Some Peculiarities of Arable Soil Organic Matter Detection Using Optical Remote Sensing Data

remote sensing

Article Some Peculiarities of Arable Soil Organic Matter Detection Using Optical Remote Sensing Data

Elena Prudnikova 1,2,* and Igor Savin 1,2

1 V. V. Dokuchaev Soil Science Institute, 119017 Moscow, Russia; [email protected] 2 Ecology Faculty, Peoples’ Friendship University of Russia (RUDN University), 117198 Moscow, Russia * Correspondence: [email protected]

Abstract: Optical remote sensing only provides information about the very thin surface layer of soil. Rainfall splash alters soil surface properties and its spectral reflectance. We analyzed the impact of rainfall on the success of soil organic matter (SOM) content (% by mass) detection and mapping based on optical remote sensing data. The subject of the study was the arable soils of a test field located in the Tula region (Russia), their spectral reflectance, and Sentinel-2 data. Our research demonstrated that rainfall negatively affects the accuracy of SOM predictions based on Sentinel-2 data. Depending on the average precipitation per day, the R2cv of models varied from 0.67 to 0.72, RMSEcv from 0.64 to 1.1% and RPIQ from 1.4 to 2.3. The incorporation of information on the soil surface state in the model resulted in an increase in accuracy of SOM content detection based on Sentinel-2 data: the R2cv of the models increased up to 0.78 to 0.84, the RMSEcv decreased to 0.61 to 0.71%, and the RPIQ increased to 2.1 to 2.4. Further studies are necessary to identify how the SOM content and composition of the soil surface change under the influence of rainfall for other soils, and to determine the relationships between rainfall-induced SOM changes and soil surface spectral reflectance.

Citation: Prudnikova, E.; Savin, I. Keywords: rainfall impact; soil surface; soil organic matter; Sentinel-2 Some Peculiarities of Arable Soil Organic Matter Detection Using Optical Remote Sensing Data. Remote Sens. 2021, 13, 2313. https://doi.org/ 1. Introduction 10.3390/rs13122313 Organic matter is an important indicator of soil state as it forms and maintains the Academic Editors: Travis W. Nauman main regimes, properties, and functions of soil [1]. Evaluation of the soil state and quality and Jonathan J. Maynard is based mainly on the content and quality of soil organic matter (SOM) and its fractions: humus, total organic carbon, water soluble carbon, microbial biomass carbon, etc. [2–6]. Received: 14 March 2021 Information on organic matter content is included in soil databases and monitoring systems Accepted: 10 June 2021 from regional to global scales [7–10]. According to a previous review [11], SOM is the Published: 12 June 2021 most frequently used indicator of soil quality. A decrease in SOM content, caused by human-induced activities, is acknowledged to be one of the five global soil threats as it Publisher’s Note: MDPI stays neutral results in the disruption of soil functions, including soil fertility, and regulation of CO2 in with regard to jurisdictional claims in the atmosphere as a main driver of global warming [12,13]. published maps and institutional affil- Optical remote sensing data have been widely used to map the organic matter content iations. in agricultural soils. Numerous studies conducted in various regions and with different types of remote sensing data have demonstrated that SOM content can be detected with different degrees of accuracy [14–23]. Due to short wavelengths, the optical range only provides information about the very thin soil surface layer [24]. Therefore, the measured Copyright: © 2021 by the authors. spectral reflectance of the surface of arable soils in the optical range only contains informa- Licensee MDPI, Basel, Switzerland. tion about the properties of this surface layer. In a previous study, Liang [24] determined This article is an open access article that the sensible optical depth is about 3, which corresponds to a geometric depth of four distributed under the terms and to five times the particle effective radius. For example, for clay particles, the geometric conditions of the Creative Commons depth was 2.2 µm at 0.5 µm and 3.15 µm at 1.105 µm. Attribution (CC BY) license (https:// The moisture content and surface roughness are considered the main soil-related creativecommons.org/licenses/by/ limiting factors when explaining the low performance of remote sensing-based SOM 4.0/).

Remote Sens. 2021, 13, 2313. https://doi.org/10.3390/rs13122313 https://www.mdpi.com/journal/remotesensing Remote Sens. 2021, 13, 2313 2 of 26

detection models [25–29]. However, one more important interfering factor that is rarely considered or discussed is the change in bare soil surface properties such as the soil mineralogical composition, soil texture, and SOM content and chemical composition under the impact of rainfall. Rainfall splash causes the breakdown of soil surface aggregates and the redistribution of formed soil material which is partially transported as a lateral surface flow and partially washed in soil surface pores, causing the sealing of the soil surface and the formation of the soil crust [30–34]. Studies on the impact of rainfall on the soil surface show that it results in changes in the soil surface organic matter content, mineral composition, and texture as well as soil surface reflectance [35–41]. The longer the surface of non-irrigated arable soils is left untreated under the influence of atmospheric precipitation, the more its spectral properties and material composition change and the greater the difference between its properties and the properties of the underlying part of the arable horizon becomes [40,42]. Most studies using optical satellite data or field spectral reflectance data to map SOM content are based only on a single date, and they do not register the soil surface state at the time of spectral data acquisition. However, the extent to which the soil surface transformed by rainfall provides information on the properties of the arable horizon is not yet clear. This discrepancy is one of the main reasons why models obtained in the laboratory fail in usage when applied to field measurements. It also limits the reproducibility of models predicting the SOM content based on satellite data. The aim of the research was to study the impact of rainfall-induced soil surface spectral changes, in particular, the formation of soil crust, on the accuracy of SOM content mapping based on optical remote sensing data.

2. Materials and Methods 2.1. Study Area and Field Data Collection Remote Sens. 2021, 13, 2313 3 of 27 The study area is a test field located in the Yasnogorsky district of the Tula region (Russia) with grey forest arable soil (Albic Luvisols [43]) (Figure1). The area of the test field isusing 13.5 radiance ha. The reflected test plot from is located a sample on and a flat radiance part of incidence the slope, reflected 2 to 3 degreesfrom a 100% of the re- western aspect.flectance In reference 2019, the target field (white was reference complete panel) fallow, [47]. which The white is a reference part of crop panel rotation. was po- Such a fieldsitioned is free horizontally of crops near during the thesample growing and was season. oriented It isto treatedreceive all throughout of the available the warm inci- period ofdent the illumination year, maintaining while at the same soil intime a looseavoiding state, shadowing carrying and out light extermination reflected from measures againstsurrounding weeds, objects. pests Since and pathogensthe reflected of radiance crops. Periodicalfrom the panel tilling is very as a close part ofto completethe total fallow cultivationenergy falling agrotechnology on the sample, this is performed reading can tobe exterminateused in the denominator emerging weedsof the reflec- and to break downtance measurement. the soil crust toSample accumulate reflectance soil was moisture calculated for the by nextHandHeld-2 crop and instrument reduce the soft- risk of soil ware. The location of each sample point was registered by GPS attached to the spectrora- degradation development. While this practice has some disadvantages, it is still widely diometer (GlobalSat BU-353 GLONASS: https://www.globalsat.ru/catalog/bu-353-glonass used in Russia, especially in dry regions, to prepare soil for the sowing of winter crops [44]. (accessed on 25 May 2021)).

FigureFigure 1. GeographicalGeographical location location of the of thetest testplot plotand sa andmpling sampling scheme: scheme: the boundaries the boundaries of the test of field the test ﬁeld are delineated in red color. are delineated in red color. In this research, we also used additional spectral data collected from this test field in our previous study: spectral reflectance of the soil crust (Figure 2a) and spectral reflectance of the sample of the ploughed soil layer. Spectral reflectance of these two additional samples was also measured with a HandHeld-2 field spectroradiometer, the same way as spectral reflectance of dry surface and wet subsurface layer as described above. The soil sample of the ploughed soil layer represents a soil surface that forms in the field after the ploughing and has not experienced the impact of rainfall yet.

2.2. Soil Crusting Process There exist three main types of soil crust: physical, biological and chemical [48]. In our research, we are dealing with the physical soil crust, which is a very thin layer formed as a result of the breakdown of soil surface aggregates under the impact of atmospheric precipitation and the redistribution of the formed soil material. According to our previous research on the formation of soil crust and data acquired in micromorphological and microtomographic analysis, the thickness of the well-developed soil crust formed on soils of the European part of Russia is about 1 to 2 mm, and it varies spatially (Figure 2b,c) [41]. We think that it would be more correct to call such a thin layer a microlayer. The degree of soil crust development depends on how long the soil surface is bare, undisturbed by agricultural practices, and on the intensity of rainfall events as well as on a number of soil properties (soil texture, salinity, SOM content, percentage of calcium carbonate, percentage of exchangeable sodium) [48,49]. Moreover, the soil crust is comprised of heterogenous soil material (residues of soil aggregates and washed soil material, as well as, in some cases, water-resistant soil aggregates) (Figure 2a).

Remote Sens. 2021, 13, 2313 3 of 26

During field work conducted on the test field on 15 August 2019, we collected 30 mixed soil samples from the upper part of the arable horizon (0 to 5 cm). Each mixed sample consisted of five single sub-samples of the upper part of the arable horizon collected from the area around the sample point (approximately 1 m). Sub-samples were collected using the diagonal method at five points. We used the simple random sampling pattern. The sampling scheme is presented in Figure1. SOM content was determined in the collected soil samples as a percentage of mass by the Tyurin method [45], which is the analogue of the Walkley–Black method [46]. The spectral reflectance of the dry surface and wet subsurface (2 to 5 cm depth) layer was measured in the field at the sample points with a HandHeld-2 field spectroradiometer (proximal data) operating in the range of 325 to 1075 nm with steps of 1 nm. To measure the spectral reflectance of the soil subsurface layer, we removed 2 cm of the soil surface after measuring the soil surface spectral reflectance. At each sample point, the spectral reflectance was measured with fivefold repetition. Reflectance measurements were made using radiance reflected from a sample and radiance incidence reflected from a 100% reflectance reference target (white reference panel) [47]. The white reference panel was positioned horizontally near the sample and was oriented to receive all of the available incident illumination while at the same time avoiding shadowing and light reflected from surrounding objects. Since the reflected radiance from the panel is very close to the total energy falling on the sample, this reading can be used in the denominator of the reflectance measurement. Sample reflectance was calculated by HandHeld-2 instrument software. The location of each sample point was registered by GPS attached to the spectroradiometer (GlobalSat BU-353 GLONASS: https://www.globalsat.ru/catalog/bu-353-glonass (accessed on 25 May 2021)). In this research, we also used additional spectral data collected from this test field in our previous study: spectral reflectance of the soil crust (Figure2a) and spectral reflectance of the sample of the ploughed soil layer. Spectral reflectance of these two additional samples was also measured with a HandHeld-2 field spectroradiometer, the same way as spectral reflectance of dry surface and wet subsurface layer as described above. The soil sample of the ploughed soil layer represents a soil surface that forms in the field after the ploughing and has not experienced the impact of rainfall yet. There exist three main types of soil crust: physical, biological and chemical [48]. In our research, we are dealing with the physical soil crust, which is a very thin layer formed as a result of the breakdown of soil surface aggregates under the impact of atmospheric precipitation and the redistribution of the formed soil material. According to our previous research on the formation of soil crust and data acquired in micromorphological and microtomographic analysis, the thickness of the well-developed soil crust formed on soils of the European part of Russia is about 1 to 2 mm, and it varies spatially (Figure2b,c) [ 41]. We think that it would be more correct to call such a thin layer a microlayer. The degree of soil crust development depends on how long the soil surface is bare, undisturbed by agricultural practices, and on the intensity of rainfall events as well as on a number of soil properties (soil texture, salinity, SOM content, percentage of calcium carbonate, percentage of exchangeable sodium) [48,49]. Moreover, the soil crust is comprised of heterogenous soil material (residues of soil aggregates and washed soil material, as well as, in some cases, water-resistant soil aggregates) (Figure2a). Remote Sens. 2021, 13, 2313 4 of 26 Remote Sens. 2021, 13, 2313 4 of 27

(a) (b)

(c)

FigureFigure 2. 2. SoilSoil crust crust on onthe thesurface surface of the of studied the studied soil: (a soil:) photo (a) of photo soil crust of soil of grey crust forest of grey soil forest collected soil collected on watershed of the test field. The photo was made on an iPhone XR located above the soil surface; on watershed of the test ﬁeld. The photo was made on an iPhone XR located above the soil surface; (b) microstructure of vertical thin section of upper 5 mm by an Olympus BX51 polarizing microscope(b) microstructure with an Olympus of vertical DP26 digital thin sectioncamera; of (c) upper model 5 of mm micromonolith by an Olympus of upper BX51 4 polarizingcm soil layer microscope showingwith an inner Olympus structure DP26 and digital pore space. camera; The ( modec) modell was of produced micromonolith by X-ray ofmicrotomography upper 4 cm soil using layer showing ainner SkyScan structure 1172G. and Red pore lines space. on (b,c The) delineate model the was boundary produced between by X-ray soil microtomography crust and the underlying using a SkyScan layer. 1172G. Red lines on (b,c) delineate the boundary between soil crust and the underlying layer.

2.2.The Soil studied Crusting soil Process did not contain calcium carbonate or water-soluble salts, and the percentage of exchangeable sodium is very low (Table 1). Therefore, the main soil-related factorsThe affecting studied soil soilcrust did formation not contain for the calciumstudied soil carbonate are SOM orcontent water-soluble and soil texture. salts, and the Aspercentage SOM acts ofas exchangeablean aggregate stabilizing sodium agent, is very its low content (Table affects1). Therefore, the speed and the degree main soil-related of soilfactors crust affecting development. soil crustSoil texture formation predeter formines the studied the mechanism soil are of SOM soil contentcrust formation. and soil texture. As SOM acts as an aggregate stabilizing agent, its content affects the speed and degree of Table 1. Chemical propertiessoil of crust grey forest development. soils of the test Soil field texture (the data predetermines for 30 samples collected the mechanism on 15 August of soil2015). crust formation. Exchangeable Cations Table 1. Chemical propertiespH of Salt of grey Mobile forest Mobile soils of theK SOM test ﬁeld (the data for 30 samples collected on 15 August 2015). Descriptive Statistics Ca++ Mg++ Extract P(mg/kg) (mg/kg) (%) Na+ (mg/kg) K+ (mg/kg) (mmol/100 g) (mmol/100Exchangeable g) Cations Descriptive pH of Salt Mobile P Mobile K Minimum 4.9 85.2 141.6 SOM1.5 (%) 9.7 ++ 1.9 ++ 0.0 +0.3 + Statistics Extract (mg/kg) (mg/kg) Ca Mg Na K Maximum 6.9 3541.4 395.0 7.3 47.6(mmol/100 g) 3.3 (mmol/1000.1 g) (mg/kg)0.8 (mg/kg) MinimumAverage 4.95.3 85.2 536.0 141.6227.0 3.2 1.5 19.6 9.72.7 1.90.1 0.00.4 0.3 MaximumStandard deviation 6.9 0.6 3541.4 943.3 395.0 61.8 1.5 7.3 10.2 47.6 0.4 3.3 0.0 0.1 0.1 0.8 Average 5.3 536.0 227.0 3.2 19.6 2.7 0.1 0.4 Standard 0.6 943.3 61.8 1.5 10.2 0.4 0.0 0.1 deviation Coefﬁcient of 11.9 176.0 27.2 46.1 52.4 15.0 26.2 29.6 variation (%) Remote Sens. 2021, 13, 2313 5 of 27

Coefficient of variation 11.9 176.0 27.2 46.1 52.4 15.0 26.2 29.6 (%)

The analyzed soil was formed on loess. The composite sample collected from the

Remote Sens. 2021, 13, 2313 ploughed horizon (0 to 30 cm) of the studied field using the diagonal5 of 26 method contained 6.5% of sand (0.05 to 1 mm), 75.6% of silt (0.001 to 0.05 mm) and 17.9% of clay (< 0.001 mm) [50]. This is silt loam in the USDA classification. The conversion from the Russian classificationThe analyzed to soil the was USDA formed classification on loess. The composite was made sample based collected on [51]. from the ploughedThe horizon impact (0 to of 30 atmospheric cm) of the studied precipitation field using the diagonal on soils method with contained such a texture 6.5% causes the break- of sand (0.05 to 1 mm), 75.6% of silt (0.001 to 0.05 mm) and 17.9% of clay (<0.001 mm)[50]. Thisdown is silt of loam soil in surface the USDA aggregates, classification. Theand conversion the formed from thematerial Russian is classification washed and moved into the poresto the USDA of the classification soil surface. was madeIn Figure based on 2b, [51 this]. washed material has a light brown color, while the darkThe impact brown of atmosphericcolor corresponds precipitation to onthe soils remains with such of the a texture soil aggregates. causes the In the field, when breakdown of soil surface aggregates, and the formed material is washed and moved into wethe poreslook of at the the soil surface surface. Inof Figure the 2soilb, this crust, washed the material washed has asoil light material brown color, has a very light color, whileclose the to darkwhite brown (Figure color corresponds 2a). The formation to the remains of of the the soil aggregates. crust in the In the field field, goes through certain whenstages. we lookThey at thehave surface been of thedefined soil crust, micromorpholog the washed soil materialically has through a very light the color, analysis of changes in close to white (Figure2a). The formation of the soil crust in the field goes through certain thestages. microstructure They have been defined [36]. micromorphologically through the analysis of changes in the microstructureIn 2017, we [36 ].conducted a model experiment, wherein we exposed the samples of ploughedIn 2017, horizons we conducted of three a model soil experiment, types, including wherein we the exposed studied the samplessoil from of the test field, to the ploughed horizons of three soil types, including the studied soil from the test field, to the impact of of atmospheric atmospheric precipitation precipitation [40,41]. We [40,41]. put soil samplesWe put in rectangularsoil samples plastic in rectangular plastic boxes and and left left them them outside, outside, periodically periodically taking photos ta ofking the soil photos surface of (Figure the so3).il surface (Figure 3).

(a) (b) (c) (d)

(e) (f) (g) Figure 3. 3.Photos Photos of soilof surfacesoil surface of grey of forest grey soil forest in model soil experiment in model [40 experiment]: (a) 7.06.2017 [40]: (the (a) 7.06.2017 (the start ofstart the of theexperiment); experiment); (b()b 16.06.2017;) 16.06.2017; (c) 26.06.2017; (c) 26.06.2017; (d) 30.06.2017; (d) (30.06.2017;e) 07.07.2017; (f()e 11.07.2017;) 07.07.2017; (f) 11.07.2017; (g) (g) 25.07.2017. 25.07.2017. According to our results, which are in accordance with [33], the longer bare soil surface was subjectedAccording to rainfall, to our the lessresults, soil aggregates which are were in left accordance there and the morewith washed [33], the soil longer bare soil sur- material formed as a result of the breakdown of soil surface aggregates (Figures3 and4 ). Therefore,face was the subjected amount of washedto rainfall, soil material the less on thesoil soil ag surfacegregates can be were used as left an indicatorthere and the more washed soilof soil material crust development. formed Asas thea result soil surface of the on 11breakdown July 2017 was of wet soil (Figure surface3f), the aggregates area (Figures 3 and 4).of washed Therefore, soil material the wasamount underestimated of washed due to soil the interferencematerial ofon soil the moisture soil contentsurface can be used as an (Figure4). The cracks did not appear after the ﬁrst rainfall. We registered the ﬁrst cracks indicatorafter the cumulative of soil rainfallcrust reacheddevelopment. 100 mm (Figure As the4). soil surface on 11 July 2017 was wet (Figure 3f), the area of washed soil material was underestimated due to the interference of soil moisture content (Figure 4). The cracks did not appear after the first rainfall. We registered the first cracks after the cumulative rainfall reached 100 mm (Figure 4). Rainfall-induced changes in the soil surface microlayer affect the soil spectral reflectance. For example, the soil crust of the studied soil (Figure 2a), collected from the watershed of the test field in our previous studies, had higher reflectance intensity compared to the sample of the ploughed soil layer (Figure 5). The difference in spectral reflectance increased with the increase in wavelength. Moreover, the average spectral reflectance of dry surface measured in the field on 15 August 2019 is very close to the spectral reflectance of

Remote Sens. 2021, 13, 2313 6 of 27

the sample of the ploughed soil layer (Figure 5). It shows that tillage occurred close to the date of our field trip, meaning that the upper layer was rather uniform. Thus, the main Remote Sens. 2021, 13, 2313 6 of 26 difference between the soil surface microlayer and subsurface layer on 15 August 2019 was due to the moisture content.

Remote Sens. 2021, 13, 2313 6 of 27

the sample of the ploughed soil layer (Figure 5). It shows that tillage occurred close to the date of our field trip, meaning that the upper layer was rather uniform. Thus, the main difference between the soil surface microlayer and subsurface layer on 15 August 2019 was due to the moisture content.

FigureFigure 4. 4.Change Change in in soil soil surface surface status status of grey of grey forest forest soil in soil model in experiment;model experiment; R1—the amountR1—the of amount of rainfallrainfall from from the the last last measurement, measurement, R2 —accumulated R2 —accumulat rainfalled rainfall from the from start the of thestart experiment. of the experiment. The The projectedprojected area, area, occupied occupied by washedby washed soil material,soil material, was determined was determined from photos from ofphotos the soil of surface the soil surface takentaken during during the the experiment experiment [40]. [40]. The The information information on rainfall on rainfall was collected was collected from the from meteorological the meteorological stationstation near near the the experiment experiment site site (http://www.pogodaiklimat.ru/weather.php?id=27605 (http://www.pogodaiklimat.ru/weather.php?id=27605 (accessed (accessed on on 2525 May May 2021)). 2021)).

Rainfall-induced changes in the soil surface microlayer affect the soil spectral reflectance. For example, the soil crust of the studied soil (Figure2a), collected from the watershed of the test field in our previous studies, had higher reflectance intensity compared to the sample of the ploughed soil layer (Figure5). The difference in spectral reflectance increased with the increase in wavelength. Moreover, the average spectral reflectanceFigure 4. Change of dry in soil surface surface measured status of grey in theforest field soil onin model 15 August experiment; 2019 R1—the is very amount close toof the spectralrainfall from reflectance the last measurement, of the sample R2 of—accumulat the plougheded rainfall soil layerfrom the (Figure start of5). the It showsexperiment. that The tillage occurredprojected area, close occupied to the date by washed of our fieldsoil material, trip, meaning was determined that the upperfrom photos layer of was the rather soil surface uniform. taken during the experiment [40]. The information on rainfall was collected from the meteorological Thus, the main difference between the soil surface microlayer and subsurface layer on station near the experiment site (http://www.pogodaiklimat.ru/weather.php?id=27605 (accessed on 1525 AugustMay 2021)). 2019 was due to the moisture content.

Figure 5. Spectral reflectance of soil samples with different state of soil surface. SR_ds and SR_ws— average spectral reflectance of dry soil surface and wet subsurface layer of 30 sample points, respectively, measured in the field on 15 August 2019; SR_sc—spectral reflectance of soil crust; SR_pl— spectral reflectance of sample of ploughed soil layer.

Soil crust is the initial stage of soil degradation. This microlayer is comprised of the remains of soil aggregates, so its original structure is lost. As it is compacted, the water penetration decreases while the surface runoff increases. The latter adds up to soil erosion FigureFiguredevelopment. 5. SpectralSpectral reflectance reflectanceThe loss of of soil soilaggr samples samplesegate with withstructure, diffe differentrent state compaction, state of soil of soil surface. surface. and SR_ds SR_dsincreased and SR_ws— and SR_ws— surface runoff averageaveragecontribute spectral to reflectance reflectancethe proliferation of of dry dry soil soil surface of surface soil and degradation. and wet wet subsurface subsurface layer layer of 30 of sample 30 sample points, points, respec- respec- tively, measured in the field on 15 August 2019; SR_sc—spectral reflectance of soil crust; SR_pl— tively, measured in the field on 15 August 2019; SR_sc—spectral reflectance of soil crust; SR_pl— spectral reflectance of sample of ploughed soil layer. spectral reflectance of sample of ploughed soil layer. Soil crust is the initial stage of soil degradation. This microlayer is comprised of the remains of soil aggregates, so its original structure is lost. As it is compacted, the water penetration decreases while the surface runoff increases. The latter adds up to soil erosion development. The loss of aggregate structure, compaction, and increased surface runoff contribute to the proliferation of soil degradation.

Remote Sens. 2021, 13, 2313 7 of 26

Soil crust is the initial stage of soil degradation. This microlayer is comprised of the remains of soil aggregates, so its original structure is lost. As it is compacted, the water Remote Sens. 2021, 13, 2313 7 of 27 penetration decreases while the surface runoff increases. The latter adds up to soil erosion development. The loss of aggregate structure, compaction, and increased surface runoff contribute to the proliferation of soil degradation. 2.3. General Strategy of the Research 2.3. GeneralFirst of Strategyall, we started of the Research with the development of the model for SOM content prediction fromFirst the data of all, collected we started in the with field the on development 15 August 2019 of the (Figure model 6). for For SOM that, content we used prediction spectral fromreflectance the data of the collected dry soil in surface the field microlayer. on 15 August As 2019a result, (Figure we determined6). For that, informative we used spectral spec- reflectancetral parameters of the that dry can soil be surface used microlayer.for the mode Aslling a result, of SOM we determinedin the upper informative layer of the spec- test tralfield parameters from optical that remote can be usedsensing for thedata. modelling Detailed of description SOM in the is upper given layer in the of theSections test field 2.6 fromand 2.7. optical remote sensing data. Detailed description is given in the Sections 2.6 and 2.7.

FigureFigure 6.6. ProcessProcess flow flow chart: chart: (1) (1) modelling modelling of of SOM SOM content content from from the fieldthe field data data and selectionand selection of Sentinel-2 of SenTable images 2. withimages dry with soil surface;dry soil (2)surface; modelling (2) modelling SOM content SOM from content Sentinel-2 from Sentinel-2 images with images dry soil with surface; dry soil (3) surface; accounting (3) accounting for soil surface for soil state surface when state when modelling SOM content from Sentinel-2 data for the dates with dry soil surface. SR_ds and SR_ws—spectral modelling SOM content from Sentinel-2 data for the dates with dry soil surface. SR_ds and SR_ws—spectral reflectance of reflectance of dry soil surface and wet subsurface layer, respectively, measured in the field on 15 August 2019; SR_sc— dry soil surface and wet subsurface layer, respectively, measured in the field on 15 August 2019; SR_sc—spectral reflectance spectral reflectance of soil crust, SR_pl—spectral reflectance of sample of ploughed soil layer. of soil crust, SR_pl—spectral reflectance of sample of ploughed soil layer. As soil moisture can negatively affect the accuracy of SOM content prediction as well As soil moisture can negatively affect the accuracy of SOM content prediction as well as result in the underestimation of the soil crust area, we used spectral reflectance of the as result in the underestimation of the soil crust area, we used spectral reflectance of the dry soil surface and wet subsurface layer measured inin thethe fieldfield onon 1515 August 20192019 toto select Sentinel-2 images with the dry soilsoil surface.surface. The selection was based on thethe intensityintensity ofof spectral reflectancereflectance and its closeness to the spectralspectral reflectancereflectance ofof thethe drydry surfacesurface andand wetwet subsurface layer. Detailed descriptiondescription isis givengiven inin thethe SectionSection 2.82.8.. At the next step, we modeled SOM content from Sentinel-2 images forfor thethe datesdates withwith the dry soil surface and estimated the accuracy of the developed models. For all the dates, we usedused the the same same predictors—the predictors—the informative informative spectral spectral parameter parameter (parameters) (parameters) determined deter- duringmined during the analysis the analysis of spectral of spectral reflectance reflecta of thence dry of soilthe dry surface soil measuredsurface measured in the field in the on 15field August on 15 August 2019. We 2019. assumed We assumed that difference that differ inence the accuracy in the accuracy of models of models at different at different dates fromdates thefrom model the model based based on spectral on spectral reflectance reflectance of the of dry the soildry surfacesoil surface was was caused caused by the by differencethe difference in the in statethe state of the of soilthe soil surface surface microlayer microlayer on theseon these dates date ands and onthe on datethe date of the of fieldthe field trip trip in 2019.in 2019. We We looked looked at at the the average average spectral spectral reflectance reflectance for for eacheach analyzedanalyzed date, spectral reflectance of spectral mixtures with different percentages of soil crust and the rainfall data to support our assumption. A detailed description is given in the Section 2.9. Finally, we accounted for the changes in the soil surface microlayer in models for SOM content estimation by adding as predictors the area of the soil crust and error of spectral unmixing, indicating the area of shadows (including cracks). These parameters

Remote Sens. 2021, 13, 2313 8 of 26

spectral reflectance of spectral mixtures with different percentages of soil crust and the rainfall data to support our assumption. A detailed description is given in the Section 2.9. Finally, we accounted for the changes in the soil surface microlayer in models for SOM content estimation by adding as predictors the area of the soil crust and error of spectral unmixing, indicating the area of shadows (including cracks). These parameters were determined for each date with a dry soil surface as a result of spectral unmixing. A detailed description is given in the Section 2.10. As at the time of field trip in 2019, there was not a soil crust on the soil surface and we did not catch the soil surface immediately after the ploughing; we used additional spectral data, collected during our previous studies and described in Section 2.1, representing these two soil surface states (spectral reflectance of soil crust and of sample of ploughed soil layer) to model spectral mixtures with different percentages of soil crust and to perform spectral unmixing.

2.4. Sentinel-2 Data Collection and Preprocessing Seven Sentinel-2 scenes were chosen for the studied region: 2 April 2019, 17 April 2019, 20 April 2019, 5 May 2019, 6 June 2019, 19 June 2019, and 28 August 2019 (https: //earthexplorer.usgs.gov/ (accessed on 25 May 2021)) (Table2). Atmospheric correction of the chosen scenes was performed with Sen2Cor in SNAP. Afterwards, we imported all of the scenes in ILWIS 3.3 Academic [52] and extracted the reﬂectance values for pixels where we had collected spectral data and soil samples in the ﬁeld.

Table 2. Sentinel-2 band characteristics.

Sentinel-2 Bands Central Wavelength (nm) Band Width (nm) Resolution (m) Band 1—coastal aerosol 443 20 60 Band 2—blue 490 65 10 Band 3—green 560 35 10 Band 4—red 665 30 10 Band 5—vegetation red edge 705 15 20 Band 6—vegetation red edge 740 15 20 Band 7—vegetation red edge 783 20 20 Band 8—near infrared 842 115 10 Band 8a—vegetation red edge 865 20 20 Band 9—water vapor 945 20 60 Band 10—shortwave infrared—Cirrus 1375 30 60 Band 11—shortwave infrared 1610 90 20 Band 12—shortwave infrared 2190 180 20

2.5. Meteorological Data Collection and Analysis Meteorological data for 2019 for the test ﬁeld were downloaded from the web-service Vega (http://pro-vega.ru/ (accessed on 25 May 2021)) (reanalysis data). The data were used to study rainfall patterns in between dates of selected Sentinel-2 data acquisition. We also calculated the average precipitation per day (APD) for the periods between image acquisition dates as the ratio of atmospheric precipitation accumulated during the period to the length of the period in days. The APD was further used as a proxy of the degree of impact of rainfall on the bare soil surface.

2.6. The Preprocessing and Analysis of Field Spectral Data All spectral reflectance curves used in this study were averaged and smoothed with the Savitzky–Golay function of the prospectr package of R software [53]. Due to the high level of noise, regions shorter than 350 nm and longer than 900 nm were excluded from further analysis. After that, smoothed field spectral reflectance was convolved into Sentinel-2 bands assuming Gaussian distribution with the spectralResampling function of the hsdar package of R software [54] by analogy with [55]. Considering the spectral range of the HandHeld-2 Remote Sens. 2021, 13, 2313 9 of 26

spectroradiomener, as a result of spectral generalization of the field spectral data, we had spectral reflectance values only for nine Sentinel-2 bands (from Band 1 to Band 8a) (Table2). All further analysis and calculations were made for spectrally generalized data. As the soil moisture content has a masking effect on SOM spectral features [26,29], we started the analysis with spectral reflectance of the dry surface. Using these spectrally generalized spectral reflectance data, we calculated a number of spectral indices and ratios: spectral ratios of analyzed Sentinel-2 bands, BI, SI, HI, CI, RI, AVI, GSI, CRUST, BSCI, NDVI, SAVI, EVI2, and NDWI [56–64] (Table A1). Then, we checked the correlations between the calculated parameters to reduce the number of predictors for further regression modelling and to avoid multicollinearity. Only the indices and ratios with correlations between each other lower than 0.6 were further used in linear regression modelling. The selection of such parameters was made in R software using the function vifcor of the usdm package with a threshold of 0.6 [65].

2.7. Modelling SOM Content with Transformed Field Spectral Data We used the parameters selected in the previous step to model SOM content with the linear regression method. First, all of the selected parameters were included in the model. Then, step by step, we excluded the parameters with a p-value above 0.05 until the model included only the parameters with a p-value below 0.05. Regression modelling was performed with the lm function of the stats package [66]. To estimate the accuracy and quality of the obtained models, we repeated 10-fold cross-validation 100 times and calculated the cross-validated coefficient of determination (R2cv), the cross-validated root-mean-squared error (RMSEcv), and the ratio of the interquartile range to the error of prediction (RPIQ). The model developed from spectral reflectance of the dry soil surface with the highest R2cv and RPIQ and the lowest RMSEcv was considered the best, and the parameter (or parameters) included in such a model was considered the most informative parameter for the prediction of SOM content. Cross-validation was performed with the package caret [67] by analogy with the example presented in (http://www.sthda.com/english/articles/38-regression-model-validation/ 157-cross-validation-essentials-in-r/ (accessed on 25 May 2021)). The same informative parameter (parameters) was used to model SOM content from spectral reflectance of the wet subsurface layer. This model was also cross-validated. We also checked for spatial autocorrelation of the residuals for models from both spectral reflectance of the dry surface and wet subsurface layer to ensure that it did not affect our results. For this, we used a semivariogram by analogy with [68]. First, we calculated the residuals for the best regression model with the stats package [66]. Then, we added this information to the attribute table with coordinates of sample points in ILWIS Academic 3.3 and calculated the semi-variance with a lag spacing of 10 m. The semivariogram was created in Microsoft Excel.

2.8. Assessment of Soil Surface Status at the Time of Satellite Data Acquisition As soil moisture has a masking effect on the spectral features of other soil properties (including SOM content and mineralogy), its interference can result in underestimation of the soil crust area. In our model experiment in 2017, described in Section 2.2, for the date with a wet soil surface (Figure3f), we registered a signiﬁcant drop in the area of washed soil material based on the analysis of soil surface photos (Figure4). As the soil surface of the soil sample was kept undisturbed during the whole experiment, such a one-time decrease in the detected area of washed soil material was caused by the masking effect of soil moisture. Therefore, we decided to exclude Sentinel-2 images from further analysis where soil surface moisture content may interfere signiﬁcantly. We suggested that the soil surface microlayer (upper 1 to 2 mm) is most often dry at the times when satellite images in an optical domain are acquired, except on the days when images were acquired after rainfall and if soils are hydromorphic (with the capillary rise from ground water). As recent research has shown, the soil surface microlayer dries out quicker than Remote Sens. 2021, 13, 2313 10 of 26

the underlying soil horizon [69]. Therefore, to correctly estimate the soil moisture content in the soil surface microlayer, it is necessary to collect samples that include the material only from that microlayer. Otherwise, the results will be unreliable. This is a considerably challenging task. In [70], the authors collected soil samples from the very first mm of the soil surface for the analysis of soil moisture content, but they pointed out that it is not always possible to collect only this thin layer without including the material of the underlying soil horizon, which can result in the overestimation of the soil moisture content. Unfortunately, they did not describe how they did this. Considering the above, for the current research, we decided to choose Sentinel-2 images with a dry soil surface microlayer based on the intensity of spectral reflectance. We based our assessment on spectral reflectance of the dry surface and wet subsurface layer collected in the field, which we subsequently named “dry” and “wet”. After a rain event, the soil surface is wet. Then, it starts drying, and the spectral reflectance curve corre- spondently changes. At the moment when the spectral reflectance curve stops changing, the soil surface becomes dry. We created a graph wherein we positioned sample points according to their spectral reflectance values measured in the field (spectral reflectance of dry surface and wet subsurface layer) as well as extracted from all Sentinel-2 images. Axes were chosen from informative Sentinel-2 bands included in the regression model to predict SOM content, which was developed in the previous step (Section 2.6). In the studied spectral range, soil moisture mainly affects the intensity of spectral reflectance in the optical range, meaning that wetter soils have lower spectral reflectance while drier soils have higher spectral reflectance. Thus, on the graph, they occupy two different regions. Dates when the spectral reflectance of sample points extracted from Sentinel data was close to the spectral reflectance of wet subsurface layer were excluded from further analysis. Otherwise, the image was included for further study. We also looked at meteorological data for the test field for 2019. For each date, we mainly checked when the closest rainfall was and its amount. Nocita et al. 2012 [71] showed that the application of the model developed based on spectral reflectance of dry soil to predict organic carbon content in wet soil with a moisture content from 0.05 to more than 0.15 gW gS−1 results in a significant drop in model accuracy. We also checked this in our data. For this, SOM content for each selected Sentinel-2 image was predicted from models for both the dry surface and wet subsurface layer developed from field spectral data. The RMSE calculated for each date for each set of models was used as an indicator of model accuracy.

2.9. Assessment of the Influence Rainfall-Induced Soil Surface Changes on Its Spectral Reflectance and on the Modelling of SOM Content (Only for the Dates with Dry Soil Surface) The influence of rainfall on a bare soil surface results in the development of a soil crust. Therefore, the longer the soil surface experiences the impact of rainfall, the more washed soil material will be found on its surface and the more developed the soil crust will be. Spectral reflectance and other properties will also change (Figure5). We considered the spectral reflectance of the sample of the ploughed soil layer as a starting point and the spectral reflectance of the soil crust as the finishing point of soil surface rainfall-induced changes (Figure5). The spectral reflectance between these two points is a spectral mixture of spectral reflectance of the soil surface with and without a soil crust. We calculated the spectral reflectance of the spectral mixtures with different fractions of soil crust, using a model for the linear spectral mixture:

SRss = a ∗ SR_pl + (1 − a) ∗ SR_sc, (1)

where SRss is the modeled spectral reflectance of spectral mixture in Sentinel-2 bands, SR_pl is the spectral reflectance of the sample of the ploughed soil layer convolved into Sentinel-2 bands, SR_sc is the spectral reflectance of the soil crust convolved into Sentinel-2 bands, a is the fraction of the soil surface without soil crust, and (1 − a) is the fraction of the soil crust. Remote Sens. 2021, 13, 2313 11 of 26

Then, for each sample point, we created spectral curves from the spectral reflectance data extracted from Sentinel-2 images. Afterwards, for each point, we calculated the deviation of its spectral reflectance from the spectral reflectance of the dry surface microlayer measured in the field on 15 August 2019 and averaged the results by the date of Sentinel-2 data acquisition. We also calculated the deviation between the spectral reflectance of the dry soil surface microlayer measured in the field and spectral reflectance of the spectral mixtures with various fractions of the soil crust to relate the spectral deviation for each Sentinel-2 date to the degree of rainfall-induced soil surface changes. The higher the deviation of spectral reflectance for a certain date was from the field spectral reflectance of dry surface microlayer, the more soil surface on this date differed from the soil surface at the time of field data collection. The spectral reflectance of spectral mixtures with different fractions of the soil crust as well as the APD was used to relate the observed differences to the soil surface transformation under the impact of rainfall. Then, we modeled the SOM content for each Sentinel-2 image using informative parameters obtained from the field spectral data analysis (Section 2.6). We further analyzed how the deviations in the spectral reflectance of the soil surface at different dates of Sentinel-2 image acquisition from the spectral reflectance of dry soil surface microlayer measured in the field on 15 August 2019 impacted the parameters, accuracy, and quality of the developed models. The comparison of the models’ accuracy and quality was based on parameters such as the R2cv, RMSEPv, and RPIQ (see Section 2.7). We also checked for the spatial correlation of the residuals in the same way as described in Section 2.7.

2.10. Accounting for Soil Surface State When Mapping SOM Content Based on Sentinel-2 Data As rainfall splash results in the formation of the soil crust, we need to know its area to account for the soil surface state on a particular date. As we excluded dates with a wet soil surface microlayer (see Section 2.8), we assumed the soil surface microlayer to be dry on the rest of the images and the impact of soil moisture content on spectral reflectance, if there was any, to be minimal. To determine the area of the soil crust, we performed linear spectral unmixing (function unmix of package hsdar of R software) for each analyzed Sentinel-2 image. As endmembers, we used spectral reflectance of the soil crust and soil sample of the ploughed horizon (Figure5) convolved into Sentinel-2 bands. A least square solution was performed during spectral unmixing. The error representing the difference between the image spectral reflectance and spectral reflectance predicted by the linear spectral mixture model for each pixel was also calculated. The error value gives the Euclidean norm of the error vector after least square error minimization. This error is caused by the components that were not included as endmembers in the linear unmixing procedure. In our model, we did not account for shadows and soil surface cracks. Therefore, we suggest that the error can be used as an indicator of the area of shadows and/or cracks on the soil surface. As a result, for each date, we obtained the area of soil crust and error for the test field. Then, we extracted the values of these parameters for pixels where we had collected spectral data and soil samples in the field. Next, we included the error and the area of soil crust as predictors into the models for SOM content prediction from Sentinel-2 data together with the informative parameter (parameters) determined during the analysis of spectral reflectance of the dry surface microlayer (see Section 2.7). Then, we removed the parameters with p-values above 0.05 from the models. We also checked for the spatial correlation of the residuals. The cross- validation, estimation of the new models’ accuracy and the check of residuals for spatial correlation were performed the same way as described in Section 2.6. Periodical tillage is a part of complete fallow cultivation agrotechnology. Thus, the soil crust microlayer forming on the surface of complete fallow field is periodically disrupted and mixed with the rest of the ploughed soil layer. As a result, there are several cycles of soil crust formation during the season. If the tillage happened close to the date of Sentinel-2 Remote Sens. 2021, 13, 2313 12 of 26 Remote Sens. 2021, 13, 2313 12 of 27

of Sentinel-2 imageimage acquisition, acquisition, we we would would ob observeserve low low values values of of crust crust area and high valuesval- of error due ues of error dueto to the the increase increase in in soil soil surface surface roughness. roughness. 3. Results and Discussion 3. Results and Discussion 3.1. Modelling of SOM Content Based on Field Spectral Data 3.1. Modelling of SOM Content Based on Field Spectral Data The SOM content of soil samples of the upper soil layer of the test field varied from The SOM 1.5%content to 7.3%of soil with samples an average of the upper value ofsoil 3.2% layer (Table of the1). test The field best varied model from for SOM content 1.5% to 7.3% withprediction an average developed value basedof 3.2% on (Table the field 1). spectralThe best reflectance model for of SOM the drycontent surface microlayer prediction developedincluded based only on one the parameter, field spectral the colorationreflectance index of the (CI), dry whichsurface is microlayer calculated using spectral included only onereflectance parameter, in the the green coloration and red index spectral (CI), bands:which is calculated using spectral reflectance in the green and red spectral bands: CI = (Band 4 − Band 3)/(Band 4 + Band 3) (2) CI = (Band4-Band3)/(Band4+Band3) (2) where Band3 andwhere Band4Band 3refer and toBand the4 field refer tospectral the field reflectance spectral reflectancerecalculated recalculated into corre- into correspond- sponding Sentinel-2ing Sentinel-2 bands. bands. When we usedWhen CI to model we used SOM CI content to model from SOM the contentfield spectral from reflectance the field spectral of the wet reflectance of the subsurface layer,wet we subsurface found that layer, two models we found had that a similar two models accuracy had and a similarprediction accuracy ability and prediction (Figure 7). Naturalability log (Figure was used7). in Natural the mode logl wasbecause used the in relationship the model because between the SOM relationship con- between tent and CI wasSOM better content described and CIby wasthis betterfunction. described by this function.

Figure 7. RelationshipFigure 7. Relationship between CI between and SOM CI content and SOM for thecontent dry surfacefor the anddry wetsurface subsurface and wet layer subsurface based onlayer the ﬁeld data based on the field data collected on 15 August 2019. collected on 15 August 2019. Originally, CI wasOriginally, related CIto wassoil color related [72]. to soilLater, color Galvao [72]. and Later, Vitorello Galvao [73] and showed Vitorello [73] showed that organic matterthat organichas the matterhighest has negative the highest correlation negative with correlation the soil spectral with the reflectance soil spectral reﬂectance in the region betweenin the region550 and between 700 nm. 550Green and and 700 red nm. spectral Green regions and redinvolved spectral in CI regions cal- involved in culation were CIalso calculation found to be were informative also found for toSOM be informativeprediction in for a SOMnumber prediction of studies in a number of [16,74–76]. studies [16,74–76]. Residuals for modelsResiduals developed for models for both developed dry and for wet both subsurface dry and wet layers subsurface did not show layers did not show spatial autocorrelationspatial autocorrelation (Figure 8). (Figure8).

Remote Sens. 2021, 13, 2313 Remote Sens. 2021, 13, 2313 13 of 27 13 of 26

Remote Sens. 2021, 13, 2313 13 of 27

(a) (b) (a) (b) Figure 8. Spatial autocorrelation within model residuals for: (a) dry surface layer; (b) wet subsurface FigureFigure 8. Spatial 8. Spatial autocorrelation autocorrelationlayer. within within modelmodel residuals residuals for: for: (a) dry (a) surfacedry surface layer; ( blayer;) wet subsurface(b) wet subsurface layer. layer. 3.2.3.2. AssessmentAssessment ofof SoilSoil SurfaceSurface StatusStatus atat thethe TimeTime of of Satellite Satellite Data Data Acquisition Acquisition 3.2. Assessment of Soil SurfTheTheace dry dryStatus soil soil surface atsurface the layerTime layer hasof has Satellite a highera higher spectralData spec tralAcquisition reﬂectance reflectancethan than that that of theof the wet wet subsur- sub- facesurface layer layer (Figure (Figure9). The 9). deviationThe deviation of spectral of spectral reﬂectance reflectance from thefrom average the average in both in cases both The dry soil surfaceincreasescases layerincreases when has when moving a higher moving from spec from thetral visible the reflectance visible to near-infrared to near-infrared than spectralthat spectral of rangethe range wet as this sub-as this range range is surface layer (Figuremoreis 9). more The sensitive sensitive deviation to theto the variations of variations spectral in in soil reflectancesoil moisture moisture content. content.from the Therefore, Therefore, average the the visiblein visible both range range is is cases increases whensuitable suitablemoving toto distinguishfromdistinguish the visible between between to the the near-infrared dry dry and and wet wet soil soil spectralsurface surface independent rangeindependent as this of of differencesrange differences in in is more sensitive to themoisturemoisture variations content.content. in soil moisture content. Therefore, the visible range is suitable to distinguish between the dry and wet soil surface independent of differences in moisture content.

FigureFigure 9.9.Spectral Spectral reflectancereflectance ofof drydry surfacesurface andand wetwet subsurfacesubsurface layerslayers measuredmeasured inin thethe field:field: dry,dry, wet—averagewet—average spectralspectral reflectancereflectance of dry surface surface and and wet wet subsurface subsurface layers layers correspondingly correspondingly for for 30 sample points; dry+sd, wet+sd—average spectral reflectance of dry surface and wet subsurface lay- 30 sample points; dry+sd, wet+sd—average spectral reflectance of dry surface and wet subsurface ers plus standard deviation of soil surface spectral reflectance of the corresponding layer; dry-sd, layers plus standard deviation of soil surface spectral reflectance of the corresponding layer; dry- wet-sd—average spectral reflectance of dry surface and wet subsurface layers minus standard de- sd,viation wet-sd—average of soil spectral spectral reflectance reflectance of the corresponding of dry surface andlayer. wet subsurface layers minus standard deviation of soil spectral reflectance of the corresponding layer. Figure 9. Spectral reflectanceWhen of dry reflectance surface valuesand wet for subsall Sentinel-2urface layers data and measured proximal in sensing the field: data dry, were posi- When reflectance values for all Sentinel-2 data and proximal sensing data were posi- wet—average spectral reflectancetioned in the of Band dry 3surface to Band and 4 spectral wet subsurface space, we foundlayers that correspondingly two distinct clouds for 30 of points tioned in the Band 3 to Band 4 spectral space, we found that two distinct clouds of points sample points; dry+sd, wet+sd—averagewere formed (Figure spectral 10). The re firstflectance cloud wasof dry concentrated surface and around wet subsurface the points correspond- lay- were formed (Figure 10). The first cloud was concentrated around the points corresponding ing to the wet subsurface layer; the second cloud was located around the points corre- ers plus standard deviationto the wetof soil subsurface surface layer; spectral the secondreflectance cloud of was the located corresponding around the layer; points dry-sd, corresponding sponding to dry surface layer. Therefore, we considered that the first cloud was formed wet-sd—average spectralto dry reflectance surface layer. of dry Therefore, surface and we consideredwet subsurface that the layers first cloudminus was standard formed de- from the from the points where the soil surface was wet at the time of Sentinel-2 data acquisition viation of soil spectral pointsreflectance where of the the soil corresponding surface was wetlayer. at the time of Sentinel-2 data acquisition and the and the second cloud was formed from the points where the soil surface was dry. second cloud was formed from the points where the soil surface was dry. The point cloud for the wet soil surface showed less scattering than the point cloud When reflectancefor values the dry for soil all surface. Sentinel-2 This can data be explaine and proximald by the masking sensing effect data that were the posi-soil moisture tioned in the Band 3 tohas Band on spatial 4 spectral variations space, of the we soil found spectral that reflectance. two distinct clouds of points were formed (Figure 10). The first cloud was concentrated around the points corresponding to the wet subsurface layer; the second cloud was located around the points corre-

sponding to dry surface layer. Therefore, we considered that the first cloud was formed from the points where the soil surface was wet at the time of Sentinel-2 data acquisition and the second cloud was formed from the points where the soil surface was dry. The point cloud for the wet soil surface showed less scattering than the point cloud for the dry soil surface. This can be explained by the masking effect that the soil moisture has on spatial variations of the soil spectral reflectance.

Remote Sens. 2021, 13, 2313 14 of 27

Scattering within the point cloud for the dry soil surface could have been caused by the variation in the state of the soil surface between different dates of Sentinel-2 data ac- Remote Sens. 2021, 13, 2313 quisition and, for each date, by the differences in the dry soil surface state14 of 26 between sample points.

FigureFigure 10. 10.Spectral Spectral reflectance reflectance for sample for sample points measured points withmeas aured HandHeld-2 with a field HandHeld-2 spectroradiome- field spectroradiom- tereter and and registered registered by Sentiniel-2 by Sentiniel-2 data in the dataBand in3 totheBand Band4 spectral 3 to Band space: 4 dry spectral soil surface space: layer—red dry soil surface layer— points;red points; wet subsurface wet subsurface layer—black layer—black points; 2 April points; 2019—green 2 April points; 2019—green 17 April 2019—yellow points; 17 points; April 2019—yellow 20points; April 2019—orange20 April 2019—orange points; 5 May points; 2019—purple 5 May points; 2019— 6 Junepurple 2019—brown points; 6 points;June 2019—brown 19 June points; 19 2019—darkJune 2019—dark green points; green 28 Augustpoints; 2019—dark 28 August blue 2019—dark points; the red blue oval points; delineates thepoints red oval with delineates a dry points with soila dry surface; soil surface; black oval—with black ov a wetal—with soil surface. a wet soil surface. The point cloud for the wet soil surface showed less scattering than the point cloud for theThe dry soilanalysis surface. of This meteorological can be explained data by theshowed masking that effect our that assignment the soil moisture of the soil surface to hasdry on and spatial wet variations classes based of the soil on spectralits reflectance reflectance. in Band 3 and Band 4 was correct (Figure 11). TheScattering dates 17 withinApril the2019, point 6 June cloud 2019, for the 19 dry June soil 2019, surface and could 28 haveAugust been 2019 caused fell in the middle by the variation in the state of the soil surface between different dates of Sentinel-2 data acquisitionof the dry and,period for eachin between date, by rainfall. the differences The date in thes 2 dry April soil 2019 surface and state 17 betweenApril 2019 were in the samplemiddle points. or immediately after the long rain period; thus, the soil surface did not have enoughThe analysis time to of dry meteorological out. As for data 5 showedMay 2019, that our it fell assignment at the ofend the of soil a surface short toperiod with low dryprecipitation. and wet classes Rainfall based onat itsthat reflectance time was in Band registered3 and Band in 4the was early correct morning (Figure 11 or). in the second Thehalf dates of the 17 Aprilday, 2019, meaning 6 June 2019,that 19the June soil 2019, surf andace 28 was August dry 2019 during fell in the middlemajority of daylight of the dry period in between rainfall. The dates 2 April 2019 and 17 April 2019 were in thehours. middle or immediately after the long rain period; thus, the soil surface did not have enough time to dry out. As for 5 May 2019, it fell at the end of a short period with low precipitation. Rainfall at that time was registered in the early morning or in the second half of the day, meaning that the soil surface was dry during the majority of daylight hours. The highest accumulated atmospheric precipitation was observed in the period from 19 June 2019 to 28 August 2019, meaning that the soil surface had undergone maximal rainfall-induced transformation by 28 August 2019 compared to the previous dates. The application of the regression model developed for the dry surface layer for the dates that fell in the point cloud for the wet surface resulted in an RMSE almost two times higher than when the model developed for the wet subsurface layer was used for the same dates (Table3). The application of the regression model developed for the wet subsurface layer for the dates that fell into the point cloud for the dry surface led to greater RMSE values than when we used the model developed for the dry surface layer for these dates.

Figure 11. Rainfall plot for the year 2019 for the test field: red arrows indicate Sentinal-2 image acquisition on 2 April 2019—day 92; 17 April 2019—day 107; 20 April 2019—day 110; 5 May 2019— day 125; 6 June 2019—day 157; 19 June 2019—day 170; 28 August 2019—day 240. Values above the

Remote Sens. 2021, 13, 2313 14 of 27

Scattering within the point cloud for the dry soil surface could have been caused by the variation in the state of the soil surface between different dates of Sentinel-2 data acquisition and, for each date, by the differences in the dry soil surface state between sample points.

Figure 10. Spectral reflectance for sample points measured with a HandHeld-2 field spectroradiometer and registered by Sentiniel-2 data in the Band 3 to Band 4 spectral space: dry soil surface layer— red points; wet subsurface layer—black points; 2 April 2019—green points; 17 April 2019—yellow points; 20 April 2019—orange points; 5 May 2019—purple points; 6 June 2019—brown points; 19 June 2019—dark green points; 28 August 2019—dark blue points; the red oval delineates points with a dry soil surface; black oval—with a wet soil surface.

The analysis of meteorological data showed that our assignment of the soil surface to dry and wet classes based on its reflectance in Band 3 and Band 4 was correct (Figure 11). The dates 17 April 2019, 6 June 2019, 19 June 2019, and 28 August 2019 fell in the middle of the dry period in between rainfall. The dates 2 April 2019 and 17 April 2019 were in the middle or immediately after the long rain period; thus, the soil surface did not have enough time to dry out. As for 5 May 2019, it fell at the end of a short period with low Remote Sens. 2021, 13, 2313 precipitation. Rainfall at that time was registered in the early morning or in the second15 of 26 half of the day, meaning that the soil surface was dry during the majority of daylight hours.

FigureFigure 11.11. Rainfall plot for for the the year year 2019 2019 for for the the test test fiel ﬁeld:d: red red arrows arrows indicate indicate Sentinal-2 Sentinal-2 image image acquisition on 2 April 2019—day 92; 17 April 2019—day 107; 20 April 2019—day 110; 5 May 2019— acquisition on 2 April 2019—day 92; 17 April 2019—day 107; 20 April 2019—day 110; 5 May 2019— day 125; 6 June 2019—day 157; 19 June 2019—day 170; 28 August 2019—day 240. Values above the day 125; 6 June 2019—day 157; 19 June 2019—day 170; 28 August 2019—day 240. Values above the colored blocks show atmospheric precipitation accumulated between the studied dates. Rainfall amount (mm) is presented on the OY axis and the number of a day in the year on the OX axis.

Table 3. Variations in the root-mean-square error of prediction (RMSE, %) of the SOM content modelling due to differences in the soil surface wetness.

RMSE When Using Regression Model RMSE When Using Regression Acquisition Date of Surface Class from Developed from Spectral Reﬂectance Model Developed from Spectral Sentinel-2 Images Figure 10 of Dry Soil Surface Microlayer Reﬂectance of Wet Subsurface Layer 2/04/2019 6.78 6.69 wet 17/04/2019 1.95 1.64 wet 20/04/2019 1.97 3.98 dry 5/05/2019 1.08 1.43 dry 6/06/2019 1.21 1.53 dry 19/06/2019 1.61 0.87 dry 28/08/2019 1.10 2.03 dry

The image acquired on 19 June 2019 was the exception. The soil surface on that date was classiﬁed as dry but the application of the regression model developed for the dry surface resulted in a higher RMSE than when we used the model developed for the wet subsurface layer for this date.

3.3. Influence of Rainfall-Induced Soil Surface Changes on Its Spectral Reflectance and on the Modelling of SOM Content The spectral reflectance of the soil surface microlayer on 5 May 2019 and 19 June 2019 in bands used for SOM prediction was very close to the spectral reflectance of the dry soil surface microlayer measured in the field (Figure 12). The models of SOM content for these dates had the highest accuracy and the lowest standard errors of the model parameters (Table4). The APD was the lowest: 1.15 for 5 May 2019 and 2.09 for 19 June 2019. Remote Sens. 2021, 13, 2313 16 of 26 Remote Sens. 2021, 13, 2313 16 of 27

Figure 12. AverageFigure deviation 12. Average of spectral deviation reﬂectance of spectral of thereflectance bare soil of surface the bare microlayer soil surface of samplemicrolayer points of sample at the dates of Sentinel-2 datapoints acquisition at the fromdates the of spectralSentinel-2 reﬂectance data acquisition of dry soilfrom surface the spectral microlayer reflectance on 15 Augustof dry soil 2019, surface and spectral mixtures with differentmicrolayer soil on crust 15 August fractions. 2019, Central and wavelengthspectral mixtur for eaches with analyzed different Sentinel-2 soil crust band fractions. is given inCentral brackets. wavelength for each analyzed Sentinel-2 band is given in brackets. Table 4. Variation in model parameters and accuracy when modelling SOM content of the soil surface layer from Sentinel-2 Table 4. Variation in model parameters and accuracy when modelling SOM content of the soil sur- data (only for the dates with a dry soil surface). face layer from Sentinel-2 data (only for the dates with a dry soil surface).

AcquisitionAcquisition Date of Date of Sentinel-2 Im- InterceptIntercept Ln (CI) Ln (CI) R2cv RMSEcvR2cv RMSEcvRPIQ RPIQ Sentinel-2 Images ages Value St. Error St. Error Value St. Value Error St. Error model based on spectralmodel reflectance based on ofspectral reflectance dry soil surface microlayerof dry soil measured surface in microlayer0.81 meas- 0.640.81 0.64 2.3 2.3 −–13.3513.35 1.65 1.65 –7.77 −7.770.77 0.77 the field on 15ured August in the 2019 field on 15 August 2019 20.04.2019 0.67 1.1 1.4 −10.43 3.23 −5.84 1.38 05.05.201920.04.2019 0.77 0.680.67 1.1 2.2 1.4 −–10.4312.25 3.23 1.73 –5.84 −7.621.38 0.85 06.06.201905.05.2019 0.75 0.850.77 0.68 1.7 2.2 −–12.2517.44 1.73 3.00 –7.62 −10.180.85 1.48 19.06.201906.06.2019 0.78 0.640.75 0.85 2.3 1.7 −–17.4411.45 3.00 1.49 –10.18 −7.541.48 0.77 − − 28.08.201919.06.2019 0.71 1.040.78 0.64 1.4 2.3 –11.4513.07 1.49 3.32 –7.54 7.800.77 1.59 28.08.2019 0.71 1.04 1.4 –13.07 3.32 –7.80 1.59 The average deviation of spectral reflectance of the soil surface microlayer from The spectralthe average reflectance spectral for 20 curve April derived 2019 and from 6 June field 2019 data was close the highest to the spectral on 28 August 2019 reflectance of(Figure the spectral 12). At mixes the same with time, 10% itof was the soil between crust the(Figure spectral 12). reflectanceThe model param- of the spectral mix eters for thesewith dates 20% differed and 10% the of most the soilfrom crust the parameters in Band 3 and of Bandthe model4 used based for SOM on the content field modeling. spectral reflectanceThe model of the parameters dry soil surface for this micr dateolayer, were similar and the to standard the parameters errors of of the the model model developed parameters fromwere fieldrather spectral high (Table data for 4). the Th drye regression surface microlayer, model for while 20 April the standard2019 had errors the for model lowest accuracyparameters among wereall models twice asdeveloped high and from the model Sentinel-2 accuracy data. was The 12% APD lower for 20 (Table April4). The APD 2019 was 2.43.was As the for highest 6 June for2019, this despite date (3.7). having an R2cv of 0.75, the RPIQ was rather low, demonstrating theThe poor spectral performance reflectance of the for mo 20del April for 2019this date. and 6The June APD 2019 for was 6 June close 2019 to the spectral was 2.7. reflectance of the spectral mixes with 10% of the soil crust (Figure 12). The model parame- Spatial tersautocorrelation for these dates of differedthe residuals the most was fromobserved the parameters for the models of the developed model based for on the field 20 April 2019spectral and 28 reflectance August 2019 of the (Figure dry soil 13a,e). surface Residuals microlayer, of the and models the standard for the errorsrest of of the model the dates didparameters not show spatial were rather autocorrelation. high (Table (Figure4). The 13b–d). regression model for 20 April 2019 had the lowest accuracy among all models developed from Sentinel-2 data. The APD for 20 April 2019 was 2.43. As for 6 June 2019, despite having an R2cv of 0.75, the RPIQ was rather low,

Remote Sens. 2021, 13, 2313 17 of 26

demonstrating the poor performance of the model for this date. The APD for 6 June 2019 was 2.7. Spatial autocorrelation of the residuals was observed for the models developed for 20 Remote Sens. 2021, 13, 2313 17 of 27 April 2019 and 28 August 2019 (Figure 13a,e). Residuals of the models for the rest of the dates did not show spatial autocorrelation. (Figure 13b–d).

(a) (b)

(e)

FigureFigure 13. Spatial 13. Spatial autocorrelation autocorrelation within model with residualsin model for: residuals (a) 20 April for: 2019; (a) 20 (b )April 5 May 2019; 2019; ( (bc) 65 JuneMay 2019; 2019; (d ()c 19) June 2019; (6e )June 28 August 2019; 2019.(d) 19 June 2019; (e) 28 August 2019.

3.4. Accounting for Soil Surface State When Modelling SOM Content Based on Sentinel-2 Data 3.4. Accounting for Soil Surface State When Modelling SOM Content Based on Sentinel-2 Data The highest mean and maximum area of the soil crust at sample points were registered The highest meanon 28 Augustand maximum 2019 followed area of by the 6 June soil 2019 crust (Table at sample5). The points APD was were also regis- the highest for tered on 28 Augustthese 2019 dates. followed The lowest by 6 June mean 2019 area of(Table the soil 5). crustThe APD was estimated was also on the 5 Mayhighest 2019 followed for these dates. Theby lowest 19 June mean 2019 (Figure area of 14 thea). ThesoilAPD crust for was these estimated dates was on also 5 May the lowest. 2019 fol- The variation lowed by 19 Junecoefﬁcient 2019 (Figure for the 14a). dates The with APD a low for crust these area dates was rather was highalso (400the tolowest. 100%), The demonstrating variation coefficienthigh for spatial the dates heterogeneity with a low of soil crust crust area development was rather (Figure high 14 (400c). As to for 100%), the error, it was demonstrating highthe spatial highest heterogeneity for the dates with of asoil low crust crust areadevelopment and APD value(Figure (Figure 14c). 14 Asb). for the error, it was the highest for the dates with a low crust area and APD value (Figure 14b).

Table 5. Temporal variations in the area of soil crust and error estimated from Sentinel-2 data for sample points and for the test field (1400 pixels): results for sample points/results for the test field.

Acquisition Date Crust (%) Error of Sentinel-2 Im- Standard Variation Standard Variation APD Maximum Mean Maximum Mean ages Deviation Coefficient Deviation Coefficient 20.04.2019 17/30 5.0/7.0 5/6 100/86 0.006/0.007 0.003/0.002 0.001/0.0008 33/39 2.43 5.05.2019 12/30 0.5/9 2/9 400/100 0.015/0.015 0.005/0.005 0.004/0.003 80/60 1.15 6.06.2019 22/60 10.0/13.0 6/8 60/62 0.009/0.032 0.003/0.003 0.002/0.004 67/133 2.70 19.06.2019 11/60 3.0/6.0 3/8 100/133 0.014/0.033 0.005/0.005 0.003/0.004 60/80 2.09

Remote Sens. 2021, 13, 2313 18 of 26 Remote Sens. 2021, 13, 2313 18 of 27

Table28.08.2019 5. Temporal 34/40 variations 20.0/22.0 in the area 11/9 of soil crust55/41 and error 0.007/0.023 estimated 0.003/0.003 from Sentinel-2 0.001/0.002 data for sample33/67 points 3.70 and for the test field (1400 pixels): results for sample points/results for the test field. At the field level, we observed a similar tendency. A lower mean area of the soil crust Acquisition Date was registeredCrust (%) for the dates with lower APD (Table 5, Figure Error 14a), while higher values of Sentinel-2 were determinedStandard for the datesVariation with the higher APD (Table 5, FigureStandard 14c). As forVariation the error, APD Maximum Mean Maximum Mean Images the oppositeDeviation pattern wasCoefficient determined. Maximum values of the parametersDeviation wereCoefficient higher 20.04.2019 17/30when 5.0/7.0 calculated 5/6 at the field 100/86 level because0.006/0.007 more pixels 0.003/0.002 were included0.001/0.0008 in the 33/39analysis. 2.43 5.05.2019 12/30Higher 0.5/9 maximum 2/9 values of 400/100 the soil crust0.015/0.015 area for the 0.005/0.005 dates with 0.004/0.003 lower APD mainly80/60 oc- 1.15 6.06.2019 22/60curred 10.0/13.0 in the area 6/8 along northern 60/62 field0.009/0.032 boundary (Figure 0.003/0.003 14a), which 0.002/0.004 is close to67/133 the road 2.70 19.06.2019 11/60 3.0/6.0 3/8 100/133 0.014/0.033 0.005/0.005 0.003/0.004 60/80 2.09 (Figure 1). 28.08.2019 34/40 20.0/22.0 11/9 55/41 0.007/0.023 0.003/0.003 0.001/0.002 33/67 3.70

(a) (b)

FigureFigure 14. 14.SpatialSpatial distribution distribution of error of ( errorb,d) and (b, dthe) and area the of soil area crust of soil(a,c) crust determined (a,c) determined as a result of as a result of spectral unmixing on the test field for 5 May 2019 (a,b) and 28 August 2019. spectral unmixing on the test field for 5 May 2019 (a,b) and 28 August 2019. When we included the information on the soil surface state in the models, their qual- At the field level, we observed a similar tendency. A lower mean area of the soil crust ity and accuracy increased (Table 6). For the scene from 20 April 2019, the R2cv increased was registered for the dates with lower APD (Table5, Figure 14a), while higher values were from 0.67 to 0.80, the RPIQ increased from 1.4 to 2.1, and the RMSEcv decreased from 1.1 to determined0.71 (Tables 4 for and the 6). datesFor scenes with from the higher5 May 2019 APD and (Table 6 June5, Figure2019, the 14 Rc).2cv As increased for the error, the fromopposite 0.77 and pattern 0.75 to was 0.79 determined.and 0.81, respectively; Maximum thevalues RPIQ increased of the parameters from 2.2 and were 1.7 higherto when 2.4calculated and 2.3, respectively; at the field and level the RMSEcv because decreased more pixels from were 0.68 and included 0.85 to in0.63 the and analysis. 0.61, Higher respectivelymaximum (Tables values 4 ofand the 6). soilFor crustthe scene area from for the28 August dates with2019, lowerthe R2cv APD and mainlyRPIQ in- occurred in creasedthe area from along 0.71 and northern 1.4 to 0.84 field and boundary 2.4, respectively, (Figure 14whilea), whichthe RMSEcv is close decreased to the road from (Figure1). 1.04 to 0.62When (Tables we included 4 and 6). the information on the soil surface state in the models, their quality and accuracy increased (Table6). For the scene from 20 April 2019, the R 2cv increased from 0.67 to 0.80, the RPIQ increased from 1.4 to 2.1, and the RMSEcv decreased from 1.1 to 0.71 (Tables4 and6). For scenes from 5 May 2019 and 6 June 2019, the R 2cv increased from 0.77 and 0.75 to 0.79 and 0.81, respectively; the RPIQ increased from 2.2 and 1.7 to 2.4 and 2.3, respectively; and the RMSEcv decreased from 0.68 and 0.85 to 0.63 and 0.61, respectively (Tables4 and6). For the scene from 28 August 2019, the R 2cv and RPIQ increased from Remote Sens. 2021, 13, 2313 19 of 26

0.71 and 1.4 to 0.84 and 2.4, respectively, while the RMSEcv decreased from 1.04 to 0.62 (Tables4 and6). Remote Sens. 2021, 13, 2313 19 of 27

Table 6. Variation in model parameters and accuracy when modelling SOM content with information on soil surface state (only for the dates with a dry soil surface). Table 6. Variation in model parameters and accuracy when modelling SOM content with information on soil surface state Acquisition Date of (only for the dates with a dry soil surface).Intercept Ln (CI) Error Crust R2cv RMSEcv RPIQ Sentinel-2 Images Value St. Error Value St. Error Value St. Error Value St. Error Acquisition Date of Sentinel-2 Intercept Ln (CI) Error Crust R−2cv RMSEcv RPIQ − 20.04.2019 0.80Images 0.71 2.1 3.7 2.2 Value2.0 St. Error 1.0 Value 891.3 St. Error 130.4 Value St. -Error Value - St. Error 05.05.2019 0.78 0.63 2.4 −5.8 3.2 −4.0 1.8 175.1 73.9 - - 06.06.2019 0.8120.04.2019 0.61 2.4 −0.803.0 0.71 3.0 2.1 −−3.13.7 2.2 1.4 −2.0 293.1 1.0 60.2891.3 −130.49.0 2.3- - 19.06.2019 0.8105.05.2019 0.64 2.3 −0.786.6 0.63 3.0 2.4 −−4.65.8 3.2 1.7 −4.0 157.9 1.8 82.3175.1 73.9 - - - - 28.08.2019 0.8406.06.2019 0.62 2.40.81 5.4 0.61 0.3 2.4 − -3.0 3.0 - −3.1 -1.4 - 293.1 −11.260.2 −9.0 1.1 2.3 19.06.2019 0.81 0.64 2.3 −6.6 3.0 −4.6 1.7 157.9 82.3 - - 28.08.2019The error related0.84 to the0.62 area of2.4 shadows 5.4 (including0.3 - soil surface- cracks)- was- informative−11.2 1.1 for most analyzed dates. The area of soil crust became a significant parameter only when its mean value reachedThe error 10% related (Table5 to). the The area APD of shadows for these (including dates was soil the surface highest. cracks) Moreover, was informa- in the case of 28tive August for most 2019, analyzed the area dates. of The soil area crust of was soil thecrust only became parameter a significant in the parameter model. only when its mean value reached 10% (Table 5). The APD for these dates was the highest. As for 19 June 2019, the quality of the model did not improve after adding new predictors; Moreover, in the case of 28 August 2019, the area of soil crust was the only parameter in it stayed at the level of the model based on the field spectral reflectance of the dry soil the model. As for 19 June 2019, the quality of the model did not improve after adding new surface microlayer.predictors; Incorporation it stayed at the of informationlevel of the model on the based soil on surface the field state spectral in thereflectance model of the resulted in andry increase soil surface in model microlayer. accuracy Incorporation up to the level of information of the model on the based soil surface on the fieldstate in the spectral reflectancemodel of resulted the dry in soilan increase surface in microlayer. model accuracy Model up accuracyto the level stayed of the model at the based same on the level independentfield ofspectral the date reflectance of Sentinel-2 of the imagedry soil acquisition. surface microlayer. Model accuracy stayed at the We did notsame observe level independent spatial autocorrelation of the date of Sentinel-2 of the residuals image acquisition. for most models developed with informationWe did on not the observe soil surface spatial state autocorrelation (Figure 15 b–e).of the Onlyresiduals the residualsfor most models of the devel- model for 20 Apriloped 2019with showedinformation an autocorrelationon the soil surface pattern state (Figure (Figure 15b–e). 15a). Only However, the residuals it was of the weaker when comparedmodel for 20 to April the model 2019 showed without an information autocorrelation on pattern the soil (Figure surface 15a). state However, for this it was date (Figure 13weakera). Residuals when compared of the model to the model with anwithout area information of soil crust on as the a singlesoil surface predictor state for this date (Figure 13a). Residuals of the model with an area of soil crust as a single predictor developed for 28 August 2019 did not show spatial autocorrelation in contrast to the model developed for 28 August 2019 did not show spatial autocorrelation in contrast to the including only CI as a predictor of SOM content for this date (Figure 13e). model including only CI as a predictor of SOM content for this date (Figure 13e).

(a) (b)

Figure 15. Cont.

Remote Sens. 2021Remote, 13, 2313 Sens. 2021, 13, 2313 20 of 26 20 of 27

(e)

Figure 15. FigureSpatial 15. autocorrelationSpatial autocorrelation within within model model residuals residuals for for mode modelsls incorporating incorporating data data on on soil soil surface surface state for: (a) 20 April 2019;state (b) 5 for: May (a) 2019; 20 April (c) 2019;6 June (b )2019; 5 May (d 2019;) 19 (Junec) 6 June 2019; 2019; (e) (28d) August 19 June 2019;2019. (e ) 28 August 2019.

4. Discussion 4. Discussion The soil crust formed under the impact of rainfall is a microlayer, which differs in microstructure, texture,The soil and crust mineralogy formed from under the underlyingthe impact soilof rainfall horizon is as a well microlayer, as which differs in in soil spectral reflectancemicrostructure, [36,38, 41texture,,77]. Several and mineralogy studies tried from to connect the underlying the changes soil in horizon as well as in soil soil surface propertiesspectral due reflectance to the formation [36,38,41,77]. of soil crustSeveral to the stud changesies tried in soilto connect spectral the changes in soil sur- reflectance. Forface example, properties [38] related due to the the difference formation in spectralof soil cr reflectanceust to the changes between in the soil spectral reflectance. soil crust and underlyingFor example, layer in[38] the related range of the 1.2 difference to 2.5 µm to in changes spectral in reflectance mineralogy andbetween the soil crust and soil texture duringunderlying crust formation. layer in They the basedrange theirof 1.2 conclusions to 2.5 μm on to thechanges analysis in ofmineralogy the and soil texture absorption peaks.during They alsocrust demonstrated formation. They that changes based their in soil conc spectrallusions reflectance on the dueanalysis of the absorption to soil crust formation can be used to model the infiltration rate. The authors of [77] also peaks. They also demonstrated that changes in soil spectral reflectance due to soil crust related observed changes in soil spectral reflectance during crust formation to changes in soil texture and mineralogy.formation can For be their used research, to model they choosethe infiltration soils of different rate. The texture authors and of [77] also related ob- carbonate contentserved and low changes SOM content in soil inspectral the range reflectance of 0.2 to 0.8%.during crust formation to changes in soil texture In contrast,and the authorsmineralogy. of [78] For in their their study research, found nothey difference choose insoils soil of texture different and texture and carbonate organic matter contentcontent between and low soil SOM crust content and control in the sample. range Thus,of 0.2 theyto 0.8%. attributed the observed changes in soilIn contrast, spectral reflectance the authors solely of to[78] the in impact their ofstudy soil surfacefound roughness.no difference in soil texture and However, they didorganic not specify matter the content depth of between the sampling soil of crust the soil and crust control for the sample. analysis Thus, of they attributed the SOM content andobserved soil texture, changes only mentioning in soil spectral that the reflectance depth of the solely layer to subjected the impact to the of soil surface roughness. artificial rainfall was 2 cm. Therefore, the absence of a difference in SOM content as well as However, they did not specify the depth of the sampling of the soil crust for the analysis in soil texture between the soil crust and control sample could be due to the fact that the analyzed samplesof includedSOM content both a and soil crustsoil texture, microlayer only and mentioni the underliningng that layer.the depth When of the layer subjected to the process is happeningthe artificial at a microlayer, rainfall was the 2 collection cm. Therefore, of soil samples the absence for the of estimation a difference in SOM content as of soil propertieswell of that as microlayerin soil texture should between be very the precise. soil crust Otherwise, and control the results sample can be could be due to the fact misleading. Previousthat the studies analyzed showed samples that both included SOM content both a andsoil itscrust composition microlayer affect and the underlining layer. soil’s spectral reflectanceWhen the [14 process,72,79,80 is]. Athappening the same time,at a microlayer atmospheric, the precipitation collection alters of soil samples for the esti- soil’s surface composition,mation of including soil properties soil organic of that matter microlayer [39,81]. Usually, should when be very developing precise. Otherwise, the results models for the predictioncan be misleading. of SOM content Previous based studies on optical showed remote that sensingboth SOM data, content the and its composition relationship is established between the SOM content of the mixed sample of the upper layer affect soil’s spectral reflectance [14,72,79,80]. At the same time, atmospheric precipitation (0 to 5 cm, 0 to 15 cm, or 0 to 20 cm) and spectral reflectance of the soil surface measured once [15,27,82]. Moreover,alters soil’s in surface some cases, composition, satellite data including used in modellingsoil organic were matter acquired [39,81]. Usually, when de- several years beforeveloping the collection models of for the the field prediction data. This of makes SOM the co developedntent based models on optical not remote sensing data, only site-specificthe but relationship also not reproducible. is established The between reproducibility the SOM of the content models of for the SOM mixed sample of the upper content estimationlayer is very (0 to important 5 cm, 0 to for 15 soil cm, monitoring or 0 to 20and cm) in an precisiond spectral agriculture. reflectance of the soil surface meas- In the currentured study, once the [15,27,82]. CI obtained Moreover, from proximal in some and satellite cases, remotesatellite sensing data used data in modelling were ac- was related to thequired content several and composition years before of SOMthe collection in the upper of the soil field layer data. of the This test field.makes the developed mod- Raindrop action (splash)els not only causes site-specific the breakdown but of also soil surfacenot reproducible. aggregates, containingThe reproducibility about of the models for 90% of soil surface organic carbon [83,84]. As different conditions are formed inside and SOM content estimation is very important for soil monitoring and in precision agriculture. outside soil aggregates, the organic matter of the surface of the aggregates is different from the organic matterIn ofthe their current inner study, part [85 the]. Therefore, CI obtained we from suggest proximal that the and registered satellite remote sensing data was related to the content and composition of SOM in the upper soil layer of the test field. Raindrop action (splash) causes the breakdown of soil surface aggregates, containing about 90% of soil surface organic carbon [83,84]. As different conditions are formed inside and outside soil aggregates, the organic matter of the surface of the aggregates is different from the organic matter of their inner part [85]. Therefore, we suggest that the registered deterioration of the accuracy and quality of models developed from Sentinel-2 data was

Remote Sens. 2021, 13, 2313 21 of 26

deterioration of the accuracy and quality of models developed from Sentinel-2 data was caused by the differences in SOM content and composition between soil surface at the time of satellite data acquisition and the upper soil layer. Moreover, in those cases, we observed a shift in the average satellite spectral curves from the average spectral curve derived from the field spectral reflectance of dry soil surface microlayer to spectral mixes with different fractions of soil crust, which generally demonstrated higher spectral reflectance. The degree of deviation increased with the APD. The increase in spectral reflectance of the soil surface microlayer as a result of soil crust formation and development was observed by [76] in their experiment with four soil types differing in soil texture. While field spectra were measured on 15 August 2019, almost at the end of the study, it did not represent the endpoint with maximum soil crusting as can be expected (Figure 12). This is also true for Sentinel-2 data. For example, the average spectral curves for 20 April 2019 and 6 June 2019 are close to the spectral mix with 10% of soil crust, while the spectral curves for 5 May 2019 and 19 June 2019 are close to the average spectral curve derived from the field spectral reflectance of the dry soil surface microlayer (Figure 12). This was probably caused by periodical tillage operations occurring several days before the Sentinel- 2 image acquisition on 5 May 2019 and 19 June 2019 as well as prior to field data collection. The accuracy of the models for these dates is the closest to the accuracy of the model derived from the field spectral reflectance of dry soil surface microlayer (Table4). According to the results of spectral unmixing, the average area of soil crust was the lowest for these dates, while the average error, related to the area of shadows and/or cracks, was the highest (Table5). The soil crust microlayer can be easily destroyed during tillage operations, and mixed with the rest of ploughed horizon. After tillage, when the soil surface experiences the impact of rainfall, the soil crust microlayer starts to form again. There can be several cycles of soil crust microlayer formation during one season. Depending on how long the soil surface is left bare, and the intensity, and the amount of rainfall, the degree of soil crust microlayer development in each cycle can vary. This makes it even more important to estimate the soil surface state at the time of spectral data acquisition to ensure the stability and reproducibility of the developed models. The results presented in Table3 are in accordance with the results obtained by [ 71]. They show that the application of the model developed for the dry surface to predict SOM content for the wet surface resulted in a significant drop in the model’s accuracy. This supports our decision to exclude certain images prior to the analysis to minimize the interference of soil moisture content. The results of spectral unmixing showed that variations of the soil surface state both in time and within the field can be rather high. We attribute the observed temporal variations to rainfall patterns. The area of soil crust on the studied field increased with the APD. We observed similar results in our model experiment (Figures3 and4), in which the area of the soil crust increased with cumulative rainfall. At the same time, the APD had the opposite effect on the error, which is related to the area of shadows (including surface cracks). As rainfall causes the breakdown of soil surface aggregates, the amount and size of large aggregates that can produce shadows decreases under its impact, while the soil crust becomes more pronounced [36,40]. Additionally, cracks in the soil surface can disappear after heavy rainfall, and the cracking process will begin again when the soil surface starts to dry. Vindeker et al. [40] demonstrated that the position of cracks changes in time, and the soil surface usually undergoes several cycles of transformation during the season. Within-field heterogeneity of the soil surface state of the test field is mainly connected to the SOM content. The SOM content affects the speed of soil surface transformation and soil crust development as it determines the stability of soil aggregates. At the beginning of soil crust development, the difference between soils with various SOM content will be low. While the soil crust develops, the difference between areas with higher and lower SOM content will increase due to the variation in the aggregate stability. Aggregates of soils with higher SOM content are more resistant to the impact of raindrops, and soil crust formation will be slower compared to soils with lower SOM content. Therefore, the longer the soil surface is exposed to the rainfall, the Remote Sens. 2021, 13, 2313 22 of 26

higher the spatial heterogeneity of soil crust development will be. As this heterogeneity is caused by the difference in SOM content, at a certain stage of soil crust development, the soil crust area will be directly related to SOM content. We observed such a situation for 28 August 2019, when the soil crust area became the only predictor in the model for SOM content estimation (Table6). The lowest area of the soil crust in the test field on this date was observed as expected in the areas with high SOM content (Figure 14c). Our results on the informativeness of the area occupied by the shadows at certain dates when modelling SOM content are in accordance with the results obtained by [27]. They showed that correction of the effect of shadows allowed for the improvement of models for soil organic carbon by 27% for field spectral data and by 25% for airborne spectral data. However, they used only the data collected during the period of 5 to 6 October and did not study how the impact of the shadows on the modelling of soil organic carbon changes in time when the soil surface experiences a more prolonged impact of rainfall. According to our results (Tables4 and6), the incorporation of the information on the area occupied by the shadows in the models leads to a decrease in the RMSEcv by 7 to 35% depending on the date of image acquisition. Here, we used a simple approach to account for the rainfall-induced changes in the soil surface when mapping SOM content from Sentinel-2 data. The approach needs further development and testing in different environments and on soils with different textures and mineralogy before we suggest its wide application. However, it is valuable for loamy soils formed on loess in similar climatic conditions (the temperate climatic zone), as soil texture in non-saline soils predetermines the mechanism of soil crust formation [33]. In our model experiment in 2017, we studied two different loamy soils (grey forest (from the test field) and leached chernozem (Luvic Chernozems according to [43]) and found that the transformations that occurred under the impact of rainfall were common [40,41]. The difference was only in the degree of soil aggregate disruption and the amount of washed material. This depends mainly on the organic matter content, because the higher the organic matter content is, the more stable the soil aggregates are. Additional studies are necessary to identify how the SOM content and composition of the soil surface microlayer changes under the influence of rainfall. It is also important to determine the relationships among rainfall-induced SOM, mineralogy changes, and soil surface spectral reflectance.

5. Conclusions The impact of rainfall leads to a discrepancy in the spectral reﬂectance and properties between the soil surface microlayer and the rest of the ploughed horizon. Our study showed that this rainfall-driven discrepancy negatively affects the accuracy and quality of models developed to map SOM content from Sentinel-2 data. The higher the rainfall impact was, the more the soil surface differed from the rest of the ploughed layer. The incorporation of information on the soil surface state in the model allowed the compensation of negative rainfall impact and the development of stable models with higher quality and accuracy. Accounting for the rainfall-induced changes in the soil surface is necessary to ensure the reproducibility of the developed models and the correct assessment of organic matter in the soil ploughed layer. This is a crucial issue in SOM detection and mapping based on remote sensing data.

Author Contributions: Conceptualization, methodology, writing—review and editing, supervision, project administration, resources, data curation, and funding acquisition, I.S.; writing—original draft preparation, methodology, visualization, validation, formal analysis, and investigation, E.P. All authors have read and agreed to the published version of the manuscript. Funding: This research was funded by the Ministry of Science and Higher Education of Russia (agreement No 075-15-2020-909). Acknowledgments: This paper has been supported by the RUDN University Strategic Academic Leadership Program. We would like to thank the Center of collective usage of equipment of the V.V. Remote Sens. 2021, 13, 2313 23 of 26

Dokuchaev Soil Science Institute for their help with laboratory analysis of soil samples and field data collection. Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Appendix A

Table A1. Spectral indices used in the study.

Index Equation Equation Adopted to Sentinel-2 bands Reference q q BI (Brightness Index) (B2 + G2 + R2) (Band 22 + Band 32 + Band 42) [58] 3 3 CI (Colouration Index) (R − G) (Band 4 − Band 3) [58] (R + G) (Band 4 + Band 3) HI (Hue Index) (2 ∗ R − G − B) (2 ∗ Band 4 − Band 3 − Band 2) [58] (G − B) (Band 3 − Band 2) SI (Saturation Index) (R − B) (Band 4 − Band 2) [58] (R + B) (Band 4 + Band 2) RI (Redness Index) R2 Band 42 [58] (B ∗ G3) (Band 2 ∗ Band 33) 1 − 2 ∗ |Band 4 − Band 3| Band 4 + Band 3 + Band 5 ; ( 3 ) 1 − 2 ∗ |Band 4 − Band 3| ( Band 4 + Band 3 + Band 6 ) ; BSCI (biological soil crust 3 1 − 2 ∗ |R − G| 1 − 2 ∗ |Band 4 − Band 3| [59] ( R + G + NIR ) Band 4 + Band 3 + Band 7 ; index) 3 ( 3 ) 1 − 2 ∗ |Band 4 − Band 3| Band 4 + Band 3 + Band 8 ; ( 3 ) 1 − 2 ∗ |Band 4 − Band 3| Band 4 + Band 3 + Band 8a ( 3 ) GSI (Grain Size Index) (R − B) (Band 4 − Band 2) [60] (R + B + G) (Band 4 + Band 3 + Band 2) CRUST − (R − B) − (Band 4 − Band 2) [57] 1 (R + B) 1 (Band 4 + Band 2) AVI (Advanced vegetation p p 3 (NIR + 1) ∗ (1 − R) ∗ (NIR − R) 3 (Band 8 + 1) ∗ (1 − Band 4) ∗ (Band 8 − Band 4) [63] index) (Band 8 − Band 4) (Band 8 + Band 4) ; ( − ) Band 8 Band 5 ; NDVI (Normalized (Band 8 + Band 5) difference (NIR − R) (Band 8 − Band 6) [64] (NIR + R) (Band 8 + Band 6) ; vegetation index) (Band 8 − Band 7) (Band 8 + Band 7) ; (Band 8 − Band 8a) (Band 8 + Band 8a) ; ∗ (Band 5 − Band 4) 1.5 (Band 5 + Band 4 + 0.5) ; ∗ (Band 6 − Band 4) 1.5 (Band 6 + Band 4 + 0.5) ; SAVI (Soil-adjusted ( − ) ( − ) 1.5 ∗ NIR R 1.5 ∗ Band 7 Band 4 ; [56] vegetation index) (NIR + R + 0.5) (Band 7 + Band 4 + 0.5) ∗ (Band 8 − Band 4) 1.5 (Band 8 + Band 4 + 0.5) ; ∗ (Band 8a − Band 4) 1.5 (Band 8a + Band 4 + 0.5) ; ∗ (Band 5 − Band 4) 2.5 (Band 5 + 2.4 ∗ Band 4 + 1) ; ∗ (Band 6 − Band 4) 2.5 (Band 6 + 2.4 ∗ Band 4 + 1) ; EVI2 (Enhanced ( − ) ( − ) 2.5 ∗ NIR R 2.5 ∗ Band 7 Band 4 ; [61] vegetation index) (NIR + 2.4 ∗ R + 1) (Band 7 + 2.4 ∗ Band 4 + 1) ∗ (Band 8 − Band 4) 2.5 (Band 8 + 2.4 ∗ Band 4 + 1) ; ∗ (Band 8a − Band 4) 2.5 (Band 8a + 2.4 ∗ Band 4 + 1) ; NDWI (Normalized (G − NIR) (Band 3 − Band 8) [62] difference water index) (G + NIR) (Band 3 + Band 8) B—spectral reflectance in the blue part of the spectrum; G—spectral reflectance in the green part of the spectrum; R—spectral reflectance in the red part of the spectrum; NIR—spectral reflectance in the near-infrared part of the spectrum. Remote Sens. 2021, 13, 2313 24 of 26

References 1. Krull, E.S.; Skjemstad, J.O.; Baldock, J.A. Functions of Soil Organic Matter and the Effect on Soil Properties; Cooperative Research Centre for Greenhouse Accounting Canberra: Canberra, Australia, 2004. 2. Halvorson, A.D.; Wienhold, B.J.; Black, A.L. Tillage, nitrogen, and cropping system effects on soil carbon sequestration. Soil Sci. Soc. Am. J. 2002, 66, 906–912. [CrossRef] 3. Orlov, D.S.; Biryukova, O.N.; Rozanova, M.S. Revised System of the Humus Status Parameters of Soils and Their Genetic Horizons. Eurasian Soil Sci. 2004, 37, 918–926. 4. Álvaro-Fuentes, J.; López, M.V.; Cantero-Martínez, C.; Arrúe, J.L. Tillage Effects on Soil Organic Carbon Fractions in Mediterranean Dryland Agroecosystems. Soil Sci. Soc. Am. J. 2008, 72, 541–547. [CrossRef] 5. Melero, S.; López-Garrido, R.; Murillo, J.M.; Moreno, F. Conservation tillage: Short- and long-term effects on soil carbon fractions and enzymatic activities under Mediterranean conditions. Soil Tillage Res. 2009, 104, 292–298. [CrossRef] 6. Kogut, B.M. Assessment of the humus content in arable soils of Russia. Eurasian Soil Sci. 2012, 45, 843–851. [CrossRef] 7. Batjes, N.H. A world dataset of derived soil properties by FAO? UNESCO soil unit for global modelling. Soil Use Manag. 1997, 13, 9–16. [CrossRef] 8. Heineke, H.J.; Eckelmann, W.; Thomasson, A.J.; Jones, R.J.A.; Montanarella, L.; Buckley, B. Land Information Systems: Developments for Planning the Sustainable Use of Land Resources; Office for Official Publications of the European Communities: Luxembourg, 1998. 9. Arrouays, D.; Deslais, W.; Badeau, V. The carbon content of topsoil and its geographical distribution in France. Soil Use Manag. 2006, 17, 7–11. [CrossRef] 10. Morvan, X.; Saby, N.; Arrouays, D.; Le Bas, C.; Jones, R.; Verheijen, F.; Bellamy, P.; Stephens, M.; Kibblewhite, M. Soil monitoring in Europe: A review of existing systems and requirements for harmonisation. Sci. Total Environ. 2008, 391, 1–12. [CrossRef] [PubMed] 11. Bünemann, E.K.; Bongiorno, G.; Bai, Z.; Creamer, R.E.; De Deyn, G.; de Goede, R.; Fleskens, L.; Geissen, V.; Kuyper, T.W.; Mäder, P.; et al. Soil quality—A critical review. Soil Biol. Biochem. 2018, 120, 105–125. [CrossRef] 12. Montanarella, L.; Badraoui, M.; Chude, V.; Costa, B.; Dos Santos, I.; Mamo, T.; Yemefack, M.; Singh Aulakh, M.; Yagi, K.; Young Hong, S. Status of the World’s Soil Resources Main Report. 2015. Available online: http://www.fao.org/3/i5199e/i5199e.pdf (accessed on 25 May 2021). 13. Minasny, B.; Malone, B.P.; McBratney, A.B.; Angers, D.A.; Arrouays, D.; Chambers, A.; Chaplot, V.; Chen, Z.-S.; Cheng, K.; Das, B.S.; et al. Soil carbon 4 per mille. Geoderma 2017, 292, 59–86. [CrossRef] 14. Vinogradov, B.V. Remote Sensing of the Humus Content of Soils; FAO: Rome, Italy, 1981. 15. Chen, F.; Kissel, D.E.; West, L.T.; Adkins, W. Field-Scale Mapping of Surface Soil Organic Carbon Using Remotely Sensed Imagery. Soil Sci. Soc. Am. J. 2000, 64, 746–753. [CrossRef] 16. Rossel, R.V.; Walvoort, D.; McBratney, A.; Janik, L.; Skjemstad, J. Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma 2006, 131, 59–75. [CrossRef] 17. Zhuo, L.; Liu, Y.; Chen, J.; Hu, C.; Wu, J. Quantitative retrieving of soil organic matter using field spectrometer and hyperspectral remote sensing. In Proceedings of the International Conference on Earth Observation Data Processing and Analysis, Wuhan, China, 28–30 December 2008. 18. He, T.; Wang, J.; Lin, Z.; Cheng, Y. Spectral features of soil organic matter. Geo-Spat. Inf. Sci. 2009, 12, 33–40. [CrossRef] 19. Vibhute, A.D.; Dhumal, R.K.; Nagne, A.; Surase, R.; Varpe, A.; Gaikwad, S.; Kale, K.V.; Mehrotra, S.C. Assessment of soil organic matter through hyperspectral remote sensing data (VNIR spectroscopy) using PLSR method. In Proceedings of the 2017 2nd International Conference on Man and Machine Interfacing (MAMI), Bhubaneswar, India, 21–23 December 2017; pp. 1–6. 20. Dou, X.; Wang, X.; Liu, H.; Zhang, X.; Meng, L.; Pan, Y.; Yu, Z.; Cui, Y. Prediction of soil organic matter using multi-temporal satellite images in the Songnen Plain, China. Geoderma 2019, 356, 113896. [CrossRef] 21. Medhioub, E.; Bouaziz, M.; Bouaziz, S. Spatial Estimation of Soil Organic Matter Content Using Remote Sensing Data in Southern Tunisia. In Advances in Remote Sensing and Geo Informatics Applications; El-Askary, H.M., Lee, S., Heggy, E., Pradhan, B., Eds.; Advances in Science, Technology & Innovation; Springer International Publishing: Cham, Switzerland, 2019; pp. 215–217. 22. Angelopoulou, T.; Balafoutis, A.; Zalidis, G.; Bochtis, D. From Laboratory to Proximal Sensing Spectroscopy for Soil Organic Carbon Estimation—A Review. Sustainability 2020, 12, 443. [CrossRef] 23. Nanni, M.R.; Demattê, J.A.M. Spectral Reflectance Methodology in Comparison to Traditional Soil Analysis. Soil Sci. Soc. Am. J. 2006, 70, 393–407. [CrossRef] 24. Liang, S. An investigation of remotely-sensed soil depth in the optical region. Int. J. Remote Sens. 1997, 18, 3395–3408. [CrossRef] 25. Cierniewski, J.; Verbrugghe, M.; Marlewski, A. Effects of farming works on soil surface bidirectional reflectance measurements and modelling. Int. J. Remote Sens. 2002, 23, 1075–1094. [CrossRef] 26. Minasny, B.; McBratney, A.B.; Bellon-Maurel, V.; Roger, J.-M.; Gobrecht, A.; Ferrand, L.; Joalland, S. Removing the effect of soil moisture from NIR diffuse reflectance spectra for the prediction of soil organic carbon. Geoderma 2011, 167–168, 118–124. [CrossRef] 27. Denis, A.; Stevens, A.; van Wesemael, B.; Udelhoven, T.; Tychon, B. Soil organic carbon assessment by field and airborne spectrometry in bare croplands: Accounting for soil surface roughness. Geoderma 2014, 226–227, 94–102. [CrossRef] 28. Vaudour, E.; Gomez, C.; Loiseau, T.; Baghdadi, N.; Loubet, B.; Arrouays, D.; Ali, L.; Lagacherie, P. The Impact of Acquisition Date on the Prediction Performance of Topsoil Organic Carbon from Sentinel-2 for Croplands. Remote Sens. 2019, 11, 2143. [CrossRef] Remote Sens. 2021, 13, 2313 25 of 26

29. Rienzi, E.A.; Mijatovic, B.; Mueller, T.G.; Matocha, C.J.; Sikora, F.J.; Castrignanò, A. Prediction of Soil Organic Carbon under Varying Moisture Levels using Reflectance Spectroscopy. Soil Sci. Soc. Am. J. 2014, 78, 958–967. [CrossRef] 30. Rabinovich, I.E. K voprosu o vliyanii meteorologicheskikh faktorov na pochvennuyu korku (On the effect of meteorological factors on the soil crust). Nauchnyi Byulleten’ Vsesoyuznogo Nauch. Issl. Instituta po Khlopku 1936, 3, 28–34. 31. Kaspirov, A.I. Prichiny obrazovaniya pochvennoi korki (Causes of soil crust formation). Pochvovedenie 1940, 7, 60–67. 32. Abramov, F.G. Opredelenie vodopronitsaemosti pochvy pri dozhdevanii (Determination of water permeability of the soil during sprinkling). Pochvovedenie 1954, 11, 71–79. 33. Valentin, C.; Bresson, L.-M. Morphology, genesis and classification of surface crusts in loamy and sandy soils. Geoderma 1992, 55, 225–245. [CrossRef] 34. Assouline, S. Rainfall-Induced Soil Surface Sealing: A Critical Review of Observations, Conceptual Models, and Solutions. Vadose Zone J. 2004, 3, 570–591. [CrossRef] 35. Kondrat’ev, K.Y.; Fedchenko, P.P. Vliyanie obrabotki na spektral’nye otrazhatel’nye svoistva pochvy (Effect of tillage on soil spectral reflectance). Pochvovedenie 1980, 12, 47–52. 36. Le Bissonnais, Y.; Bruand, A.; Jamagne, M. Laboratory experimental study of soil crusting: Relation between aggregate breakdown mechanisms and crust stucture. Catena 1989, 16, 377–392. [CrossRef] 37. Savin, I.Y. Vliyanie livnevogo dozhdya na integral’nuyu otrazhatel’nuyu sposobnost’ poverkhnosti chernozemnykh pochv (Influence of heavy rains on the integral reflectance properties of chernozemic soil surface). Pochvovedenie 1995, 8, 976–980. 38. Ben-Dor, E.; Goldlshleger, N.; Benyamini, Y.; Agassi, M.; Blumberg, D.G. The Spectral Reflectance Properties of Soil Structural Crusts in the 1.2- to 2.5-µm Spectral Region. Soil Sci. Soc. Am. J. 2003, 67, 289. [CrossRef] 39. Niu, J.; Yang, K.; Tang, Z.; Wang, Y. Relationships between Soil Crust Development and Soil Properties in the Desert Region of North China. Sustainability 2017, 9, 725. [CrossRef] 40. Vindeker, G.V.; Prudnikova, E.Y.; Savin, I.Y. Transformation of open soil surface under the impact of rainfall in model experiment. Dokuchaev Soil Bull. 2018, 95, 23–40. [CrossRef] 41. Prudnikova, E.Y.; Varlamov, E.B.; Churilin, N.A.; Churilina, A.E. Differentiation and balance of minerals during transformation of the open surface of arable soils under the impact of rain. Dokuchaev Soil Bull. 2019, 98, 105–131. [CrossRef] 42. Savin, I.Y.; Prudnikova, E.Y. About optimal dates of satellite images acquisition for arable soil mapping. Dokuchaev Soil Bull. 2014, 74, 66–77. [CrossRef] 43. IUSS Working Group WRB. World reference base for soil resources 2014: International soil classification system for naming soils and creating legends for soil maps. Update 2015. World Soil Resour. Rep. 2015, 106, 188. 44. Kiryushin, V.I. Ecological Foundations of Agriculture; Kolos: Moscow, Russia, 1996; 347p. 45. GOST 26213-91. Soils Methods for Determination of Organic Matter; Publisher of Standards: Moscow, Russia, 1991; 8p. 46. Walkley, A.; Black, I.A. An examination of the degtjareff method for determining soil organic matter, and a proposed modification of the chromic acid titration method. Soil Sci. 1934, 37, 29–38. [CrossRef] 47. Prudnikova, E.; Savin, I.; Vindeker, G.; Grubina, P.; Shishkonakova, E.; Sharychev, D. Influence of Soil Background on Spectral Reflectance of Winter Wheat Crop Canopy. Remote Sens. 2019, 11, 1932. [CrossRef] 48. Williams, A.J.; Pagliai, M.; Stoops, G. Physical and Biological Surface Crusts and Seals. In Interpretation of Micromorphological Features of Soils and Regoliths, 2nd ed.; Stoops, G., Marcelino, V., Mees, F., Eds.; Elsevier: Amsterdam, The Netherlands, 2018; pp. 419–440. 49. Mees, F.; Tursina, T.V. Salt Minerals in Saline Soils and Salt Crusts. In Interpretation of Micromorphological Features of Soils and Regoliths; Elsevier: Amsterdam, The Netherlands, 2010; pp. 441–469. 50. GOST 12536-79. Soils Methods of Laboratory Granulometric (Grain-Size) and Microaggregate Distribution; Publisher of Standards: Moscow, Russia, 1980; 18p. 51. Shein, E. The Particle-Size Distribution in Soils: Problems of the Methods of Study, Interpretation of the Results, and Classification. Eurasian Soil Sci. 2009, 42, 284–291. [CrossRef] 52. Available online: http://ilwis.itc.utwente.nl/wiki/index.php/Main_Page (accessed on 4 May 2021). 53. Stevens, A.; Ramirez-Lopez, L.; Stevens, M.A.; Rcpp, L. Package ‘Prospectr’. R Package Version 2020. Available online: https://cran.r-project.org/web/packages/prospectr/prospectr.pdf (accessed on 25 May 2021). 54. Lehnert, M.L.W. Package ‘Hsdar’. 2019. Available online: https://www.jstatsoft.org/article/view/v089i12 (accessed on 25 May 2021). 55. Fongaro, C.T.; Demattê, J.A.M.; Rizzo, R.; Safanelli, J.L.; Mendes, W.D.S.; Dotto, A.C.; Vicente, L.E.; Franceschini, M.H.D.; Ustin, S.L. Improvement of Clay and Sand Quantification Based on a Novel Approach with a Focus on Multispectral Satellite Images. Remote Sens. 2018, 10, 1555. [CrossRef] 56. Huete, A. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 1988, 25, 295–309. [CrossRef] 57. Karnieli, A.; Karnieli, A. Development and implementation of spectral crust index over dune sands. Int. J. Remote Sens. 1997, 18, 1207–1220. [CrossRef] 58. Mathieu, R.; Pouget, M.; Cervelle, B.; Escadafal, R. Relationships between Satellite-Based Radiometric Indices Simulated Using Laboratory Reflectance Data and Typic Soil Color of an Arid Environment. Remote Sens. Environ. 1998, 66, 17–28. [CrossRef] 59. Chen, J.; Zhang, M.Y.; Wang, L.; Shimazaki, H.; Tamura, M. A new index for mapping lichen-dominated biological soil crusts in desert areas. Remote Sens. Environ. 2005, 96, 165–175. [CrossRef] Remote Sens. 2021, 13, 2313 26 of 26

60. Xiao, J.; Shen, Y.; Tateishi, R.; Bayaer, W. Development of topsoil grain size index for monitoring desertification in arid land using remote sensing. Int. J. Remote Sens. 2006, 27, 2411–2422. [CrossRef] 61. Jiang, Z.; Huete, A.R.; Didan, K.; Miura, T. Development of a two-band enhanced vegetation index without a blue band. Remote Sens. Environ. 2008, 112, 3833–3845. [CrossRef] 62. McFeeters, S.K. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. Int. J. Remote Sens. 1996, 17, 1425–1432. [CrossRef] 63. Rikimaru, A.; Roy, P.S.; Miyatake, S. Tropical forest cover density mapping. Trop. Ecol. 2002, 43, 39–47. 64. Deering, D.W. Rangeland Reflectance Characteristics Measured by Aircraft and Spacecraft Sensors. Ph.D. Thesis, Texas A&M Universtiy, College Station, TX, USA, 1978. 65. Naimi, B. usdm: Uncertainty Analysis for Species Distribution Models. R Package Version 1.1-15. R Documentation. 2015. Available online: https://rdrr.io/cran/usdm/ (accessed on 11 June 2021). 66. The R Development Core Team. Package “Stats” R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing: Vienna, Austria, 2013. 67. Kuhn, M. Building predictive models in R using the caret package. J. Stat. Softw. 2008, 28, 1–26. [CrossRef] 68. Ploton, P.; Mortier, F.; Réjou-Méchain, M.; Barbier, N.; Picard, N.; Rossi, V.; Dormann, C.; Cornu, G.; Viennois, G.; Bayol, N.; et al. Spatial validation reveals poor predictive performance of large-scale ecological mapping models. Nat. Commun. 2020, 11, 1–11. [CrossRef] 69. Savin, I.Y.; Vindeker, G.V. About some peculiarities of the soil surface optical properties usage to detect soil moisture. Eurasian Soil Sci. 2021, 7, 806–814. 70. Stevens, A.; Udelhoven, T.; Denis, A.; Tychon, B.; Lioy, R.; Hoffmann, L.; van Wesemael, B. Measuring soil organic carbon in croplands at regional scale using airborne imaging spectroscopy. Geoderma 2010, 158, 32–45. [CrossRef] 71. Nocita, M.; Stevens, A.; Noon, C.; van Wesemael, B. Prediction of soil organic carbon for different levels of soil moisture using Vis-NIR spectroscopy. Geoderma 2013, 199, 37–42. [CrossRef] 72. Escadafal, R.; Huete, A. Improvement in remote sensing of low vegetation cover in arid regions by correcting vegetation indices for soil ’noise’ (Etude des proprietes spectrales des sols arides appliquee a l’amelioration des indices de vegetation obtenus par teledetection). Comptes Rendus Acad. Sci. 1991, 312, 1385–1391. 73. Galvao, L.S.; Vitorello, I. Role of organic matter in obliterating the effects of iron on spectral reflectance and colour of Brazilian tropical soils. Int. J. Remote Sens. 1998, 19, 1969–1979. [CrossRef] 74. Agbu, P.A.; Fehrenbacher, D.J.; Jansen, I.J. Soil Property Relationships with SPOT Satellite Digital Data in East Central Illinois. Soil Sci. Soc. Am. J. 1990, 54, 807–812. [CrossRef] 75. Daniel, K.W.; Tripathi, N.K.; Honda, K.; Apisit, E. Analysis of VNIR (400–1100 nm) spectral signatures for estimation of soil organic matter in tropical soils of Thailand. Int. J. Remote Sens. 2004, 25, 643–652. [CrossRef] 76. Brown, D.J.; Shepherd, K.D.; Walsh, M.G.; Mays, M.D.; Reinsch, T.G. Global soil characterization with VNIR diffuse reflectance spectroscopy. Geoderma 2006, 132, 273–290. [CrossRef] 77. Goldshleger, N.; Ben-Dor, E.; Benyamini, Y.; Agassi, M. Soil reflectance as a tool for assessing physical crust arrangement of four typical soils in Israel. Soil Sci. 2004, 169, 677–687. [CrossRef] 78. Anderson, K.; Kuhn, N.J. Variations in soil structure and reflectance during a controlled crusting experiment. Int. J. Remote Sens. 2008, 29, 3457–3475. [CrossRef] 79. Shields, J.A.; Paul, E.A.; Arnaud, R.J.S.; Head, W.K. Spectrophotometry measurement of soil color and its relationship to moisture and organic matter. Can. J. Soil Sci. 1968, 48, 271–280. [CrossRef] 80. Karmanov, I.I.; Rozhkov, V.A. Experimental determination of quantitative relationships between color characteristics of soils and soil constituents. Sov. Soil Sci. 1972, 4, 666–674. 81. Kuhn, N.J. Erodibility of soil and organic matter: Independence of organic matter resistance to interrill erosion. Earth Surf. Process. Landf. 2007, 32, 794–802. [CrossRef] 82. Mirzaee, S.; Ghorbani-Dashtaki, S.; Mohammadi, J.; Asadi, H.; Asadzadeh, F. Spatial variability of soil organic matter using remote sensing data. Catena 2016, 145, 118–127. [CrossRef] 83. Jastrow, J.D.; Miller, R.M.; Boutton, T. Carbon Dynamics of Aggregate-Associated Organic Matter Estimated by Carbon-13 Natural Abundance. Soil Sci. Soc. Am. J. 1996, 60, 801–807. [CrossRef] 84. Six, J.; Elliott, E.T.; Paustian, K. Soil Structure and Soil Organic Matter II. A Normalized Stability Index and the Effect of Mineralogy. Soil Sci. Soc. Am. J. 2000, 64, 1042–1049. [CrossRef] 85. Ellerbrock, R.H.; Gerke, H.H. Characterizing organic matter of soil aggregate coatings and biopores by Fourier transform infrared spectroscopy. Eur. J. Soil Sci. 2004, 55, 219–228. [CrossRef]