r × ncomp = 5 × 10 = 50 resampling data set splits, each of the 10 possible held-out and model fitting set combinations (folds) was subjected to candidate model building at the respective ncomp, using k − 1 = 9 out of 10 subsets and remaining held-out samples were predicted based on the fitted models. The root 175 mean square error (RMSE, eq. (1)) of the held-out samples was calculated by aggregating all repeated K-fold cross-validation predictions (ŷ i ) and corresponding observed values (y i ) grouped by ncomp, which resulted in a cross-validated performance profile RMSE vs. ncomp.
Based on this performance profile, the minimal ncomp among the models whose performance was within a single standard 180 error ("One standard error" rule, (Breiman et al., 1984)) of the lowest numerical value of RMSE was selected.
Model assessment was done with the best factors for each property using cross-validation hold outs. We reported the crossvalidated measures RMSE, R 2 (coefficient of determination) obtained via linear least-squares regression, and ratio of perfor-mance to deviation (RPD), after averaging predictions across repeats. The RPD index is the ratio of the chemical reference data standard deviation (s y ) to the RMSE of prediction.
Besides calculating the above listed performance measures, the uncertainty of spectral estimates was graphically reported for each soil sample, using prediction means and 95% confidence intervals derived from cross-validation repeats (n = r = 5; Eq. 3 and 4).
In order to cover the full training data space in the models for future sample predictions, the final PLSR models were rebuilt using the entire training set and the respective values of optimal final number of PLSR components determined by the procedure described above.

Model interpretation 195
The mid-IR spectra contain complex information about soil composition and properties. To establish a predictive relationship, statistical models need to find relevant spectral features for each soil property. Model interpretation requires a variable importance assessment to decide on the contribution of spectral variables to prediction and to explain spectral mechanisms.
Therefore, we conducted model interpretation based on the variable importance in projection (VIP) method (Wold et al., 1993;Chong and Jun, 2005), using the model at respective best number of factors (ncomp). The VIP measure v j was calculated for where w aj are the PLSR weights for the a th component for each of the wavenumber variables and SS a is the sum of squares explained by the a th component: where q a are the scores of the predicted variable y and t a are the scores of the predictors X. These VIP scores account for multicollinearity found in spectra and are considered as robust measure to identify relevant predictors. Important wavenumbers were classified with a VIP score above 1. A variable with VIP above 1 contributes more than the average to the model prediction.
For model interpretation, we only computed VIP at the respective finally chosen number of PLS components a final for each considered model. We focused on a selection of three well performing models with R 2 ≥ 0.8 (RPD ≥ 2.3) to illustrate model 210 interpretation. These were total C, total N and clay content.

Statistical software
The entire analysis was performed using the R statistical computing language and environment (version 3.6.0) (R Core Team, 2017). We used the pls (Mevik et al., 2019) package for PLSR, as described by Martens and Naes (1989). Cross-validation resampling, model tuning, and assessment was done using the caret package (Kuhn et al., 2019). Custom functions from the 215 simplerspec package were used for spectroscopic modeling (Baumann, 2019). All data and code to reproduce the results of this study is available online via Zenodo (Baumann, 2020).

Measured properties and mid-IR estimates of yam soils
The distribution of soil properties at the yam fields showed a wide variation across the landscapes ( Figure 1). Total C concen-220 trations across all fields ranged from 2.4 g C kg −1 soil to 24.7 g C kg −1 soil. Total C values at the landscape scale were the lowest (median) in Léo and the highest in Tiéningboué. Soils from yam fields in the two landscapes from Ivory Coast (13.0 ± 5.4 g C kg −1 soil; mean ± standard deviation) had relatively higher total C compared to the fields in the landscapes in Burkina Faso (6.1 ± 3.6 g C kg −1 soil). The median value and variation of CEC eff exhibited similar patterns across the landscapes to total C. Total N concentrations across all fields ranged from 0.18 g N kg −1 soil to 2.48 g N kg −1 soil. Total N within and across 225 the four landscapes exhibited a similar pattern as total C. Generally, the landscapes in Burkina Faso were low in total N compared to those from Ivory Coast (0.44 ± 0.24 g N kg −1 soil vs. 1.09 ± 0.46 g N kg −1 soil). Median total N concentrations were almost identical for Liliyo and Tiéningboué with 1.1 g N kg −1 soil). Total S concentrations varied between 41 mg S kg −1 soil to 242 mg S kg −1 soil across all fields, and showed a similar pattern as total C and N. The yam fields in the landscapes of Bukina Faso had on average more than two times higher total S than the other landscapes. Total P concentrations were in a 230 similar range for the landscapes Léo, Midebdo, and Liliyo. In Tiéningboué, total P values were almost two times higher than the other fields (817 mg S kg −1 soil vs. 453 mg S kg −1 soil), with more within-landscape variation.
The concentrations of total Fe, total Al, total Ca, total Zn, and total Cu in the soil tended to be higher for the landscapes in   34.1 g K kg −1 soil), and lowest in Midebdo (range = 0.9-8.9 g K kg −1 soil), while the highest total K median was measured for yam fields in Léo (range = 4.1-25.0 g K kg −1 soil).

240
Soil resin-extractable P concentrations varied between 0.8 mg P kg −1 soil to 33.1 mg P kg −1 soil. In Tiéningboué, resinextractable P was on average higher than in soils of the other landscapes ( Figure 1). Median extractable Fe and its interquartile ranges were comparable across the landscapes (see Figure 1). However, there were some fields where extractable Reference measurements for total N, S, exchangeable Ca, exchangeable Mg and CEC eff. were highly correlated to total C ( Figure 2; 0.71 ≤ r ≤ 0.92 (CEC eff. )). Also, total Ca, Al, and clay content correlated strongly to total C (r > 0.70). Clay contents were weakly related to silt (r = 0.21), while sand had a markedly negative relationship to silt (r = −0.89). Bioavailable Cu 260 and Zn . Resin-extractable P was moderately correlated to total C (r = 0.53) and pH (r = 0.38). Bioavailable Zn (DTPA) was co-varying with both CEC eff. (r = 0.58) and total Zn (r = 0.59). Bioavailable Cu (DTPA) had a strongly positive association to total Cu (r = 0.90). Exchangeable K (BaCl 2 ) had the strongest relationship to total C and CEC eff. (r = 0.63, and r = 0.64).

Soil mid-IR spectroscopic models
Among the measured soil properties, mid-IR PLSR models for total K (R 2 = 0.96) and total Al (R 2 = 0.97) were best per-265 forming ( Table 1). Out of a total of 27 soil attributes, 11 were well quantified by the models when considering categorization judged upon on an R 2 cv 0.75 criterion (Figure 3). The confidence intervals derived from cross-validation prediction were very narrow, showing that all PLSR models were stable. Within this group of stable models, four soil attributes are directly related to the mineralogy (total Fe, Al, K and Ca), three are related to soil organic matter (total C, N and S), one to texture (clay fraction), one to plant nutrition (exchangeable Fe), and two related to mineralogy and plant nutrition (exchangeable Ca and 270 CEC eff ). More specifically, total C was accurately predicted, with an R 2 of 0.92 and a RMSE of 1.6 g C kg −1 soil. The models were also able to predict total N well (R 2 = 0.89; RMSE = 0.16 g N kg −1 soil). Prediction accuracy of total S was slightly lower Pearson correlation coefficients (r) were rounded to 1 digit. than for total C, but its goodness-of-fit and RMSE suggest that the model was reliable for prediction. However, exchangeable K (R 2 = 0.28), resin-extractable P (R 2 = 0.40) and BS eff (R 2 = 0.24) were poorly predicted (Table 1). Predictions for percent clay were reliable (R 2 = 0.81; RMSE = 2.1%), whereas predictions for percent sand (R 2 = 0.45; RMSE = 8.1%) and percent 275 silt (R 2 = 0.41; RMSE = 6.5%) were not accurate. Finally, chosen models of all soil attributes had between 1 and 9 PLSR components. Table 1. Descriptive summary of measured (meas.) soil reference data (see Figure 1) and evaluation results of cross-validated PLSR models.

Model interpretation
A large proportion of absorptions had VIP > 1 for each the total C, total N and clay models (Figure 4) Figure 3. Cross-validated predictions of soil properties derived from best mid-infrared (mid-IR) partial least squares regression (PLSR) models vs. laboratory reference measurements (see Figure 1). Average estimates, their confidence intervals (error bars), and evaluation metrics were derived with 5 × repeated 10-fold cross-validation. ncomp = number of PLSR components of most accurate final models, RSME = root mean square error, RPD = ratio of performance to deviation. Only soil properties modeled with R 2 > 0.75 are shown. CECeff = effective cation exchange capacity. Exchangeable (exch.) elements were determined with BaCl2. Bioavailable Fe was determined diethylenetriaminepentaacetic acid (DTPA) extraction.
continuous spectral features that were important to the models. For example, the relatively continuous and smooth spectral region between the alkyl C−H vibrations at 2855 cm −1 and 2362 cm −1 had comparable contribution to the model as peak regions associated with total C prediction. The VIP patterns across wavenumbers were almost identical for total C and N models, and its reference measurements were strongly correlated (r = 0.94; Figure 2). In contrast, the clay content model deviated from the total C model in particular regions, for example around the kaolinite OH− feature at 3620 cm . Variable importance analysis of partial least squares regression (PLSR) models for the concentrations of total soil C and total N, and clay content, including overlaid raw and preprocessed spectra. Top panel shows resampled mean sample absorbance spectra (n = 94).
Prominent peaks were identified as local maxima with a span of 10 points 20 cm −1 for the selected wavenumbers. Fundamental mid-IR vibrations that are well described in the literature (e.g., Madejová et al., 2002;Rossel and Behrens, 2010;Stevens et al., 2013) were added as labels when identified peaks matched literature assignments. (Q) stands for quartz and (K) for kaolinite. The middle panel depicts preprocessed spectra (Savitzky-Golay first derivative with a window size of 21 points (42 cm −1 ); 3rd order polynomial fit). The bottom panel shows variable importance in the projection (VIP) for three selected well performing PLSR models (total C, total N and clay; R 2 > 0.81).
The black horizontal line at VIP = 1 indicates the threshold above where absorbance at the wavenumbers explain more than average to the prediction of a certain soil property. Dashed points closely below the y = 0 line of the VIP graph visualize positive (above y = 0) and negative (below y = 0) PLSR β coefficients.

Accuracy and relevance of mid-IR spectroscopy for agronomic diagnostics
Timely and accurate estimates of multiple soil properties are required to better understand and predict soil constraints across the yam belt in West Africa. The soil spectral library from our study, which includes four landscapes of the yam belt, can 290 be practical to diagnose and monitor (and eventually manage) soil fertility that is considered to be low and therefore being a major constraint for yam production in West Africa. Specifically, our results show that properties closely related to organic matter -total amount of C, (micro)-nutrients, and exchangeable cations -can be accurately estimated using mid-IR spectra and in the selected yam growing landscapes (Figure 3). Soil organic matter plays a crucial role during vegetative growth and tuber formation phases of yam, as it guarantees among many other functions the storage and availability of essential 295 nutrients and water needed for yam and tuber growth throughout the season, and as well prevents soil erosion due to it's structural stabilization capacity. It promotes soil aggregation, which stabilizes soil organic matter and protects it from microbial decomposition (Six et al., 2006).
Fertilizers are becoming more essential to replenish mineral nutrients for prolonged cropping. Nevertheless, soil organic matter is at high risk of depletion in the regions because of the increasing land use frequencies and shorter fallows to restore the soil 300 organic C pools. While it is pivotal to develop innovative crop and soil management solutions to this problem (O'Sullivan and Jenner, 2006;Frossard et al., 2017;Kiba et al., 2020), it is also crucial to perform a separate but complementary activity to give feedback on potential soil changes: developing and applying soil conventional and proximal sensing methods. When testing sustainable soil and crop management options, for example to derive region-specific and farm-adapted nutrient management strategies, putting both validated quantitative statements on the status of soil organic carbon and local farmers' soil knowledge 305 into the equation is crucial (Wawire et al., 2021). Inevitably, both determining the inherent soil status (i.e., soil texture and organic carbon) and measuring the chemical and physical environment that regulates nutrient availability at trial sites (e.g., pH), is of agronomic and environmental importance (Foster, 1981). Maintaining and improving soil quality attributes will be paramount to sustain soils' ecosystem functions and crop yields over time. Activities to maintain and improve soil properties can for example be oriented towards fostering nutrient recycling.

310
Quick and reasonably accurate soil estimates derived from mid-IR spectra and empiric models as for example outlined in this study can inform site-adapted timing, placing and form of nutrient supply based on local soil conditions. To give a specific example, yam requires relatively large quantities of N and K (e.g., O'Sullivan, 2010); on light-textured soils, yam can attainhigh tuber yields, but at a high risk of loosing large proportions of applied N and K to the environment (e.g., Diby et al., 2011). Therefore, spectral estimates of texture can give an indication that applying larger amounts of N and K at once 315 would not improve yield potential under such situations. Hence, more frequent and local mineral applications of these nutrients after crop emergence, eventually combined with organic mulch, could improve the fertilizer efficiency and mitigate negative environmental impacts under these soil conditions. To estimate the availability of specific (micro)nutrients, however, more efforts need to be made to measure them in fine temporal and spatial resolution.
The mid-IR model accurately estimated C (RMSE = 1.6 g kg −1 soil; Table 1; Figure 3). Mostly, only field-scale spectroscopic 320 models achieve such accuracy (Nocita et al., 2015;Guerrero et al., 2016), whereas the predictive accuracy reported for largerscale application of spectroscopic models is lower than for our model (Rossel and Webster, 2012;Stevens et al., 2013;Sila et al., 2016). Models covering a wide geographical range of soils often result in high prediction errors (Stenberg and Rossel, 2010). Despite different soil types and climate regimes across a wide geographic spacing between the calibration fields, we achieved an accurate spectroscopic estimation of total C. The model was also able to reliably estimate a range of other important 325 soil properties than total C. Specifically, other soil variables eligible for a mid-IR quantification include total N, total S, total Ca, total K, total Al, exchangeable Ca, Fe DTPA, CEC eff. , and clay content (R 2 > 0.75). The high correlations of total C to N, S, exchangeable Ca, exchangeable Mg, CEC eff. , total Ca, Al, and clay content (Figure 2) are consistent with Johnson et al.
(2019), who reported very similar associations of clay content and exchangeable cations (Ca, Mg, K) as well as CEC eff. in soils from rice fields (0.54 ≤ r ≤ 0.65) -nevertheless they spectrally modeled a considerable soil variability (20 countries in sub-330 Saharan Africa; 42 study sites) and a larger sample size (n = 285) using PLS regression. At the same time, the measured range and the error in spectral estimates of CEC were larger compared to ours (RMSE = 6.7 cmol(+) kg −1 vs. 1.4 cmol(+) kg −1 ; range = 1.9-66.5 cmol(+) kg −1 vs. 0.9-14.6 cmol(+) kg). Even though, total K and Fe(DTPA) were poorly correlated to total C, their spectroscopic estimates were relatively accurate. This suggests that the mid-IR prediction of other soil properties is largely based on their correlation with total C as well as other absorption features of many organic and mineral soil components 335 having a specific IR adsorption.
We also found reasonable prediction accuracy for Cu(DTPA) (R 2 = 0.74) and Mn(DTPA) (R 2 = 0.55), despite that soil nutrients that are extraction-based or dependent on surface chemistry usually have variable predictive performance (Janik et al., 1998). Since relationships between soil composition and soil matrix exchange processes are typically complex, some properties may not be represented in the models in a straight-forward manner (Janik et al., 1998;Nocita et al., 2015). Although 340 the spectral estimates for resin-extractable P were not eligible for accurate soil diagnostics (R 2 = 0.40), they might be sufficient for the screening of site conditions. Singh et al. (2019) also reported varying applicability of available P within four regions of cocoa smallholder farms in Papua New Guinea (0.12 ≤ R 2 ≤ 0.58). A further aspect of interest is that resin-extractable P was mainly correlated to total P (r = 0.78), CEC eff. (r = 0.60) and total C (r = 0.52), while it poorly did to silt, sand and clay (−0.19 ≤ r ≤ 0.16). Despite that iron and aluminum oxides as well as clay minerals commonly cause P sorption (Gérard,345 2016), we found no correlation between resin-extractable P and total Fe or Fe (DTPA) (r = 0.05 and -0.03). This most likely is because the form rather than the abundance of iron oxides (poorly crystalline Fe) and kaolins controls the P adsorbed and available to crops in soils (Hartwig and Loeppert, 1993).
Although total elements are nor necessarily a direct proxy for plant-available nutrients -with exception of total C from organic matter -they can be related to mineralogical status, which is influenced by weathering and fertilization. For example, 350 total Fe from iron oxides can be an important control on the availability of P (Parfitt et al., 1975) and total P can be correlated to available P in other cases. For yam -which is an understudied crop with a relatively high yield gap geographicallyfertilizer response on N, P, and K is often absent under land that has been under long fallow periods (O'Sullivan, 2010). Even more importantly, the number of thoroughly conduced yam fertilizer trials in a region and for distinct soil types are often not sufficient for allowing site-specific calibration of soil tests with regard to fertilizer response and recommendations (O'Sullivan and Jenner, 2006).

Interpretation of spectral features
All mid-IR spectra that we measured for soils in the four landscapes exhibited a similar pattern of absorbance ( Figure 4).
The O-Si-O absorptions in quartz at 1080 cm −1 , 800-780 cm −1 and 700 cm −1 were a prominent feature in the spectra due to relatively high sand contents across the landscapes (range 30% to 92%, median 76%). Our spectra further had hydroxyl 360 (OH) absorptions that are typical for kaolin minerals, at 3695 cm −1 (surface OH), 3620 cm −1 (inner OH), 914 cm −1 (inner OH), and 936 cm −1 (surface OH) (Madejová et al., 2002). The spectral pattern between the hydroxyl bands at 3695 cm −1 and 3620 cm −1 was relatively consistent and the intensity ratio of these flanking peaks was close to 1. This is typical for halloysite (0.8-0.9) while the ratio for kaolinite is often higher (1.2-1.5) and dickite lower (0.6-0.8) (Lyon and Tuddenham, 1960). The two weak intermediary stretching absorptions at around 3657 cm −1 and 3670 cm −1 indicate surface hydroxyls. Together with 365 the absorption at 936 cm −1 , the spectra would suggest the presence of rather well-ordered prismatic halloysite (Hillier et al., 2016). This aligns well with the spectral patterns of soils that were assigned to the Halloysite archetype through similarity mapping (by comparison to the pure mineral spectra) by Sila et al. (2016). Our spectra confirm the presence of kaolin minerals, which reflects the advanced state of mineral weathering in these tropical soil types.
Our accurate predictions, which are comparable to field-scale calibrations, are most likely because of the relatively uniform 370 mid-IR spectra we obtained for our samples and their linear relationships to some of the key properties. This suggests a relatively homogeneous soil chemical composition, particularly with regard to the mineralogy in the sampled soils. Still, the data set presented here is relatively small and no randomized spatial sampling strategy was used for selecting field locations.
Therefore, we propose to implement a spectroscopy-driven approach to diagnose soils from other yam fields as an effort to broaden the library to achieve better spatial coverage of soil variability.

Conclusions
We developed models with mid-IR spectra to estimate soil chemical and physical properties relevant to production of yam and other staple crops in four landscapes in the yam belt of West Africa. We tested the models for the important soil properties that are applied widely for agronomic performance evaluation. We showed that mid-IR spectroscopy models have the potential to cost-effectively and rapidly determine the distribution and variability of important soil properties across highly variable yam 380 production landscapes in West Africa. Specifically, total C, total N, total S, total Fe, total Al, total K, total Ca, exchangeable Ca, CEC eff , bioavailable Fe, and clay content can be quantified with RPD > 2 and R 2 > 0.75 when aiming to predict in the range of soil property values found in the environmental conditions covered by this study. We achieved spectral estimates with quite small uncertainties, that are typically reported for libraries at the geographical extent of a field or farm. The correlation analysis of measured values together with spectral inference helps improve our understanding of how soil properties are interrelated to monitor and predict soil quality and to manage crop nutrition. Hence, we envision this pilot study as being a starting point to continuously update and adapt the mid-IR model library for more efficient site-specific and agronomically relevant soil estimates in the West African yam belt. This can bring better capacity to diagnose and and long-term monitor soils compared with traditional wet chemistry, and will hopefully ameliorate the soil conditions for sustainably meeting the demand of yam 390 and other important staple crops in the regions.
Code and data availability. All data and code to reproduce the results of this publication are publicly available under GNU General Public License v3.0, and can be accessed via the Zenodo archive and the corresponding github public repository (Baumann, 2020) Author contributions. Philipp Baumann carried out the research and analysis (soil sampling, sample preparation, soil chemical analysis, infrared spectroscopy, statistical modeling) under continuous support of the YAMSYS project team, and took the lead in writing the manuscript.
to thank Dr. Federica Tamburini for support in the CNS analysis, Éva Mészaros for her advice on the resin P protocol, and Björn Studer for the opportunity to perform XRF analyses on soils. We thank Raphael Viscarra Rossel and Marijn Van de Broek for their valuable feedback, which helped us to improve the manuscript.