LimeSoDa: A dataset collection for benchmarking of machine learning regressors in digital soil mapping

LimeSoDa: A da ase collec ion o benchma king o machine lea ning eg esso s in digi al soil mapping Jonas Schmidinge a,b,* , Sebas ian Vogel b , Viachesla Ba ko a,b , Anh-Duy Pham a,b , Robin Gebbe s b , Hamed Ta akoli b , Jose Co ea b , Tiago R. Ta a es c , Pa ick Filippi d , Edwa d J. Jones d , Voj ech Lukas e , E ic Boenecke , Joe g Ruehlmann , Ingma Sch oe e g , Ecka K ame g , S e an Pae zold h , Masakazu Kodai a i , Alexand e M.J.-C. Wadoux j , Luca B agazza k , Kon ad Me zge k , Jingyi Huang l , Domingos S.M. Valen e m , Jose L. Sa anelli n , Edua do L. Bo ega o , Rica do S.D. Dalmolin p , Csilla Fa kas q , Alexande S eige , Tacia a Z. Ho s s , Leona do Rami ez-Lopez ,u , Thomas Schol en ,w , Felix S ump x , Pablo Rosso y , Ma celo M. Cos a z , Rod igo S. Zandonadi aa , Johanna We e lind ab , Ma in A zmuelle a,ac a Osnab ück Uni e si y, Join Lab A i icial In elligence and Da a Science, Osnab ück, Ge many b Leibniz Ins i u e o Ag icul u al Enginee ing and Bioeconomy (ATB), Depa men o Ag omecha onics, Po sdam, Ge many c Uni e si y o S˜ ao Paulo (USP), Cen e o Nuclea Ene gy in Ag icul u e (CENA), Pi acicaba, B azil d The Uni e si y o Sydney, Sydney Ins i u e o Ag icul u e, Sydney, Aus alia e Mendel Uni e si y in B no, Depa men o Ag osys ems and Bioclima ology, B no, Czech Republic Leibniz Ins i u e o Vege able and O namen al C ops, Nex Gene a ion Ho icul u al Sys ems, G ossbee en, Ge many g Ebe swalde Uni e si y o Sus ainable De elopmen , Landscape Managemen and Na u e Conse a ion, Ebe swalde, Ge many h Uni e si y o Bonn, Ins i u e o C op Science and Resou ce Conse a ion (INRES)—Soil Science and Soil Ecology, Bonn, Ge many i Tokyo Uni e si y o Ag icul u e and Technology, Ins i u e o Ag icul u e, Tokyo, Japan j LISAH, Uni . Mon pellie , Ag oPa isTech, INRAE, IRD, L’Ins i u Ag o, Mon pellie , F ance k Ag oscope, Field-C op Sys ems and Plan Nu i ion, Nyon, Swi ze land l Uni e si y o Wisconsin-Madison, Depa men o Soil Science, Madison, USA m Fede al Uni e si y o Viçosa, Depa men o Ag icul u al Enginee ing, Viçosa, B azil n Woodwell Clima e Resea ch Cen e , Falmou h, USA o Fede al Uni e si y o San a Ma ia (UFSM), Academic Coo dina ion, San a Ma ia, B azil p Fede al Uni e si y o San a Ma ia (UFSM), Soil Depa men , San a Ma ia, B azil q No wegian Ins i u e o Bioeconomy Resea ch (NIBIO), Di ision o En i onmen and Na u al Resou ces, Aas, No way Uni e si y o Ros ock, Chai o Geodesy and Geoin o ma ics, Ros ock, Ge many s Fede al Technological Uni e si y o Pa an´ a, Dois Vizinhos, B azil BÜCHI Labo echnik AG, Da a Science Depa men , Flawil, Swi ze land u Impe ial College London, Impe ial College Business School, London, UK Uni e si y o Tübingen, Depa men o Geosciences, Tübingen, Ge many w Uni e si y o Tübingen, DFG Clus e o Excellence ‘Machine Lea ning o Science’, Ge many x Be n Uni e si y o Applied Sciences, Compe ence Cen e o Soils, Zolliko en, Swi ze land y Leibniz Cen e o Ag icul u al Landscape Resea ch (ZALF), Simula ion and Da a Science, Münchebe g, Ge many z Fede al Uni e si y o Ja aí, Ins i u e o Ag icul u al Sciences, Ja ai, B azil aa Fede al Uni e si y o Ma o G osso, Ins u e o Ag icul u al and En i onmen al Scinces, Sinop, B azil ab Swedish Uni e si y o Ag icul u al Sciences (SLU), Depa men o Soil and En i onmen , Ska a, Sweden ac Ge man Resea ch Cen e o A i icial In elligence (DFKI), Resea ch Depa men Plan-Based Robo Con ol, Osnab ück, Ge many ARTICLE INFO Handling Edi o : B andon Heung ABSTRACT Digi al soil mapping (DSM) elies on a b oad pool o s a is ical me hods, ye de e mining he op imal me hod o a gi en con ex emains challenging and con en ious. Benchma king s udies on mul iple da ase s a e needed o e eal s eng hs and limi a ions o commonly used me hods. Exis ing DSM s udies usually ely on a single da ase * Co esponding au ho . E-mail add ess: [email p o ec ed] (J. Schmidinge ). Con en s lis s a ailable a ScienceDi ec Geode ma jou nal homepage: www.else ie .com/loca e/geode ma h ps://doi.o g/10.1016/j.geode ma.2025.117337 Recei ed 4 Ma ch 2025; Recei ed in e ised o m 30 Ap il 2025; Accep ed 6 May 2025 Geode ma 459 (2025) 117337 A ailable online 20 May 2025 0016-7061/© 2025 The Au ho (s). Published by Else ie B.V. This is an open access a icle unde he CC BY license ( h p://c ea i ecommons.o g/licenses/by/4.0/ ). Keywo ds: Machine lea ning Benchma king Open da a Da ase collec ion Digi al soil mapping Pedome ics wi h es ic ed access, leading o incomple e and po en ially misleading conclusions. To add ess hese issues, we in oduce an open-access da ase collec ion called P ecision Liming Soil Da ase s (LimeSoDa). LimeSoDa consis s o 31 ield- and a m-scale da ase s om a ious coun ies. Each da ase has h ee a ge soil p ope ies: (1) soil o ganic ma e o soil o ganic ca bon, (2) clay con en and (3) pH, alongside a se o ea u es. Fea u es a e da ase -speci ic and we e ob ained by op ical spec oscopy, p oximal- and emo e soil sensing. All da ase s we e aligned o a abula o ma and a e eady- o-use o modeling. We demons a ed he use o LimeSoDa o benchma king by compa ing he p edic i e pe o mance o ou lea ning algo i hms ac oss all da ase s. This compa ison included mul iple linea eg ession (MLR), suppo ec o eg ession (SVR), ca ego ical boos ing (Ca Boos ) and andom o es (RF). The esul s showed ha al hough no single algo i hm was uni e sally su- pe io , ce ain algo i hms pe o med be e in speci ic con ex s. MLR and SVR pe o med be e on high- dimensional spec al da ase s, likely due o be e compa ibili y wi h p incipal componen s. In con as , Ca - Boos and RF exhibi ed conside ably be e pe o mances when applied o da ase s wi h a mode a e numbe (<20) o ea u es. These benchma king esul s illus a e ha he pe o mance o s a is ical me hods can be highly con ex -dependen . LimeSoDa he e o e p o ides an impo an esou ce o imp o ing he de elopmen and e alua ion o s a is ical me hods in DSM. 1. In oduc ion In digi al soil mapping (DSM), s a is ical modeling o he ela ion- ships be ween labo a o y measu ed soil p ope ies and seconda y ea- u es is used o c ea e soil maps o en i onmen al and ag icul u al applica ions. A key goal o DSM esea ch is o con inuously imp o e he accu acy o soil maps (Minasny and McB a ney, 2016). Gi en he s a- is ical and compu a ional na u e o he ask, mos o he ecen esea ch has ocused on imp o ing accu acy by implemen ing mo e sophis ica ed modeling app oaches—especially, since he adop ion o machine lea ning (ML) and deep lea ning algo i hms has inc eased p edic i e capabili ies (Heu elink and Webs e , 2022). Howe e , modeling pipe- lines in DSM a e usually no s aigh o wa d because se e al s a is ical challenges ha e o be add essed such as high-dimensional, noisy and in e co ela ed ea u es (Shi e al., 2023); he inco po a ion o spa ial and some imes empo al componen s (Heu elink and Webs e , 2022); o e y sca ce aining da a (Schmidinge e al., 2024b). Va ious s a is ical me hods ha e been sugges ed and de eloped in esponse. These ange om di e en ML algo i hms (Wadoux e al., 2020), p e-p ocessing echniques (Shi e al., 2023) and da a usion app oaches (Wang e al., 2022) o he de elopmen o new sampling designs (B us, 2019). While each o he sugges ed me hods may se e a speci ic pu pose, i is di icul o keep ack o he b oad and e e -e ol ing pool o a ailable ools. In addi ion, di e en s udies o en show con adic o y conclusions abou he e ec i eness o commonly used me hods, as discussed by ˇ Zíˇ zala e al. (2024) o sampling designs, Shi e al. (2023) o p e-p ocessing echniques, o Heung e al. (2016) o lea ning algo i hms. As a esul , i o en emains unclea how obus each o hese me hods is and o which con ex hey a e mos sui able. T us wo hy and in o ma i e benchma king s udies a e needed o ully unde s and hei capabili ies and s eng hs. Benchma king e e s o he me hodological compa ison o compe ing s a is ical me hods o assess and ank hei capabili ies (Nießl e al., 2022). Essen ially, in a benchma king s udy, he pe o mance o wo o mo e me hods (e.g., lea ning algo i hms) a e compa ed based on e alua ion me ics (e.g., R 2 ). Such s udies a e al eady common in DSM, as shown by ou suppo ing li e a u e e iew on benchma king s udies in 2023, p o ided in Appendix A. Howe e , he in o ma i e alue o hem is o en limi ed because he unde lying s udy designs do no sup- po comp ehensi e and gene alizable conclusions. In he ollowing, we will e e o he main indings om he li e a u e e iew on DSM benchma king (Appendix A) o highligh cu en ly p esen sho comings acco ding o se e al guidelines on benchma king (e.g., Boules eix e al., 2017; Nießl e al., 2022; Webe e al., 2019). We obse ed ha o e 95 % o DSM s udies elied on a single da ase o hei benchma king (see Appendix A, Fig. A2). These single da ase s we e ei he s udy-speci ic p op ie a y da ase s o one o he ew a ail- able open-access op ions, such as he Land Use and Co e age A ea F ame Su ey (LUCAS) (O giazzi e al., 2018). Howe e , gi en ha he pe o mance o me hods s ongly depend on he con ex and inhe en pa e ns wi hin da ase s, benchma king should include a subs an ially la ge numbe o da ase s (Boules eix e al., 2015; S obl and Leisch, 2024). The la ges numbe o da ase s used o DSM benchma king in 2023 was h ee (Appendix A, Fig. A2). In con as , classical ML s udies may ea u e ens o o e a hund ed da ase s o hei benchma king (Shmuel e al., 2024). Including a la ge numbe o da ase s ensu es a mo e obus analysis, as esul s will be less in luenced by indi idual da ase -speci ic pa e ns. Mo eo e , i allows he e alua ion o me hods unde a ying condi ions, which helps o unco e hei con ex -speci ic s eng hs and limi a ions (G insz ajn e al., 2022; Shmuel e al., 2024). Ano he c i ical aspec o obus benchma king is he open accessi- bili y o bo h code and da a. Fewe han 10 % o DSM benchma king s udies p o ided open da ase s and e en less han 5 % sha ed hei code (see Appendix A, Fig. A1), which is ma ginal compa ed o o he compu a ional scien i ic ields (Lau ina ichyu e e al., 2022; Pineau e al., 2021). While many s udies (48 %) included a s a emen ha indica ed a willingness ‘ o sha e da a upon eques ’, p io s udies on ep oducibili y ha e demons a ed ha only a mino i y o esea che s e en ually comply wi h his commi men (K a z and S asse , 2014; Lau ina ichyu e e al., 2022). The e o e, mos DSM benchma k s udies mus be deemed i ep oducible. Rep oducibili y in he con ex o modeling is c ucial no only o iden i ying po en ial e o s h ough code e iew (G eene e al., 2017) bu also because i allows esea che s o build upon exis ing wo k o cumula i e scien i ic p og ess (Kapoo and Na ayanan, 2022). Ye , ep oducibili y becomes e en mo e c i ical in he con ex o benchma king. Benchma king is o en conduc ed along- side he in oduc ion o new me hods o p o e hei me i s. I has been a gued ha his can make he benchma king inhe en ly biased owa ds he newly p oposed me hod (Nießl e al., 2022; Webe e al., 2019). This is because au ho s can ha e a compe i i e d i en in e es in publishing esul s ha show hei no el me hod as supe io o common compe i o s (B own e al., 2017). Such supe io i y may be achie ed h ough un ai s udy design choices, al hough o en unin en ionally (Nießl e al., 2022). In con as , ep oducible esea ch wi h sha ed code and sha ed da a encou ages mo e igo ous and less biased s udy designs. We a ibu e he wo a o emen ioned sho comings (i.e., he eliance on oo ew da ase s in benchma king and he lack o ep oducibili y), in pa icula o he lack o open da ase s o DSM pu poses. This issue has been add essed in o he academic ields wi h he es ablishmen o benchma k da ase collec ions (e.g., Mo is e al., 2020; Romano e al., 2022). Howe e , a compa able collec ion does no ye exis o DSM. While p og ess has been made o imp o e he da a-a ailabili y as indi- ca ed by he open soil spec oscopy lib a y (OSSL) by Sa anelli e al. (2025) and o he soil-da abases (Gobezie e al., 2024), he ew exis ing open soil da ase s a e p ima ily ocused on la ge-scale labo a o y-based soil spec oscopy o include only he a ge soil p ope ies wi hou accompanying p edic o ea u es. Fu he mo e, some open da ase s (e. g., LUCAS) we e published unde es ic i e licenses, which p ohibi s J. Schmidinge e al. Geode ma 459 (2025) 117337 2 sha ing o he da ase wi hin a code eposi o y. On he o he hand, open da ase s om a ield- and a m-scale o high- esolu ion p ecision ag i- cul u e a e almos comple ely missing, despi e he ac ha p ecision ag icul u e wi h p oximal soil senso s is an impo an applica ion o DSM (Gebbe s and Adamchuk, 2010). Field-scale DSM wi h p oximal soil senso s has i s own dis inc i e challenges such as aining da a sca ci y (Schmidinge e al., 2024b) nex o high-dimensional da a usion (Schmidinge e al., 2024a; Wang e al., 2022). To add ess hese issues, we in oduce P ecision Liming Soil Da ase s (LimeSoDa). LimeSoDa is an open da ase collec ion wi h small-scale da ase s, ha a e eady- o- use o modeling. This expands he possibili y o DSM p ac i ione s o benchma k s a is ical me hods ac oss a a ie y o soil mapping con ex s. In Sec ion 2, we p o ide an o e iew o he da ase s included in LimeSoDa. Addi ionally, we p esen an example benchma k s udy in which we compa ed ou lea ning algo i hms (Sec ion 3), o demon- s a e how LimeSoDa can lead o mo e comp ehensi e and nuanced conclusions. Las ly, Sec ion 4 o e s an ou look on he po en ial appli- ca ions o LimeSoDa beyond he p esen s udy. 2. LimeSoDa 2.1. O e iew LimeSoDa con ains 31 da ase s a ield- o a m-scale. Da ase s a e eady- o-use o modeling, which means ha hey con ain con inuous a ge soil p ope ies and ea u es in a abula o ma , usable o eg ession asks. Th ee a ge soil p ope ies a e included in all da ase s: soil o ganic ma e (SOM) o soil o ganic ca bon (SOC), pH and clay. Since SOC and SOM a e di ec ly ela ed, we ea ed hem in e change- ably as a single soil p ope y. These h ee soil p ope ies a e among he mos commonly e alua ed soil p ope ies in DSM (Chen e al., 2022) and a e c ucial pa ame e s o assessing soil quali y. We e e o his benchma king collec ion as “LimeSoDa” because hese h ee a ge soil p ope ies a e no exclusi ely bu especially ele an o lime equi e- men calcula ions, acco ding o bes managemen p ac ices in he UK and Ge many (Ag icul u e Ho icul u e De elopmen Boa d, 2023; B¨ onecke e al., 2021). Fea u es o modeling a e da ase -speci ic and o igina e om labo a o y-based spec oscopy, in-si u p oximal soil sensing, and emo e sensing. Da ase s a e eleased unde a pe missi e open-access license, allowing implemen a ion in a code- eposi o y o inc ease he possibili y o code sha ing. Addi ionally, p e-de e mined olds o c oss alida ion (CV) a e p o ided o each da ase , o enable compa abili y wi h ou benchma king and u u e benchma king esul s. In o al, he combined da ase s include 3,174 soil samples bu sample sizes o indi idual da ase s ange om 30 o 460 samples. P e-p ocessing is no manda o y when wo king wi h LimeSoDa da ase s. Howe e , o da ase s wi h high-dimensional spec al da a, ea u e dimensionali y educ ion is s ongly encou aged. Ne e heless, u he p e-p ocessing may imp o e modeling pe o mances, an aspec ha should i sel be benchma ked using his da ase collec ion. Table 1 Lis o da ase s included in LimeSoDa. Da ase ID Loca ion S udy A ea (ha) Numbe o Samples Numbe o Fea u es Senso Da a* P e ious Usage** SSP.460 S a e o Sao Paulo, B azil 473 460 830 is-NIR Rami ez-Lopez e al. (2019) BB.250 B andenbu g, Ge many 52 250 17 DEM, ERa, Gamma, pH-ISE, RSS, VI Schmidinge e al. (2024b) SP.231 Sai ama P e ec u e, Japan 3.1 231 272 is-NIR Kodai a and Shibusawa (2020) B.204 Bahia, B azil 204 204 16 DEM, RSS, VI Pe ei a e al. (2022) G.150 Goias, B azil 79 150 17 DEM, ERa, RSS, VI Valen e e al. (2024) H.138 Hubei, China 420 138 2,489 MIR Wadoux e al. (2025) SL.125 Skåne L¨ an, Sweden 78 125 2,082 ERa, is-NIR We e lind e al. (2010) UL.120 Uppsala L¨ an, Sweden 97 120 2,082 ERa, is-NIR We e lind e al. (2010) NRW.115 No h Rhine-Wes phalia, Ge many 17 115 1,686 MIR Leenen e al. (2022) MG.112 Ma o G osso, B azil 111 112 17 DEM, ERa, RSS, VI Valen e e al. (2024) SA.112 Saxony-Anhal , Ge many 27 112 1,412 DEM, ERa, Gamma, NIR, pH- ISE, VI − G.104 Goias, B azil 95 104 16 DEM, RSS, VI − MGS.101 Ma o G osso do Sul, B azil 95 101 16 DEM, RSS, VI − CV.98 Can on o Vaud, Swi ze land 28 98 2,151 is-NIR Me zge e al. (2024) SC.93 San a Ca a iMG-na, B azil 108 93 2,146 is-NIR Ho s e al. (2018) BB.72 B andenbu g, Ge many 3.4 72 17 DEM, ERa, Gamma, pH-ISE, RSS, VI − NRW.62 No h Rhine-Wes phalia, Ge many 0.6 62 1,686 MIR Leenen e al. (2019) RP.62 Rhineland-Pala ina e, Ge many 3.3 62 1,410 ERa, Gamma, NIR, pH-ISE, VI Ta akoli e al. (2022) SSP.58 S a e o Sao Paulo, B azil 0.7 58 351 is-NIR Ta a es e al. (2020) NSW.52 New Sou h Wales, Aus alia 1,158 52 5 DEM, RSS Filippi e al. (2019)Jones e al. (2021) BB.51 B andenbu g, Ge many 40 51 4 DEM, ERa, pH-ISE − SC.50 San a Ca a ina, B azil 13 50 3 DEM, ERa Bo ega e al. (2022) W.50 Wisconsin, USA 80 50 15 DEM, ERa, VI, XRF Cha e jee e al. (2021) PC.45 Pes Coun y, Hunga y 4.5 45 4 CSMois u e, ERa Ris olainen e al. (2006) MG.44 Ma o G osso, B azil 13 44 351 is-NIR Ta a es e al. (2020) NRW.42 No h Rhine-Wes phalia, Ge many 1.5 42 1,686 MIR Leenen e al. (2019) SM.40 Sou h Mo a ia, Czechia 53 40 3 DEM, ERa Lukas e al. (2009) MWP.36 Mecklenbu g-Wes e n Pome ania, Ge many 18 36 5 DEM, RSS S eige e al. (2025) O.32 Occi ania, F ance 1.5 32 1,637 MIR Weh le e al. (2022) BB.30_1 B andenbu g, Ge many 19 30 8 DEM, ERa, pH-ISE, VI − BB.30_2 B andenbu g, Ge many 1.4 30 13 DEM, ERa, Gamma, RSS, VI − *Abb e ia ions: Capaci i e soil mois u e (CSMois u e), Digi al ele a ion model and e ain pa ame e s (DEM); Appa en elec ical esis i i y (ERa); Gamma- ay ac- i i y (Gamma); Mid in a ed spec oscopy (MIR); Nea in a ed spec oscopy (NIR); Ion selec i e elec odes o pH de e mina ion (pH-ISE); Remo e sensing de i ed spec al da a (RSS), X- ay luo escence de i ed elemen al concen a ions (XRF), Vege a ion indices (VI), Visible- and nea in a ed spec oscopy ( is-NIR). ** The da ase s used in he e e enced s udies does no always comple ely align wi h he da ase s o LimeSoDa, because only a subse o addi ional da a may ha e been used. J. Schmidinge e al. Geode ma 459 (2025) 117337 3 All da ase s we e collec ed in he con ex o p e ious s udies o esea ch p ojec s. The as majo i y o hese da ase s had no been publicly a ailable p io o his e o . Hence, LimeSoDa is buil upon olun a ily submissions om esea che s and esea ch ins i u es. An o e iew o each da ase is p o ided in Table 1. Da ase s we e included based on he ollowing ou c i e ia: P ecision liming con ex : A da ase had o con ain SOM/SOC, pH, and clay as opsoil (<30 cm) a ge p ope ies a ield- and a m-scale (<2,000 ha). We chemis y: Ta ge soil p ope ies had o be de e mined h ough we chemis y echniques ins ead o being in e ed om spec al models. P edic i e pe o mance: Ta ge soil p ope ies had o be p edic - able wi h he gi en da ase -speci ic se o ea u es. We excluded da ase s whe e no lea ning algo i hm ou pe o med he null model (i.e., wi h an R 2 <0) o any o he h ee a ge soil p ope ies gi en he modeling pipeline in Sec ion 3.1. Sample size: A da ase had o con ain a minimum size o a leas 30 soil samples o all h ee soil p ope ies a he same dis inc sampling loca ions. Da ase s we e no excluded because o hei sampling design. Fo example, some da ase s we e based on a a ge ed sampling design (see Appendix B, Table B1), which e e s o sampling ocused on speci ic a eas o ea u es o in e es a he han andom o sys ema ic sampling. The alida ion o soil maps gene a ed om a ge ed- o spa ially clus- e ed sampling wi hou an addi ional independen es ing sample se is con o e sial because he absolu e pe o mance may no be eliably es ima ed h ough CV o da a spli ing (Piikki e al., 2021). Howe e , we a gue ha o he pu pose o benchma king, he ela i e pe o mance di e ences emain in o ma i e. None heless, use s o LimeSoDa can decide i hey p e e o he alida ion s a egies ha ake in o accoun he spa ial dependency o pa e n o a da ase . Spa ial coo dina es a e a ailable o mo e han hal o he da ase s (see Appendix B, Fig. B1). In o he cases, coo dina es had o be excluded o anonymized mos ly because o p i acy conce ns. 2.2. P ocessing o da ase s Only a minimal amoun o p ocessing was conduc ed o da ase s o LimeSoDa o ensu e ha he da a was as close o he o iginal o m as possible while s ill enabling e ec i e benchma king. This gi es use s o LimeSoDa he lexibili y o apply hei own p e-p ocessing on he da ase s. O dina y k iging was employed o da ase s collec ed wi h on- he-go p oximal soil senso s o ma ch sensing loca ions wi h soil sam- pling loca ions. Fo ea u es a ailable in as e o ma (DEM, RSS and VI), ea u e alues we e ex ac ed a he soil sampling loca ions. Some senso s had an unde lying da a-p ocessing s ep wi hin hei in e nal so wa e. This led o some op ical spec oscopy da a being esampled o di e en wa ebands and spec al esolu ions compa ed o he measu ed aw da a. Fo da ase s wi h limi ed aining samples (<60 samples), we included ewe emo e sensing-based ea u es o main ain a a o able sample- o- ea u e a io. P ocessing s eps a e documen ed in mo e de ail in he me ada a o he da ase s. Samples wi h missing alues in ei he he ea u e o a ge soil p ope y ma ix we e always disca ded. Two spec oscopy da ase s (H.138 and NRW.115) had a ew samples wi h e lec ance alues abo e 100 % o ce ain bands (see Appendix B, Fig. B2). Whe he hese anomalies a e due o noise o ins umen al calib a ion is unclea . Ne e heless, we did no disca d hese samples o bands om he da ase . I is impo an o no e ha da ase s ac oss LimeSoDa we e delibe - a ely no ha monized. While his educes he possibili y o c oss-da ase lea ning, i inc eased he lexibili y o including da ase s om a ious con ex s and domains wi hou elying on many assump ions necessa y o da a ha moniza ion. The aim o LimeSoDa is no o build a uni ied da abase bu o p o ide mul iple smalle da ase s ha a e independen en ies o a benchma king s udy. As a consequence, a ge soil p op- e ies a e some imes exp essed in di e en uni s depending on he measu emen me hod (see Appendix B, Fig. B2). This dis inguishes LimeSoDa om spec al lib a ies, ha ely on ha moniza ion o spec al ea u es and soil p ope ies (Sa anelli e al., 2025). None heless, o cohe ence wi hin LimeSoDa, we ans o med he appa en elec ical conduc i i y o ERa and spec al measu emen s a e exp essed as e lec ance. 2.3. Accessibili y LimeSoDa is closely aligned o he FAIR-p inciples (Wilkinson e al., 2016). I is eely accessible h ough Zenodo (h ps://doi.o g/10.5 281/zenodo.14932573), is licensed unde CC BY-SA 4.0 and con ains ex ensi e documen a ion in he o m o da ase -speci ic me ada a. Addi ionally, an R- and Py hon da ase package, called likewise Lime- SoDa downloadable om Gi Hub, was c ea ed. I can be accessed h ough gi hub.com/JonasSchmidinge /LimeSoDa o R and h ps:// gi hub.com/a11 o1n3/LimeSoDa o Py hon. 3. Demons a i e benchma k s udy 3.1. Me hodology A benchma king s udy was conduc ed o demons a e he use o LimeSoDa. Fou lea ning algo i hms we e compa ed based on hei p edic i e pe o mance. These included andom o es (RF) (B eiman, 2001) and suppo ec o eg ession (SVR) due o hei widesp ead use in DSM (Khaledian and Mille , 2020), ca ego ical boos ing (Ca Boos ) (P okho enko a e al., 2018) because o i s s ong pe o mances in ecen ML benchma king s udies on abula da ase s (McEl esh e al., 2023; Shmuel e al., 2024) and mul iple linea eg ession (MLR) o include a simple baseline model. We ained and e alua ed each algo- i hm independen ly on SOC/SOM, pH and clay o each o he 31 da ase s o LimeSoDa. Consequen ly, a o al o 93 p edic ion asks (31 da ase s ×3 a ge soil p ope ies) we e u ilized in he benchma king. Ac oss hese 93 p edic ion asks, algo i hms we e compa ed on he R 2 and he o dinally anked oo mean squa e e o (RMSE) (see Appendix C.1). Addi ionally, he Wilcoxon Signed-Rank Tes was used o de e - mine whe he he mean o dinal anks based on he RMSE signi ican ly di e ed ( α =0.1) among he lea ning algo i hms. The lea ning algo i hms we e e alua ed and hype pa ame e s uned, using a nes ed K- old c oss alida ion (CV), wi h he ou e loop (K =10) used o model es ing and he inne loop (K’ =5) o hype pa ame e selec ion. The op imal hype pa ame e combina ions o SVR, Ca Boos and RF we e selec ed h ough a andom sea ch wi h 400 i e a ions based on he lowes agg ega ed RMSE o he inne loop. The hype pa ame e sea ch-space is gi en in Table 2. Fo da ase s wi h high-dimensional op ical spec oscopy da a ( is- NIR, NIR o MIR), modeling wi h he aw da a is s ongly limi ed due o he un a o able sample o ea u e a io. Hence, we duplica ed he hype pa ame e sea ch space o 800 i e a ions and added wo unsu- pe ised bu compu a ionally e icien me hods o dimensionali y educ ion as addi ional hype pa ame e b anch: p incipal componen analysis (PCA) and a co ela ion ma ix il e (CMF) (Pe ez-Ri e ol e al., 2017). CMF and PCA we e applied only o he is-NIR, NIR, o MIR ea u es a e he sepa a ion o aining and es ing o alida ion olds. Wi h CMF, spec al ea u es we e disca ded when he absolu e pai wise Pea son co ela ion coe icien exceeded a de ined cu o , ollowing he indCo ela ion algo i hm o he R-package ca e (Kuhn, 2008). The cu o o CMF anged om 0.7 o 1, whe e CMF =1 e e s o no dimensionali y educ ion, and he numbe o sea ched p incipal com- ponen s o PCA om 5 o 20. Algo i hm 1 p esen s he modeling pipeline as pseudocode. The ac ual R-code is a ailable on gi hub.com /JonasSchmidinge /LimeSoDa_benchma king. J. Schmidinge e al. Geode ma 459 (2025) 117337 4 3.2. Resul s & Discussion The R 2 dis ibu ion ob ained om all p edic ion asks wi h he ou di e en lea ning algo i hms is shown in Fig. 1a. O e all, he e appea o be small di e ences in he pe o mance dis ibu ion. Al hough SVR and Ca Boos achie ed he bes a e age R 2 o 0.44 ac oss all p edic ion asks, hey we e only mode a ely be e han MLR, which had he lowes a e age R 2 o 0.40. Simila ly, his is also e lec ed by he o dinally anked RMSE in Fig. 2a. I illus a es how equen ly a p edic ion al- go i hm achie ed each ank compa ed o o he s, based on he ascending RMSE. Fo example, MLR had he lowes (i.e., he bes ) RMSE o 27 % o he p edic ion asks, ecei ing he i s ank. Simul aneously, i had he highes (i.e., he wo s ) RMSE o 30 % o p edic ion asks, placing i in he ou h ank. O e all, Fig. 2a esembles a uni o m-like dis ibu ion. This means ha each lea ning algo i hm had oughly he same numbe o p edic ion asks in which hey anked bes o wo s . The Wilcoxon Signed-Rank Tes u he con i med ha di e ences among he lea ning algo i hms we e no s a is ically signi ican (Fig. 2a). This migh be su p ising o some DSM p ac i ione s, because mo e sophis ica ed me hods a e o en expec ed o signi ican ly ou pe o m a simple model like MLR (Pada ian e al., 2020). Howe e , di e en lea ning algo- i hms, including simplis ic models like MLR, ha e ad an ages unde ce ain condi ions, as illus a ed in he nex pa ag aphs. G ouping he p edic ion asks based on he p esence o absence o is-NIR, NIR o MIR ea u es e eals a signi ican shi in he pe o - mance dis ibu ion among he lea ning algo i hms (Fig. 1b-c & Fig. 2b- c). The g ouping is made because da ase s wi h is-NIR, NIR and MIR a e cha ac e ized by a highly in la ed ea u e o aining sample a io, so ha dimensionali y educ ion is usually necessa y p io o he modeling. On da ase s wi hou is-NIR, NIR and MIR, ee-based lea ning Table 2 Hype pa ame e sea ch-space o SVR, Ca Boos and RF, whe e a andom ins ance is d awn om he gi en dis ibu ion. Names o he hype pa ame e s e e o he ange R-package (W igh and Ziegle , 2017) o RF, ca boos R- package o Ca Boos (Dmi ie e al., 2024) and e1071 R-package (Meye e al., 2024) o SVR. Lea ning Algo i hm Hype pa ame e Sea ch Space Dis ibu ion Ca Boos dep h [1, 10] Disc e e Uni o m lea ning_ a e [0.005, 0.5] Log-Uni o m i e a ions [50, 2000] Disc e e Uni o m l2_lea _ eg [0, 10] Uni o m sm [0.6, 1] Uni o m subsample [0.6, 1] Uni o m andom_s eng h [0.001, 10] Log-Uni o m RF num. ees =2 000 − m y [0.1, 1]* Uni o m min.node.size [1, 12] Disc e e Uni o m max.dep h [1, 10] Disc e e Uni o m sample. ac ion [0.6, 1] Uni o m SVR cos [0.01, 1000] Log-Uni o m gamma [0.001, 10] Log-Uni o m ke nel [linea , adial] Disc e e Uni o m * Gi en as ac ion ins ead o absolu e alue inpu o R- unc ion. J. Schmidinge e al. Geode ma 459 (2025) 117337 5 algo i hms (i.e., RF and Ca Boos ) conside ably ou pe o med SVR and MLR (Fig. 1b). This is u he highligh ed by he ac ha in 63 % o he p edic ion asks in hese da ase s, RF o Ca Boos had he bes pe o - mance (Fig. 2b). Meanwhile, MLR o SVR had he wo s pe o mance in 72.5 % o he cases. The ad an ages o ee-based lea ning algo i hms o egula abula da ase s ha e been highligh ed in nume ous bench- ma king s udies (G insz ajn e al., 2022; Shmuel e al., 2024; Shwa z- Zi and A mon, 2022). Thei pa icula s eng hs include e ec i e eg- ula iza ion when dealing wi h mul icollinea i y and i ele an ea u es, along wi h he abili y o lea n i egula unc ions o in e ac ions. Fo his eason, ee-based algo i hms, mos no ably RF, ha e become he mos widely used lea ning algo i hms in DSM (Khaledian and Mille , 2020). Opposi e esul s a e e iden o da ase s wi h is-NIR, NIR and MIR. He e, SVR and MLR had on a e age an R 2 conside ably la ge han ha o Ca Boos and RF (Fig. 1c). Addi ionally, hey anked i s in 69 % o he p edic ion asks (Fig. 2c), whe eas RF o Ca Boos had he wo s pe o mances in 73.8 % o he cases. The e y high dimensionali y and inhe en mul icollinea i y in spec al da a, whe e adjacen bands a e o en highly co ela ed, made ea u e dimensionali y educ ion ech- niques, such as PCA and CMF, essen ial. Fo example, he da ase O.32 had up o 1,637 ea u es bu only 32 soil samples a ailable o aining (Table 1). Lea ning algo i hms would se e ely o e i on his in la ed ea u e o aining sample a io. Howe e , CMF and especially PCA acili a ed p ope modeling, as shown by he bes selec ed Fig. 1. Violin plo showing he dis ibu ion o R 2 alues g ouped by lea ning algo i hms o p edic ion asks om (a) all da ase s, (b) da ase s wi hou is-NIR, NIR, o MIR, and (c) da ase s wi h is-NIR, NIR, o MIR. Ho izon al lines wi h labels ep esen he mean R 2 alue. Fig. 2. Ba plo showing he ela i e equency dis ibu ion o o dinal anks based on he lowes RMSE (i.e., lowe ank numbe indica es be e pe o mance) o he ou lea ning algo i hms o p edic ions asks om (a) all da ase s, (b) da ase s wi hou is-NIR, NIR, o MIR, and (c) da ase s wi h is-NIR, NIR, o MIR. Mean anks o he o dinally anked RMSE o he lea ning algo i hms a e shown, including signi icance le e s om he Wilcoxon Signed-Rank Tes ( α =0.1) o indica e whe he he pe o mance di e ences be ween he lea ning algo i hms we e s a is ically signi ican . J. Schmidinge e al. Geode ma 459 (2025) 117337 6 hype pa ame e s (see Appendix C.2, Fig. 1C). This posed a disad an age o he ee-based algo i hms because PCA has been epo ed o pe o m weake when combined wi h ee-based algo i hms (Howley e al., 2005). PCA c ea es new unco ela ed ea u es ha a e linea combina- ions o he o iginal ea u es bu ee-based models a e in lexible o adap o o a ed o ans o med da a ha do no ep esen he o iginal ea u e composi ion (G insz ajn e al., 2022). Consequen ly, MLR and SVR ou pe o med Ca Boos and RF due o hei be e compa ibili y wi h PCA. P e ious benchma king s udies in soil spec oscopy, as e iewed by Pada ian e al. (2020), ha e epo ed di e en ou comes a o ing mo e sophis ica ed me hods. One possible explana ion o his disc epancy is ha hese ea lie s udies elied on a single da ase , which may inad e en ly lead o o e in e p e a ion o inciden al esul s. Al e na i ely, an unin ended publica ion bias a o ing no el me hods o e simple ones, as obse ed in o he compu a ional ields (Buchka e al., 2021), could ha e con ibu ed o he di e ing ou comes. None- heless, we acknowledge ha ee-based models possibly pe o m be e on ull spec al da a when e y la ge aining se s a e a ailable (Clingensmi h and G unwald, 2022) o wi h e ec i e ea u e selec ion (Cane o e al., 2024). Such an ex ended analysis was no wi hin he scope o his s udy. Las ly, we ound a ela ionship be ween he mean o dinal ank o a lea ning algo i hm and he sample size a ailable o a p edic ion ask (Fig. 3). Mos no ably, MLR showed sensi i i y o he numbe o a ail- able samples, as i s mean ank wo sened o da ase s wi h mo e han 100 samples. This aligns wi h p e ious esul s, which simila ly indica ed ha MLR can ha e ad an ages o small sized da ase s bu may no be ideal i enough aining da a is a ailable (Schmidinge e al., 2024b). The opposi e beha io was obse ed o SVR, which conside ably imp o ed i s ank o da ase s wi h mo e han 100 samples. Fo RF and Ca Boos , he e ec o he sample size was no as p onounced. Ca Boos was sligh ly be e wi h mo e aining da a, whe eas RF was mo e e ec i e o smalle da ase s. Despi e he ad an ages ha ce ain lea ning algo i hms ha e, hei supe io i y is no de e minis ic. Thei pe o mances can a y based on nume ous ac o s and may no be op imal in e e y case. Fo example, RF was gene ally he wo s pe o ming lea ning algo i hm o da ase s wi h is-NIR, NIR and MIR (Fig. 1b), ye i s ill u ned ou o be bes in 10 % o he p edic ion asks (Fig. 2b). The e o e, elying on a single da ase is insu icien o es ablish he supe io i y o one me hod and may os e misleading conclusions based on non-gene alizable esul s. 4. Fu he applica ions Fo demons a i e pu poses, we es ic ed he benchma king o ou lea ning algo i hms, al hough many o he ele an me hods emain o be explo ed. The benchma king s udy o Sec ion 3.2 is ully ep oduc- ible, as he code and da ase s ha e been published. The e o e, addi ional lea ning algo i hms o di e en ea u e selec ion s a egies can eadily be in eg a ed in o he pipeline o Sec ion 3.1 o u he ex end he analysis. While neu al ne wo ks we e no included in his ini ial benchma king due o hei ela i ely poo pe o mances in la ge abula benchma king s udies (Shmuel e al., 2024), ecen ad ance- men s in neu al ne wo ks designed o abula da a such as TabPFN (Hollmann e al., 2025) ha e demons a ed p omising esul s. Fi s success ul applica ions o TabPFN in DSM can be ound by Ba ko e al. (2024) bu i s applica ion has o be u he e alua ed. Las ly, we encou age ex ending he benchma king wi h mo e open da ase s, such as hose a ailable om OSSL, in addi ion o LimeSoDa. Wi hin he scope o his s udy, we ocused en i ely on he need o open da ase s o benchma king pu poses o add ess cu en ly p esen sho comings. None heless, he usage o LimeSoDa should no be es ic ed o s a is ical benchma king o me hod de elopmen . Open da ase s om a ious geog aphical con ex wi h di e en sensing ech- niques may enhance ou gene al unde s anding o pedological p o- cesses. I allows o he c i ical e alua ion o soil mapping in ag onomical decision making and asses he obus ness o a ious soil sensing echniques. Especially he liming con ex o he da ase can be use ul o s udies in he ield o p ecision ag icul u e. Consequen ly, LimeSoDa is a aluable ool o add essing key challenges o pedo- me ics ou side o DSM benchma king (Wadoux e al., 2021). In o he academic ields, seconda y da a analysis has al eady answe ed a ious esea ch ques ions un ela ed o he o iginal esea ch pu pose o he da ase (G eene e al., 2017). Las ly, spec oscopy da a om LimeSoDa can be ha monized and added o o he spec al lib a ies as addi ional aining da a o imp o e global modeling. 5. Conclusion Cu en benchma king p ac ices in DSM su e om da a limi a ions ha lead o incomple e o po en ially biased conclusions. The e is a lack o open da ase s om a ious domains, spa ial dimensions and ypes o senso s. To add ess his p oblem, we in oduced an open-access da a collec ion called LimeSoDa, which cu en ly consis s o 31 ield- and a m-scale da ase s. LimeSoDa o e s da ase s ha a e eady- o-use o modeling and co e s a ious ypes o ea u es om di e en sensing Fig. 3. Line plo showing he mean o dinal ank based on he lowes RMSE (i.e., lowe ank numbe indica es be e pe o mance) in dependence o he sample size. Lines a e di e en ia ed by he ou lea ning algo i hms and he ype o ea u es p esen in he da ase o he p edic ion ask. J. Schmidinge e al. Geode ma 459 (2025) 117337 7 echniques. This enables hose who a e wo king in he ield o DSM o benchma k s a is ical me hods on a di e se ange o soil da ase s. Fu he , he open license is in ended o imp o e he ep oducibili y. The u ili y o LimeSoDa was demons a ed h ough a benchma king s udy wi h ou lea ning algo i hms. The esul s showed ha no algo- i hm signi ican ly exceeded he o he s ac oss all da ase s. Ins ead, di e en lea ning algo i hms had ad an ages depending on he ype o ea u es and sample size o a da ase . On a e age, ee-based algo i hms, i.e., Ca Boos and RF, pe o med be e on da ase s wi hou is-NIR, NIR and MIR ea u es, p o ing hei sui abili y o con en ional abula da ase s. In con as , SVR and MLR ou pe o med Ca Boos and RF o da ase s wi h is-NIR, NIR and MIR due o hei be e compa ibili y wi h PCA- ans o med da a. Addi ionally, he aining sample size in luenced he anking o a lea ning algo i hm. The ela i e pe o mance mos no ably dec eased o MLR and inc eased o SVR wi h mo e aining samples. A benchma king s udy based on a single da ase could no ha e e ealed such con ex -dependen pe o mance o lea ning algo i hms. Mo e so, elying on singula da ase s isks o e in e p e ing inciden al ou comes and a po en ial publica ion bias a o ing newe me hods canno be uled ou . In con as , LimeSoDa acili a es mo e in-dep h analyses and enables comp ehensi e conclusions. Addi ionally, he benchma king can eadily be ex ended wi h u he lea ning algo i hms o p e-p ocessing echniques because da ase s and code a e openly a ailable. Beyond benchma king, he e a e u he applica ions and challenges in pedome ics ha bene i om open da ase s. LimeSoDa can be used o in es iga e pedological p ocesses ac oss di e se geog aphical con ex s o he spec al da a can be in eg a ed in o o he spec oscopy lib a ies o imp o e global modeling. In summa y, by p o iding a ich and di e se collec ion o open da ase s, LimeSoDa has he po en ial o signi ican ly ad ance ML applica ions in pedome ics. Funding This esea ch was suppo ed by he Lowe Saxony Minis y o Science and Cul u e (MWK), unded h ough he zukun .niede sachsen p og am o he Volkswagen Founda ion (ZN4072) as well as he Fede al Minis y o Educa ion and Resea ch o Ge many (BMBF) h ough he BonaRes p ojec I4S: In elligence o Soil (031B1069A). Compu e esou ces we e unded by he Deu sche Fo schungsgemeinscha (DFG, Ge man Resea ch Founda ion) p ojec numbe 456666331. Decla a ion o compe ing in e es The au ho s decla e ha hey ha e no known compe ing inancial in e es s o pe sonal ela ionships ha could ha e appea ed o in luence he wo k epo ed in his pape . Acknowledgemen s We a e hank ul o a lo o scien is s, ield- and echnical assis an s who helped us o ga he he da a and me ada a o he da ase s. Appendix A:. Li e a u e e iew on DSM benchma king s udies Appendix A.1: Li e a u e e iew me hodology The “Web o Science” da abase was used o he li e a u e e iew. We sea ched o “digi al soil mapping” and “p edic i e soil mapping” as keywo ds in he abs ac o documen s de ined as “da a pape ”, “ea ly access” o “a icle” and es ic ed he sea ch o publica ions om he yea 2023. This esul ed in an ini ial pool o 192 pape s. F om his pool, eigh pape s we e excluded because hey we e no w i en in English, missed DSM con ex o we e inaccessible. Ano he 73 we e emo ed because he abs ac did no indica e ha any s a is ical benchma king was conduc ed, leading o a o al o 111 pape s o he e iew on DSM benchma king p ac ices. As benchma king we conside ed a b oad de ini ion, encompassing any compa ison o compe ing s a is ical me hods based on quan i ied alida ion me ics. This included compa isons o : p edic ion algo i hms, ea u e selec ion me hods, sampling designs, enginee ed ea u es e c. We did no conside he compa ison o senso s o ea u es as benchma king, i he ocus did no lie on he s a is ical enginee ing o he ea u es (e.g., empo al s acks). In some s udies, he benchma king was no he main objec i e o he s udy bu a he supplemen a y in o ma ion. We also conside ed hese as “benchma king s udy” when he compa ison was men ioned in he abs ac . The code and da a a ailabili y in each o he 111 pape s was e alua ed. We di e en ia ed be ween “no sha ed”, “sha ed” and “pa ially sha ed”. The s a emen on ma e ial a ailabili y and any supplemen a y in o ma ion p o ided we e examined alongside keywo ds such as “ eposi o y”, “da ase ”, “Gi Hub”, “Zenodo”, “code” e c. o de e mine he a ailabili y o code o da a. We classi ied any da a o code o be “sha ed” when hey we e di ec ly accessible h ough he supplemen a y in o ma ion o a eposi o y (e.g., Zenodo o Gi Hub). In some code- eposi o ies, da ase s we e no di ec ly p o ided due o es ic i e da ase licenses (e.g., as in LUCAS) bu links o he websi es ha hos ed hese da ase s we e added. This was also conside ed as “sha ed” when i was p o ided in combina ion wi h he da ase p e-p ocessing code. In o he cases, links we e gi en bu hey expi ed since he publica ion o he da ase hos ing webpages we e no in English, which made na iga ion impossible. These cases we e conside ed “no sha ed”. Las ly, i was e alua ed how many da ase s we e used o he benchma king. We classi ied any collec ion o da a based on a cohe en sampling in a dis inc s udy a ea as da ase . Appendix A.2: Li e a u e e iew esul s A majo i y o o e 90 % nei he sha ed he da a no he code (Fig. A1). Howe e , many s udies (48 %) included a s a emen on he da a a ailabili y wi h willingness o p o ide da a on eques . In con as , such a s a emen on he a ailabili y o code was almos ne e included apa om a single excep ion (1 %). This can be explained by he ac ha mos jou nals ask o a s a emen on he a ailabili y o da a bu do no expec his o he code a ailabili y. The majo i y o benchma king was conduc ed based on a single da ase (95.5 %) and he maximum numbe o da ase s did no exceed h ee (Fig. A2). All ma e ials used in he li e a u e e iew a e p o ided a gi hub.com/JonasSchmidinge /LimeSoDa_li e a u e. e iew, wi h addi ional commen s o bounda y cases. J. Schmidinge e al. Geode ma 459 (2025) 117337 8 Fig. A1. Ba plo showing he ela i e equency dis ibu ion on he a ailabili y o da a (a) o code (c) in o he DSM benchma king publica ions om 2023, as well as he s a emen on da a (b) o code (d) a ailabili y i da ase s we e no sha ed. Fig. A2. Ba plo showing he ela i e equency dis ibu ion on how many da ase s we e e alua ed in o he DSM benchma king publica ions om 2023. J. Schmidinge e al. Geode ma 459 (2025) 117337 9 Schmidinge , J., Ba ko , V., Ta akoli, H., Co ea, J., Os e mann, M., A zmuelle , M., Gebbe s, R., Vogel, S., 2024a. Which and how many soil senso s a e ideal o p edic key soil p ope ies: A case s udy wi h se en senso s. Geode ma 450, 117017. Schmidinge , J., Sch ¨ o e , I., B¨ onecke, E., Gebbe s, R., Ruehlmann, J., K ame , E., Mulde , V.L., Heu elink, G.B.M., Vogel, S., 2024b. E ec o aining sample size, sampling design and p edic ion model on soil mapping wi h p oximal sensing da a o p ecision liming. P ecision Ag ic 25 (3), 1529–1555. Shi, X., Song, J., Wang, H., L , X., Zhu, Y., Zhang, W., Bu, W., Zeng, L., 2023. Imp o ing soil o ganic ma e es ima ion accu acy by combining op imal spec al p ep ocessing and ea u e selec ion me hods based on pXRF and is-NIR da a usion. Geode ma 430, 116301. Shmuel, A., Glickman, O., Lazebnik, T., 2024. A Comp ehensi e Benchma k o Machine and Deep Lea ning Ac oss Di e se Tabula Da ase s. h p://a xi .o g/pd / 2408.14817. Shwa z-Zi , R., A mon, A., 2022. Tabula da a: Deep lea ning is no all you need. In . Fusion 81, 84–90. S eige , A., Qaswa , M., Bill, R., Mouazen, A.M., G enzd¨ o e , G., 2025. Compa ing he handheld S enon Fa mLab soil senso wi h a Vis-NIR mul i-senso soil sensing pla o m. Sma Ag ic. Technol. 10, 100717. S obl, C., Leisch, F., 2024. Agains he “one me hod i s all da a se s” philosophy o compa ison s udies in me hodological esea ch. Biom. J. 66 (1), e2200104. Ta akoli, H., Co ea, J., Vogel, S., Gebbe s, R., 2022. RapidMappe – a mobile mul i- senso pla o m o he assessmen o soil e ili y in p ecision ag icul u e, in: , AgEng-Land.Technik 2022. In e na ional Con e ence on Ag icul u al Enginee ing ; No embe 22nd and 23 d 2022, Be lin. VDI-Be ich 2406. VDI Ve lag GmbH, Düsseldo , pp. 351–358. Ta a es, T.R., Molin, J.P., Ja adi, S.H., de Ca alho, H.W.P., Mouazen, A.M., 2020. Combined Use o Vis-NIR and XRF Senso s o T opical Soil Fe ili y Analysis: Assessing Di e en Da a Fusion App oaches. Senso s 21 (1), 148. Valen e, D.S.M., Pe ei a, G.W., Quei oz, D.M. de, Zandonadi, R.S., Ama al, L.R.d., Bo ega, E.L., Cos a, M.M., F ei as Coelho, A.L. de, G i , T., 2024. Accu acy o Va ious Sampling Techniques o P ecision Ag icul u e: A Case S udy in B azil. Ag icul u e 14 (12), 2198. Wadoux, A.-M.-C., Heu elink, G.B., La k, R.M., Lagache ie, P., Bouma, J., Mulde , V.L., Liboho a, Z., Yang, L., McB a ney, A.B., 2021. Ten challenges o he u u e o pedome ics. Geode ma 401, 115155. Wadoux, A.-M.-C., Minasny, B., McB a ney, A.B., 2020. Machine lea ning o digi al soil mapping: Applica ions, challenges and sugges ed solu ions. Ea h Sci. Re . 210, 103359. Wadoux, A.-M.-J.-C., S ump , F., Schol en, T., 2025. A ca chmen -scale da ase o soil p ope ies and hei mid-in a ed spec a. Da a B ie 60, 111510. Wang, J., Zhao, X., Deuss, K.E., Cohen, D.R., T ian a ilis, J., 2022. P oximal and emo e senso da a usion o 3D imaging o in e ile and acidic soil. Geode ma 424, 115972. Webe , L.M., Saelens, W., Cannood , R., Soneson, C., Hap elmeie , A., Ga dne , P.P., Boules eix, A.-L., Saeys, Y., Robinson, M.D., 2019. Essen ial guidelines o compu a ional me hod benchma king. Genome Biol. 20 (1), 125. Weh le, R., Coulouma, G., P¨ a zold, S., 2022. Po able mid-in a ed spec oscopy o p edic pa ame e s ela ed o ca bon s o age in ineya d soils: Model calib a ions unde a ying geopedological condi ions. Biosys . Eng. 222, 1–14. We e lind, J., S enbe g, B., S¨ ode s ¨ om, M., 2010. Inc eased sample poin densi y in a m soil mapping by local calib a ion o isible and nea in a ed p edic ion models. Geode ma 156 (3–4), 152–160. Wilkinson, M.D., Dumon ie , M., Aalbe sbe g, I.J.J., Apple on, G., Ax on, M., Baak, A., Blombe g, N., Boi en, J.-W., Da Sil a San os, L.B., Bou ne, P.E., Bouwman, J., B ookes, A.J., Cla k, T., C osas, M., Dillo, I., Dumon, O., Edmunds, S., E elo, C.T., Finke s, R., Gonzalez-Bel an, A., G ay, A.J.G., G o h, P., Goble, C., G e he, J.S., He inga, J., Hoen, P.A.C. ’ , Hoo , R., Kuhn, T., Kok, R., Kok, J., Lushe , S.J., Ma one, M.E., Mons, A., Packe , A.L., Pe sson, B., Rocca-Se a, P., Roos, M., an Schaik, R., Sansone, S.-A., Schul es, E., Sengs ag, T., Sla e , T., S awn, G., Swe z, M. A., Thompson, M., an de Lei, J., an Mulligen, E., Vel e op, J., Waagmees e , A., Wi enbu g, P., Wols enc o , K., Zhao, J., Mons, B., 2016. The FAIR Guiding P inciples o scien i ic da a managemen and s ewa dship. Sci Da a 3 (1), 160018. W igh , M.N., Ziegle , A., 2017. ange : A Fas Implemen a ion o Random Fo es s o High Dimensional Da a in C++ and R. J. S a . So w. 77 (1), 1–17. ˇ Zíˇ zala, D., P inc, T., Sk´ ala, J., Juˇ ico ´ a, A., Lukas, V., Boho ic, R., Z´ ado o ´ a, T., Minaˇ ík, R., 2024. Soil sampling design ma e s - Enhancing he e iciency o digi al soil mapping a he ield scale. Geode ma Reg. 39, e00874. J. Schmidinge e al. Geode ma 459 (2025) 117337 16