Synergy of remotely sensed data in spatiotemporal dynamic modeling of the crop and cover management factor

Pooja P.PREETHAand Ashraf Z.AL-HAMDAN

1Department of Civil Engineering,Alabama A&M University,4900 Meridian Street N,Huntsville,AL 35811-7500(USA)2Department of Civil and Environmental Engineering,Universityof Alabama in Huntsville,301 Sparkman Drive,Huntsville,AL 35899-7500(USA)

ABSTRACT Soil erosion is a threat to the water quality constituents of sediments and nutrients and can cause long-term environmental damages.One important parameter to quantify the risk of soil loss from erosion is the crop and cover management factor(C-factor),which represents how cropping and management practices affect the rates and potential risk of soil erosion.We developed remotely sensed data-driven models for dynamic predictions of C-factor by implementing dynamic land cover modeling using the SWAT(Soil and Water Assessment Tool)model on a watershed scale.The remotely sensed processed variables included the enhanced vegetation index(EVI),the fraction of photosynthetically active radiation absorbed by green vegetation(FPAR),leaf area index(LAI),soil available water content(AWC),slope gradient(SG),and ratio of area(AR)of every hydrologic response unit(HRU)to that of the total watershed,comprising unique land cover,soil type,and slope gradient characteristics within the Fish River catchment in Alabama,USA between 2001 and 2014.Linear regressions,spatial trend analysis,correlation matrices,forward stepwise multivariable regression(FSMR),and 2-fold cross-validation were conducted to evaluate whether there were possible associations between the C-factor and EVI with the successive addition of remotely sensed environmental factors.Based on the data analysis and modeling,we found a significant association between the C-factor and EVI with the synergy of the environmental factors FPAR,LAI,AWC,AR,and SG(predicted R2 (R2pred)=0.51;R2 =0.68,n=3 220,P <0.15).The results showed that the developed FSMR model constituting the non-conventional factors AWC(R2pred =0.32;R2 =0.48,n=3 220,P <0.05)and FPAR(R2pred =0.13;R2 =0.28,n=3 220,P =0.31)was an improved fit for the watershed C-factor.In conclusion,the union of dynamic variables related to vegetation(EVI,FPAR,and LAI),soil(AWC),and topography(AR and SG)can be utilized for spatiotemporal C-factor estimation and to monitor watershed erosion.

KeyWords: C-factor,enhanced vegetation index,land cover modeling,remote sensing,soil erosion,soil moisture,solar radiation

INTRODUCTION

Soil loss from water erosion can cause long-term environmental changes and affect water quality in terms of sediments and nutrients (Blanco-Canqui and Lal, 2008;Ezemonye and Emeribe,2012).The essential measures of practical erosion investigations and effective water conservation include accurate predictions of soil loss risks and their variability with time and space (Singhet al., 2008;Dymond and Vale,2018).The effectiveness of soil erosion risk assessments largely depends on the factors affecting soil erosion under various geographical conditions(Panagoset al.,2015;Preethaet al.,2019).To date,researchers have obtained large amounts of experimental and field data and developed predictive physical models and regression models for soil erosional risk assessment and hot spot identification.Conservation measures may have considerable effects on soil loss reduction, sedimentation, and the probability of floods in these areas(Luet al.,2004;Mancinoet al.,2014).The universal soil loss equation (USLE) model, revised universal soil loss equation(RUSLE)model,and their latest versions are very commonly used to predict soil loss,and are related to six erosional factors (Sadeghiet al., 2004).One important parameter in the USLE equation is the crop and cover management factor (C-factor), which characterizes the effects of different land covers,as well as that of vegetation.The C-factor estimates also show how cropping and management practices affect the rates and potential risk of soil erosion(Karaburun,2010).The C-factor component in the USLE model is defined as the ratio of soil loss from land cropped under specified conditions to the corresponding loss from clean-tilled,continuous fallow land(Karaburun,2010;Panagoset al.,2015).A study by Alewellet al.(2019)highlighted that the C-factor predicted by the USLE model was non-continuous and imprecise.The aforementioned study recommended that a selective collaboration of models,remotely sensed data,and continuous monitoring programs are a requisite for resolving the uncertainties prevailing in conventional C-factor predictions and soil erosion modeling.

Various studies have evaluated the C-factor on different terrains and watershed areas based on field data, physical and semi-physical models,statistical and survey analyses,regressions,remotely sensed satellite data,and a series of effective correlation variables(Basicet al.,2004;Karaburun,2010;Loh,2012;Panagoset al.,2015;Tanyaşet al.,2015;Ucaret al.,2018).In addition,many studies have represented C-factor using multiple approaches with the normalized difference vegetation index(NDVI),including linear correlations,natural logarithms,and exponential functions(Van der Knijffet al.,2000).Despite this,research has shown that the NDVI,with all its new enhancements,is vulnerable to atmospheric and canopy background conditions(Luoet al.,2005;Loh,2012).The enhanced vegetation index(EVI)is an alternative approach that incorporates the vital concepts of background adjustment and atmospheric resistance into the NDVI.Thus,EVI is an excellent tool for use in mixed landscapes and regions of dense vegetation and forests(Matsushitaet al.,2007;Hatfield and Prueger,2010;Loh,2012).Recently,a noticeable increase in studies devoted to remote sensing for data and modeling has been observed(El Baroudy, 2011;Preetha and Al-Hamdan,2020b).Some studies have combined models and remote sensing technology to develop new C-factor prediction methods(Denget al.,2008).They highlighted the noticeable impact of the areal resolution in altering the spatial dynamics of the C-factor while remarking the need to revisit the conventional C-factor estimation(Denget al.,2008;Guoet al.,2015).The SWAT(Soil and Water Assessment Tool)model was employed in this study for watershed modeling.Annual estimates of the C-factor were simulated using the soil,digital elevation model(DEM)and remotely sensed land cover data from the moderate resolution imaging spectroradiometer(MODIS)on the Terra and Aqua satellites for the study area.The model categorized the watershed into subunits with unique soil types, slope gradients,and land cover properties,described as hydrologic response units (HRUs) (Dechmiet al., 2012).The USLE functionality was embedded in SWAT-simulated HRU-wise C-factor with the aid of land cover-wise minimum C-factor and quantities of rainfall residue on the surface soil layer(Neitschet al.,2011).However,this C-factor approach uses static variables that cannot showcase spatial and temporal changes in the C-factor estimation over a study area.This gap can be overcome by using a collection of temporally and spatially changing variables and implementing them into the traditional equation.Nevertheless,there is no provision of dynamic variables in the conventional equation.In this case, remotely sensed satellite image data can be utilized to provide easily accessible,promptly updated,and timely information on land cover(vegetation indices,surface cover reflectance,surface water models),soil profiles(surface and root zone moisture levels,temperature,soil texture,organic carbon content),slope gradient,and elevation.Further,they can be used to develop a model that incorporates dynamic variables affecting the C-factor and estimate the variation in the C-factor both spatially and temporally in watersheds(Serraet al.,2008).

The aim of this study was to develop,validate,and verify a dynamic and straightforward C-factor model,constituting readily available remotely sensed land cover,soil,and topography factors.This study incorporated annual land cover progressions within the watershed and explored the effects of prominent,yet unintroduced,variables of soil available water content(AWC)and the fraction of photosynthetically active radiation absorbed by green vegetation(FPAR)on C-factor predictions.The dynamic C-factor model solely incorporated the remotely sensed factors from MODIS,including land cover,EVI,FPAR,and leaf area index(LAI)(annually monitored)as well as AWC(monthly monitored).

MATERIALS AND METHODS

This study was designed to develop a dynamic C-factor model for mixed landscapes based on remotely sensed geospatial data.The methodology of the study included three major phases: i) geoprocessing of remotely sensed land cover, soil, and topographical data, and development of consolidated C-factor for the study area;ii)development and validation of a novel C-factor model based on the remotely sensed factors of EVI,LAI,FPAR,AWC,ratio of HRU area to the total watershed area(AR),and slope gradient(SG);and iii)the verification of the C-factor model and development of C-factor maps of the study area to evaluate its accuracy at the spatial and temporal levels.

In this study,we incorporated the annual remotely sensed satellite imagery of EVI,LAI,FPAR,and AWC for modeling and validating C-factor in the Fish River watershed in Alabama,United States with the aid of Python programming language,ArcGIS software(works on the concept and processes of geographic information system(GIS)),SWAT,and Google Earth Engine Code Editor.The statistical approaches used in the study comprised linear regressions,spatial trend analysis,correlation matrices,forward stepwise multivariable regressions(FSMR),and 2-fold cross validations.We included annual land cover variability in the model using the maximum likelihood supervised classification of the yearly surface reflectance data from MODIS.

Case studyarea and land cover modeling

The case study area is located in the Fish River watershed in Baldwin County,Alabama,southeastern United States.The agricultural watershed has an area of 386km2.The governing land cover in the watershed is agriculture and cultivation(44%),followed by grass and shrub(34%),andforests(11%)(Fig.1).The Fish River watershed was initially modeled in SWAT using the DEM, land cover, soil type,and meteorological conditions of the year 2001, and then simulated for the years 2002–2014.The soil data, DEM,and climate data were obtained from the United States Department of Agriculture(USDA),United States Geological Survey(USGS),and United States National Weather Service,respectively(Table I).The hydrological data(stream flow and sediment yields)collected at seven water quality gauging sites in 2001–2014 were from the USGS and Alabama Department of Environmental Management(ADEM).

Fig.1 Maps showing land cover classes(a)and the corresponding values of consolidated crop and cover management factor(Cc)(b)averaged over the years 2001–2014 for the 230 hydrologic response units of the Fish River watershed in Baldwin County,Alabama,southeastern USA.

For the dynamic modeling of land cover,the annual land cover data were geoprocessed for every HRU of the Fish River watershed using the maximum likelihood supervised classification of the surface reflectance data collected from MODIS.The surface reflectance data from MODIS were obtained from the USGS under the Land Processes Distributed Active Archive Center(LP DAAC)database for the years 2001–2014(USGS,2019b).Each pixel is of 500-m resolution and contains the best possible pixel value from the year.The software and models employed in the analysis included the Python programming language,Google Earth Engine Code Editor,ArcGIS(Extract by Mask and Spatial Join tools),and SWAT(HRU Overlay and HRU Definition tools).The preprocessing involved the development of a Python script in the platform known as Google Earth Engine Code Editor for the supervised classification of satellite remotely sensed data (Google Earth Engine, 2018).The main processes involved in supervised classification were the development of training land cover points to convert unsupervised land cover data to supervised data and the import of the annual MODIS surface reflectance data.This was followed by the identification of the watershed extents,selection of training land cover points,and overlay of training points onto the MODIS data,which generated the MODIS land cover maps with supervised classification for the Fish River watershed.Maps were developed for the years 2001–2014 in Google Earth Engine Code Editor and spatiallyjoined with the HRU map of the watershed using ArcGIS.These spatially joined maps containing a land cover class corresponding to every HRU of the watershed were fed as the land cover input into the SWAT models for the respective years to obtain annual dynamic models of the Fish River watershed.In SWAT,the hydrologic model was defined,and watershed delineation was performed using DEM, which generated seven sub-basins.Afterward,the soil data,slope features,and dynamically classified land cover maps were overlaid and defined,and 230 HRUs were generated.Thus,the application of annual dynamic MODIS data in SWAT accounted for the annual changes in land cover within the HRUs of the Fish River watershed.Next,the meteorological data(precipitation,temperature,solar radiation,relative humidity,and wind speed)were fed into the model.The model was developed and used for simulation of the hydrological components of stream flows and sediment yields over 2001–2014.The calibration and validation results of the developed models are shown in Table SI(see Supplementary Material for Table SI).They suggested that the model predictions successfully represented the real hydrogeological conditions in the Fish River watershed(Nash-Sutcliffe efficiency(NS)=0.61–0.84;R2=0.63–0.93).

TABLE IData sources for the crop and cover management factor(C-factor)modeling study

Remotelysensed data and geoprocessing

The remotely sensed environmental data,including the EVI,AWC,LAI,and FPAR,were obtained for per pixel of the derived images for the southeastern United States.The data sources,spatial and temporal data resolution,and periods of analysis are described in Table I.They were extracted for the Fish River watershed from the LP DAAC database over the entire 2001–2014 period.The EVI data(product ID: MOD13Q1) had a resolution of 250 m, and the LAI and FPAR data product(MOD15A2H)had a resolution of 500 m(USGS,2019a,c).The LAI is the ratio of one-half of the total surface area of the leaves to the unit ground area;FPAR represents the fraction of photosynthetically active radiation (FPAR) absorbed by green vegetation, which is identified as in the data product subset in MODIS;and AWC represents the soil moisture percentage in the surface 0–10 cm soil layer.The remotely sensed datasets of EVI,AWC,LAI,and FPAR used in the improved C-factor estimation in this study were developed using spatial superposition of the remotely sensed satellite data for the spatial level of HRU in the Fish River watershed,as shown in Fig.2.The steps involved were importing the raw data from remote sensing data into ArcGIS,locating the watershed extents,masking and statistical zoning,and spatial joining the zoned MODIS data to the HRU map.This process was carried out annually for the years 2001–2014 for the four factors EVI, AWC,LAI,and FPAR.Each spatially joined map contained 230 attributes corresponding to the 230 HRUs in the Fish River watershed.The attributes were further used in modeling as the independent and confounding variables for each year between 2001 and 2014(n=230×14=3 220).The spatial distribution of EVI,FPAR,LAI,and AWC in the study area was averaged annually from 2001 to 2014,as shown in Fig.S1(see Supplementary Material for Fig.S1).In this study,AR represents the ratio of an HRU area to the total watershed area, and SG represents the slope gradient of an HRU in the watershed.Watershed delineation performed using DEM gave AR and SG,which were identified for the corresponding HRUs and used as the topographical variables for modeling.

Fig.2 Workflow of geoprocessing estimates of the enhanced vegetation index(EVI),the fraction of photosynthetically active radiation absorbed by green vegetation(FPAR),leaf area index(LAI),and soil available water content(AWC)from remotely sensed moderate resolution imaging spectroradiometer(MODIS)imageries.DEM=digital elevation model;HRU=hydrologic response unit.

Consolidated C-factor development

The C-factor calculated based on the USLE equation in the code inbuilt in the SWAT model was denoted asCU(Neitschet al.,2011):

whereCU,mis the minimum value ofCUfor the land cover,and rsdsis the amount of residue on the soil surface(kg ha−1).Thus,theCUin the watershed has a reasonable probability of spatial variation with changes in the land cover.However,it fails to represent a pixel unit smaller than a land cover.The main limitation ofCUis that it is non-dynamic with time since it conventionally depends on the temporal resolution of land cover, which is fed as input to the SWAT model.Hence,dynamic land cover plays a critical role in producing spatially as well as temporally significant C-factor.The gaps inCUled to the development of a better,progressive C-factor called consolidated C-factor(Cc),which is produced from annual dynamic land covers.

This study developedCc,which was used as the dependent variable for modeling in place ofCU.For this,C-factor values were collected from the surveyed data published in the literature and from empirical model estimates(Mattheus and Norton,2013;Mancinoet al.,2014;Guoet al.,2015;Panagos,2015;Phamet al.,2018;USDA,2018).Then,they were manually sorted and assigned based on the dominant land cover in every watershed HRU.The variableCcis also unique and detailed, owing to its improved depiction of the C-factor for the dominant land cover of agriculture and cultivation,grass and shrub,and forests in the study area.It is a comprehensive measure that takes into account two major classes of forests(evergreen and deciduous),four kinds of grasses,summer and winter pastures,and wetland and water.They also included agricultural lands with generic crops,row crops,and close-grown crops,as well as 26different types of crops with distinct vegetation and growth stage properties,as shown in Table SII(see Supplementary Material for Table SII).TheCcvalues for every land cover class in the study were calculated annually over the period of 2001–2014 for 230 HRUs(n=3 220).Such comprehensive and annually changing C-factor data elevated the reliability of the new C-factor function and had the strongest control on sedimentation and soil loss modeling and management(Lee,2004).The variableCcis unitless and ranges from 0 to 1.Thereafter,the study divided land cover into eight classes based onCcvalue:urban(0–0.001),evergreen forest(0.001–0.001 3),deciduous forest(0.001 3–0.001 5),rangeland(0.001 5–0.003),grass and shrub(0.003–0.06),pasture(0.06–0.1),agriculture and cultivation(0.1–0.5), and wetland and water(>0.5).The maps showing land covers and the correspondingCcaveraged over the years 2001–2014 for the 230 HRUs of the Fish River watershed are shown in Fig.1.

Dynamic C-factor model development and validation

Initially, the data from different sources of information, including DEM, land cover data, EVI, AWC, LAI,and FPAR,with different spatial and temporal resolutions,were homogenized to ensure consistency of data,integrity of analysis,and validity of results(Böhmet al.,2001).All of the provided data layers were visually analyzed for their geographic accuracy.Then,the most accurate datasets were used to georectify the other data layers.The modeling of the dynamic C-factor in the study consisted of linear regressions,spatial trend analysis,correlation matrices,and FSMRs.The general trends betweenCcand remotely sensed variables of EVI, LAI, FPAR, and AWC were quantified using linear regression analysis.Next,a spatial trend analysis was employed to establish the association of the eight categoricalCcclasses with the HRU-wise distribution of the topographical variables of AR and SG.Later,the dependent variable,Cc,and the independent variables,including six remotely sensed factors, were used to develop a correlation matrix.This study used correlation matrices developed using the non-parametric correlation indicator of Pearson’s product-moment correlation (r) (Damghani, 2013).Next,dynamic C-factor models were developed using FSMR with the union of remotely sensed satellite data, topographical variables,andCc.The 2-fold cross-validation of the FSMR model was performed to consider all observations for both model development and validation.The cross-validation estimates eliminated the chances of overfitting the model because of the diversity of variables included in it.The cross-validation was carried out by randomly shuffling the 2001–2014 dataset into two sets(K1and K2)of equal size.Then,the model was developed on K1and validated on K2,followed by development on K2and validation on K1.The model performance was assessed using the following three evaluation parameters: coefficient of determination (R2),predictedR2(),and NS(Gungoret al.,2016).The NS of the model was calculated as follows:

whereCsis the simulated C-factor using the FSMR model,andis the arithmetic average ofCsfor the 230 HRUs in the study watershed.

FSMR model verification

To evaluate the significance of the FSMR model in representing the spatial land cover distribution in the Fish River watershed,theCswas verified using the long-term averages from 2001 to 2014 for each HRU in the watershed.Then,theCswas classified into eight different sets according to the minimum and maximum estimates of the C-factor for all land cover classes in the study area.The same process wasrepeated to obtainCcandCUfor all land cover classes in the watershed.To evaluate the significance of the FSMR model in representing the annual C-factor in the watershed,theCswas computed from 2001 to 2014 and annually averaged for all HRUs in the watershed.The same process was repeated to obtain annual estimates ofCcandCUbetween 2001 and 2014 in the watershed.For both the spatial and temporal verification scenarios,the estimates ofCsandCcwere compared with those ofCU.Finally,spatiotemporal maps ofCswere generated for the 230 HRUs of the watershed averaged over 2001–2014 and were compared to the estimates ofCUandCcprediction.

RESULTS

Spatiotemporal data analysis

The descriptive statistics of land cover,soil,and topographical factors in the watershed are detailed in Table II.For the study area,the meanCcand mean EVI were assessed to be 0.12 and 0.33,respectively.These values indicate the dominance of vegetation within the watershed because Cfactor decreases and EVI increases with green area(Panagoset al.,2015).The soil analysis results showed relatively high soil moisture contents,with a maximum of 33%.The LAI showed a range from 0.3 (indicating the presence of bare soil)to 10(representing dense forests)with an average of 1.55(Neitschet al.,2011).The peaking effect of sunlight was evident from the deviation(0.59)observed in the mean FPAR from the minimum FPAR, indicating the probable effects of FPAR onCc.The watershed delineation statistics showed that the HRU distribution had a reasonable pixel resolution that captured up to 0.001 km2area.Thus, the HRUs enhanced the dynamic estimation of C-factor in the area because of their fine spatial resolution,together with the annual variation in mixed landscapes.The SG varied from 0.74 to 13.78 m m−1.Table III shows the positive association ofCcwith AWC and AR.This indicates that the C-factor tended to increase with higher levels of soil moisture and larger land cover areas of the corresponding HRUs.Negative associations ofCcwere observed with EVI,FPAR,LAI,and SG.This indicates that the C-factor tended to reduce with higher EVI and LAI,increasing proportions of FPAR,and elevated gradients of slope.The results of linear regression analysis for all the studied factors produced lowR2values despite significantPvalues except for the factor of FPAR.

Little is known about the possible association of topography with EVI and C-factor.The significance of including the topographical factors in the FSMR model was tested using the spatial trend analysis of AR and SG for the eight categorical C-factor classes (Fig.3).In this analysis, AR represented the ratio of the HRU area to the total watershed area,which fell under each category of the C-factor class;SGrepresented the average value of the slope gradients of the HRUs under each C-factor class.The AR values varied from 0.01 (pasture) to 0.6(agriculture and cultivation), which tended to significantly elevate the C-factor from 0.06to 0.1(USDA,2001).These variations could be attributed to human interferences, such as overgrazing, crop rotations,and logging(Celik,2005;Emadiet al.,2008).Overall,the dominant land cover areas exhibited a relatively higher Cfactor range(0.06–0.50)when compared with the lesser land cover areas in the watershed.The analysis showed that the C-factor classes tended to deviate the SG from 2.50 to 5.00 m m−1.The C-factor displayed a substantial elevation from 0.001 in urban spaces to 0.06in pastures,with noticeable increases in SG of the respective land covers.This trend was strengthened by studies that found significant slope gradient effects on the space-wise features of the topsoil layer and land cover with specific cover management(Asmamaw and Mohammed, 2013).Nonetheless, few studies have shown that multiple land cover management practices are relevant to the varying range of channel slope gradients(Mugaggaet al.,2012;Li and Meng,2013).Overall,the study resultsshowed positive correlations of both AR and SG with the C-factor,with high levels of significance(P <0.001)andR2of 0.527 and 0.265, respectively(Fig.3).Hence, they were included as confounding variables in the modeling process of dynamic C-factor.

Fig.3 Relationships between AR and SG and Cc of the eight land cover classes for the Fish River watershed in Baldwin County,Alabama,southeastern USA.The AR and SG values are averages of each of the eight land cover classes over the years 2001–2014 (n = 3 220).AR = ratio of the area of hydrologic response unit to that of the total watershed;SG=slope gradient;Cc =consolidated crop and cover management factor.Land cover classes(Cc value):I=urban(0–0.001);II=evergreen forest(0.001–0.001 3);III=deciduous forest(0.001 3–0.001 5);IV=rangeland(0.001 5–0.003);V=grass and shrub(0.003–0.06);VI=pasture(0.06–0.10); VII = agriculture and cultivation (0.1–0.5); VIII = wetland and water(>0.5).

TABLE IIIResults of linear regression between land cover, soil, and topographical variablesa) and consolidated crop and cover management factor(Cc)for the Fish River watershed in Baldwin County,Alabama,southeastern USA(n=3 220)

TABLE IIDescriptive statistics of land cover,soil,and topographical variablesa) for the Fish River watershed in Baldwin County,Alabama,southeastern USA averaged for the time period between 2001 and 2014

Correlations between cover management and its influencing factors

TABLE IVCorrelation coefficients between consolidated crop and cover management factor(Cc)and remotely sensed land cover,soil,and topographical variables for the Fish River watershed in Baldwin County,Alabama,southeastern USA(n=3 220)

Table IV shows the correlations betweenCcand independent variables.ThePvalues for all variables were less than 0.05,which indicates significance correlations between the factors and the C-factor as well as between the factors.TheCcand EVI exhibited a negative correlation(r=−0.46).The EVI,being a global vegetation index,provides global vegetation data that are more reliable and spatiotemporally updated (Jianget al., 2008; Wuet al., 2011; Loh, 2012).Hence, we hypothesized that the EVI has a strong influence on the C-factor for the study area, which was dominated by agriculture and permanent crops.However,other environmental components,including soil characteristics,atmospheric and meteorological changes,and topographical factors,can contribute to the prediction uncertainties in EVI and NDVI,and eventually alter cover management estimates(Matsushitaet al.,2007;Loh,2012).Hence,a comprehensive analysis of the effect of EVI on all types of land covers,including urban land,forests,pastures,shrubs,agriculture and cultivation,and water,was more descriptive and suitable to represent real C-factor than the conventional C-factor assessment methods(Dye and Tucker,2003).In addition,the combined evaluation of EVI in the presence of other environmental factors was advocated.The factor FPAR,which is a measure of sunlight on vegetation,quantifies the proportion of photosynthetically active radiation that a canopy absorbs.The significant negative correlation between FPAR andCcsupported this(r=−0.33)(Wuet al.,2011).

In contrast, the vegetative property of LAI exhibited a weak positive correlation withCc(r=0.09).However,the strong positive associations between FPAR and LAI(r=0.88)and LAI and AR(r=0.46)were noteworthy,as confirmed by Haggaret al.(2011).The absence of LAI weakened the correlations between the C-factor and EVI,FPAR, and AR (r <0.04).This finding strengthened the strong dependence of LAI on FPAR and the land cover areas of each HRU in the study area, suggesting the inclusion of FPAR followed by LAI in further analyses.Despite the low positive correlation betweenCcand AWC(r=0.25),the associations of AWC with EVI, FPAR, and LAI were steady.It is known that AWC has long-term effects on soil sedimentation and is beneficial for analyzing the quantity and quality of various soil nutrients and runoff.Thus,AWC can affect the long-term C-factor constitution in watershed units(Li and Meng,2013;Preetha and Al-Hamdan,2019,2020a).Hence, this study explored the probable effect of soil-related factor of AWC on the C-factor and EVI,which was conventionally thought of as a function of ground measurements and landscape attributes.A substantial positive correlation was observed betweenCcand AR(r=0.52).However,a poor negative connection was observed betweenCcand SG(r=−0.07)despite the robust interaction of SG with EVI and AR(Matsushitaet al.,2007).

FSMR model

In this study,six FSMR model equations were developed and validated to assess the association betweenCcand EVI in the presence of AWC, LAI, FPAR, AR, and SG in the Fish River watershed for the years 2001–2014 (Table V).The exponential and linear regression functions employedfrom Equation 1 to Equation 6were accompanied by 95%profile likelihood confidence intervals andPvalues from the chi-squared tests.Equation 1 included only the EVI factor and measured the effect of an exponential variation in EVI onCc.The constantsaandbin Equation 1 determined the form of the exponential curve ofCcand EVI.They were generated through a trial-and-error methodology.The best fit was obtained whena= 1.5 andb= 1 (R2pred= 0.07,R2= 0.21).TheR2predandR2values were considerably low,despite the favorable correlations betweenCcand EVI yielding acceptablePvalues (P <0.05).The reduction in direct sunlight was considered to improve the yields in the related landscapes (Staveret al., 2001).Thus, FPAR(r=0.14,P <0.05),which showed a high correlation and effective trend with EVI(Table IV),was added to Equation 1 to generate Equation 2.The LAI factor,being a function of the plant cover geometry,provides a better description of the C-factor dynamics in various plant covers and vegetation practices(Andersonet al.,2004).The high correlation of LAI with FPAR also made it an appropriate choice to be included in Equation 3.To identify whether there was a dominant contribution between the vegetation component of EVI and the soil component of AWC in this study,we added AWC to EVI into Equation 4.The results indicated that the combination of EVI,FPAR,and LAI with the uninvestigated element of AWC significantly affectedCc(R2pred= 0.32,R2=0.48,P <0.05).

TABLE VResults of the six forward stepwise multivariable regression(FSMR)model equations that determine the effects of enhanced vegetation index(EVI),the fraction of photosynthetically active radiation absorbed by green vegetation(FPAR),leaf area index(LAI),soil available water content(AWC),ratio of the area of hydrologic response unit to that of the total watershed(AR),and slope gradient(SG)on consolidated crop and cover management factor(Cc)in the Fish River watershed in Baldwin County,Alabama,southeastern USA(n=3 220)

In Equation 5,the topographical factor of AR was included with all of the elements in Equation 4 to reconfirm the importance of the topographical aspect of land cover areas in affecting crop management factors.The inclusion of the factor AR markedly affected the performance of the model(R2=0.63).To consider the potential effect of slopes and terrains on the likely improvement of C-factor estimates,SG was included as the confounding variable in Equation 6 with all of the variables in Equation 5(R2pred=0.51,R2=0.68).The results showed that thePvalues of Equations 1–6using the FSMR were less than 0.05 for all categories except for FPAR(Equation 2)and LAI(Equations 5 and 6).The contradicting values ofR2andPcould be attributed to the multicollinearity among the six independent variables of the models.Therefore, the validity of Equation 6was reconfirmed usingR2pred.TheR2predexplains the overfitting nature of a regression model with the successive addition of variables and also eliminates random noise from the dataset(Li,2017).In this study,R2predincreased from 0.07(Equation 1)to 0.51(Equation 6),and Equation 6displayed the best fit for the dynamic C-factor estimation.Similarly,theR2values increased from 0.21(Equation 1)to 0.68(Equation 6).

Further,a 2-fold cross-validation analysis was performed to verify the efficacy of Equation 6in predicting C-factor in the study area (Fig.4).Comparison betweenCsandCcgeneratedR2pred,R2, and NS values of 0.54, 0.72, and 0.68 for the K1fold and 0.48, 0.64, and 0.52 for the K2fold, respectively, which indicates a good fit for Equation 6.This suggests that there is a relevant and compelling association betweenCcand EVI in the active presence of FPAR, LAI, AWC, AR, and SG.A higher precision and reliability of the C-factor estimation could be achieved by the inclusion of more factors,such as human interventional aspects and atmospheric effects, as well as by capturing seasonal variations.

Fig.4 Comparison of Cs and Cc for the years 2001–2014 within the 230 hydrologic response units of the Fish River watershed in Baldwin County,Alabama,southeastern USA.Cs =simulated crop and cover management factor; Cc =consolidated crop and cover management factor.Stepwise multivariable regression model was used for the simulation(n =3 220).The 2001–2014 dataset was randomly shuffled into two sets(K1 and K2)of equal size for cross-validation.The lines are trend lines.R2 =coefficient of determination;R2pred =predicted coefficient of determination;NS=Nash-Sutcliffe efficiency.

Spatiotemporal assessment of FSMR models

The spatial and temporal estimates ofCsandCcaveraged over 2001–2014 for every HRU in the study area were compared to those ofCU(Fig.5).A reasonable closeness was observed betweenCcandCsfor all land covers except urban,rangeland,and wetland and water.Hence,we suggest either modeling theCsfor urban,rangeland,or wetland and water explicitly or reclassifying theCcusing higher spatially resolved data.TheCsshowed strong resemblance with theCcwith regard to the most dominant land covers of agriculture and cultivation, grass and shrubs, and evergreen forest,constituting 90%of the Fish River watershed.The relative changes inCUandCcwere 0.36and 0.61 for deciduous forest and pasture,respectively.The variability dropped to 0.29 and 0.45, respectively, when the C-factor was tuned for EVI,LAI,and the effects of topography,along with the newly identified factors of FPAR and AWC.This implies that the appropriately diverse combination of environmental factors positively affected the C-factor estimation.However,the range ofCsvaried considerably fromCcfor the other land covers,such as water,urban,rangeland,and pastures,based on annual temporal conditions.The incorporation of seasonality could eliminate this variation in C-factor estimation,especially for dry seasons(Loh,2012).For all models, the estimates ofPvalues were less than 0.05.The outcome confirmed an active collaboration between the variables conventionally accounted for inCU,Cc,andCs,while considering areas with different land covers.TheCUfor the seven years(2002, 2004, 2006, 2008, 2010, 2012,and 2014)differed from theCcby an average of 70%.This was becauseCUcalculations incorporated in SWAT used only the radiation and land cover management factors while neglecting some of the vital contributing factors.Conversely,no major deviations were found betweenCcandCs.TheCsdiffered from theCcby an average of 0%–8%, except for in the year 2010,which showed a 12%deviation.These results indicate the necessity of including FPAR and AWC in C-factor modeling for the Fish River watershed.

Fig.5 Spatial and temporal verification of the crop and cover management factor (C-factor) using CU,Cc, and Cs for the eight land cover classes averaged over the years 2001–2014(a)and for the seven years during 2001–2014 averaged across the 230 hydrologic response units (b) of the Fish River watershed in Baldwin County,Alabama,southeastern USA.CU =C-factor calculated from the universal soil loss equation;Cc =consolidated C-factor;Cs =C-factor simulated using the forward stepwise multivariable regression model.Land cover classes:I=urban;II=evergreen forest;III=deciduous forest;IV=rangeland;V=grass and shrub;VI=pasture;VII=agriculture and cultivation;VIII=wetland and water.

Mapping of C-factor using the USLEand FSMR models

The C-factor maps were generated for the 230 HRUs of the Fish River watershed usingCUandCsaveraged from 2001 to 2014 (Fig.6).Unlike the coarse spatiotemporal variation ofCU(once in 15 years, 500–1500 m), a good spatiotemporal resolution ofCs(annual, 36–500 m) was observed in our study.The outcome showed the creditable capability of the remotely sensed data-driven FSMR model to describe progressive soil losses with changing land cover in the watershed.In addition, in the existing studies, the procedures of annual dynamic land cover modeling and the combination of the explanatory factors of FPAR,AWC,and LAI with vegetation indices for improving cover management predictions were not introduced.In our study,these aspects were elucidated in a much more detailed way,and the processes were implemented annually in a spatial resolution range of 36–500 m.In general,the annual estimates of the FSMR C-factor of this study can be utilized as the yearly C-factor estimates in sediment and erosion modeling,with a strong emphasis on model verification.Further, studies should implement the C-factor developed using the FSMR model to evaluate its influence on sediment yield estimation and water quality in watersheds.Additionally,spatiotemporal validation of the C-factor model for other watersheds in different past,present,and future temporal scenarios would strengthen the reliability and global usability of the developed FSMR model for real-time water quality modeling.It would also help to elucidate the interconnectivity of C-factor and sedimentation at catchment scales.Our findings indicate that remotely sensed data-based dynamic models are highly suited for estimating the spatial and temporal variabilities of C-factor in agricultural watersheds such as the Fish River watershed.The study model also showcased advancements from the existing C-factor estimation approaches by addressing the annual dynamics of land use components(Panagoset al.,2015).In addition,remotely sensed satellite data and data products,which are freely available,possess adequate spatiotemporal resolution,and can be retrieved on regional and global scales,greatly compensate for the dearth of Cfactor data in different parts of the world (Mattheus and Norton,2013;Mancinoet al.,2014;Guoet al.,2015).

Fig.6 Spatiotemporal mapping of the crop and cover management factor(C-factor)using CU(a)and Cs(b)for the Fish River watershed in Baldwin County,Alabama,southeastern USA.CU =C-factor calculated from the universal soil loss equation; Cs =C-factor simulated using the forward stepwise multivariable regression model.

Studylimitations

Although the study presented some advancements in the existing C-factor predictions in river basins,the limitations of the developed model deserve close attention and further studies.Future studies should validate the developed Cfactor from the FSMR model to evaluate the reliability of the spatiotemporal sediment yield estimation in watersheds.Further,studies may consider developing a specific module to simulate sediment yield and soil erosion rates in space and time using the FSMR model,which can be integrated into a watershed scale hydrological model,such as SWAT.The application of the study model to other watersheds in varying temporal scenarios as well as in differing land cover,soil,and terrain characteristics would strengthen the reliability and global usability of the developed FSMR model for realistic water quality modeling.Future work should include more sensitive contributing factors for accurate Cfactor estimations and implement advanced remote sensing techniques to detect spatial and temporal changes at the optimal resolutions.These measures particularly enhance the genuine assessment of potential multi-temporal and multispatial changes for modeling and monitoring soil and water quality.Since the assessment of water quantity(stream flow,river runoff, groundwater flow) is interrelated with water quality concerns(erosion,sediments,nutrients),especially in regions with intensive agriculture,a connective system of water quantity with the developed water quality model would be of great relevance for integrated water management.

CONCLUSIONS

Land cover,a vital component in hydrogeological and landscape modeling, has been assessed using various vegetation indices.Dynamic interactions of EVI with easilyaccessible information from satellite images to determine C-factor have yet to be conducted in the Fish River watershed in Alabama.As an effective indicator of global vegetation,EVI can have strong contributions to variability in C-factor for the study area, being an agricultural watershed with significant spatiotemporal environmental changes.However,many other environmental factors may produce errors in EVI,altering cover management estimates.Linear regressions,spatial trend analysis,correlation matrices,FSMR,and 2-fold cross-validation were conducted to evaluate whether there were possible associations between C-factor and EVI with successive addition of remotely sensed environmental factors.The results suggested that the traditionally unaccounted factors AWC(R2pred=0.32;R2=0.48,P <0.05)and FPAR (R2pred= 0.13;R2= 0.28) advanced the connection between C-factor and EVI.Performance tests of the FSMR model predictions were conducted in the Fish River watershed for spatial(HRU level)and temporal(2001–2014)scenarios(R2pred=0.51;R2=0.68,P <0.15).The average C-factor values from the consolidated data differed from those predicted by the FSMR model by 0%–8%,whereas they differed from those predicted by the USLE model by 0%–70%.This showed that the dynamic model signified the environmental and hydrogeological carriers of erosional soil losses compared with the previous models.The spatial verification of FSMR C-factor closely approached the consolidated C-factor by 7%, 25%, and 15% for the land cover classes agriculture and cultivation,grass and shrub,and evergreen forest,respectively.Hence,the outcomes suggested the applicability of the developed model in C-factor prediction of multi-spatial patterns of land.The C-factor maps showed that the association of dynamic variables with vegetation(EVI,FPAR,and LAI),soil(AWC),and topography(AR and SG)produced a more realistic and spatially and temporally dynamic estimation of C-factor.Compared with other methods for estimating C-factor,the FSMR model possesses one main advantage:it uses an optimal and easily accessible synergy of data derived from satellite images to develop C-factor in watershed for finer spatial and temporal resolutions.Our findings indicate that remotely sensed data greatly facilitate the evaluation of watershed erosion and sediment yield distribution in any spatial and temporal scenarios.The remotely sensed data-based dynamic models are highly suitable for estimating the spatial and temporal variability of C-factor in agricultural watersheds such as the Fish River watershed.This study showcased advancements from the existing C-factor estimation approach by addressing the land use components’annual dynamics and can be applied to comparable river basins.In addition, remotely sensed satellite data and data products are freely available,possess adequate spatiotemporal resolution,and can be retrieved at regional and global scales.Therefore,they will compensate for the absence of C-factor data and literature in different parts of the world.

ACKNOWLEDGEMENT

We are extremely grateful to the reviewers and editors for their valuable comments and suggestions,which improved this article.

SUPPLEMENTARY MATERIAL

Supplementary material for this article can be found in the online version.