Estimation of Soil Surface Roughness Parameters Under Simulated Rainfall Using Spectral Reflectance in Optical Domain

The purpose of the study was to evaluate the possibility of parameterizing the state of soil surface roughness (SSR) based on proximal measurements of spectral reflectance in the VIS‐NIR range, which is important for the needs of monitoring the state of soil surfaces. SSR should be constantly monitored as it provides an insight into a range of hydrological and erosive soil processes and improves the interpretation of remote sensing data. SSR and the spectral reflectance of three texturally different soils were measured under simulated rainfall in laboratory condition. The relationship between the SSR parameters and soil spectra was determined using regression random forest models. Various spectral data processing methods were tested and the best wavelengths for SSR description after rainfall were found. Two roughness indices were used to describe the SSR: Height Standard Deviation (HSD) and T3D (Tortuosity index). Although both shared a significant correlation with SSR, the T3D index demonstrated a more pronounced rainfall effect and a closer correlation with spectral data than HSD. The best determination of T3D was obtained with the raw spectra (RAW) (R2 = 0.71), as well as with spectra transformed with the baseline alignment first derivative (BA1d) method (R2 = 0.71) or the Savitzky‐Golay (SG) method (R2 = 0.69). Different wavelengths were the best SSR predictors depending on the spectral transformation method (VIPs ‐ Variable Importance in Projection). For both roughness indices, the NIR wavelengths (725–820 nm) yielded the highest VIP Score in models based on RAW spectra, while those in the VIS region (450–772 nm) were most important in models based on transformed spectra.

Each of these processes also depends on the volume and intensity of rainfall and surface slope (Alvarez-Mozos et al., 2011).
Many studies have been conducted to measure the effect of rainfall on roughness, soil properties and soil processes. Dalla Rosa et al. (2012) investigated the influence of different soil tillage systems, and the effect of loamy and sandy/clayey textured soils, on the evolution of SSR with time in Brazil under simulated rainfall. Similar quantitative research on SSR was performed using simulated rainfall on the Red Yellow Latosol in Brazil by Bramorski et al. (2012), on sandy loam soil in Austria by Hauer et al. (2001) and on black soil in Northeast China by Xingming et al. (2017). In both studies, SSR was measured initially and then under natural rainfall. Changes in the SSR, and the rate of these changes during rainfall, are known to depend on the intensity and volume of rainfall, and on the size, shape and distribution of soil aggregates (Zobeck & Onstad, 1987). Mwendera and Feyen (1994) studied changes in SSR, bulk density and total porosity caused by rainfall on silt loam soil in Belgium.
It is also possible to predict the magnitude of these changes. Römkens et al. (2002), Takken et al. (2001), Le Bissonnais et al. (2005), Strudley et al. (2008) researched the effect of rainfall on soil erosion at different scales. Other studies have examined the impact of SSR on the partitioning of the rainwater into surface runoff, infiltration , depressional storage (Zhao et al., 2018) and soil and water loss (Bramorski et al., 2012;Hauer et al., 2001). Research to date indicates that SSR can act like mulch or vegetation cover, causing a reduction of surface runoff and increasing the infiltration of water (Prosdocimi et al., 2016;Zhao, Gao, et al., 2016) and depressional storage (Zhao et al., 2018), thus reducing soil loss by erosion (Darboux & Huang, 2005). In addition, Abban et al. (2017) examined changes in SSR regarding initial microroughness (i.e., less than 2 mm) on soil with loamy texture in the USA, under simulated rainfall. During the experiment, rain splash appeared to be responsible for increasing SSR on smooth surfaces. However, the authors emphasize that the precise conditions and mechanisms behind the observed increase in microroughness are unknown. Previous studies have also evaluated tool for describing soil microtopography (Digital Elevation Model DEM) (Stańczyk & Baryła, 2016), as well as various roughness indices such as random roughness, variogram analysis, fractional dimensions models (Vermang et al., 2013) and the limiting difference (Abban et al., 2017).
Spectral measurement is an useful and effective method for soil research and can be applied for environmental modeling, soil mapping and agriculture purposes (Soriano-Disla et al., 2014;Viscarra Rossel et al., 2016). This method is relatively inexpensive, non-destructive and allows a range of soil constituents to be determined based on a single spectrum (Viscarra Rossel & Behrens, 2010). VNIR spectra are acquired from satellite, airborne and UAV images, as well as ground and laboratory spectroscopic measurements. However, SSR has also a considerable effect on spectral reflectance, with increasing surface roughness reducing its intensity (Herodowicz, 2017;Piekarczyk et al., 2016) due to the occurrence of shadows on rough surfaces (Cierniewski, 1999;Matthias et al., 2000). In addition, larger aggregates have more irregular shapes, with breaks that trap incident light (Mikhajlova & Orlov, 1986).
Earlier studies on spectral properties of SSR subjected to rainfall concerned mainly very heavy and crusted soils (Anderson & Kuhn, 2008;Ben-Dor et al., 2003;Croft et al., 2009Croft et al., , 2012 or did not include quantitative description of surface roughness (Ballard et al., 2012;Ben-Dor et al., 2003;Chappel et al., 2005Chappel et al., , 2006Chappel et al., , 2007. However, the present study investigates the changes of texturaly light soils surface spectra under gradual smoothing caused by simulated rainfall. The objectives of this research were to develop predictive models for estimating soil roughness parameters (HSD and T 3D ) from VNIR-SWIR spectra, to evaluate the suitability of different pre-processing methods, and to indicate the rank of different spectral bands.

Soil Preparation
Rainfall simulations were carried out on a set of soil samples with different soil properties and with varying initial soil sample roughness. To ensure that they were representative of the Greater Poland region, the soil samples were taken by mixing random subsamples from approximately 25 m 2 from the A horizon of fields near Poznań, western Poland. The samples were collected just after plowing so that the aggregates were intact and underwent as little mechanical breakdown as possible. The textural classification of soil samples were loam (L), sandy loam (SL) and loamy sand (LS).
Under laboratory conditions, the soil materials were air-dried at room temperature. Following this, part of each soil material of the three textural groups was ground and sieved through a 2 mm sieve. The second part of the soil materials, consisting of natural aggregates, were used to form three SSR states: R1 as the lowest soil roughness state, R2 representing medium soil roughness and R3 the greatest roughness. These soils where prepared on circular trays, 28 cm in diameter. All samples were performed in triplicate, that is, three repetitions. The base of the trays had small holes lined with muslin material to drain water and reduce soil loss. A layer of air-dried and sieved soil material approximately 1 cm thick was placed in each tray.
The soil samples with roughness R1 consisted only of the sieved material. The R2 soil samples were created by placing aggregates ranging in size from 1 to 3 cm on the sieved soil material. Finally, the R3 samples were prepared by placing the largest aggregates (diameter 3-5 cm) on the sieved material, and then using aggregates smaller than 3 cm to fill the spaces between them. Efforts were made to make the created roughness levels related to the roughness occurring in the field so that they were representative. SSR was created by the addition of different-sized soil clods and aggregates, and these were clearly visible on the shaded relief DSMs (Figure 3). On the analyzed surfaces, the soil aggregates were scattered randomly and did not simulate any tillage practice; however, the HSD indices for the R3 soil surface (SL and LS) were similar to those of soil surfaces treated with disc harrow, while the HSD for the R3 soil (L) was similar to soil with cultivation aggregate and HSD for the R2 soil surface corresponded to a soil surface after cultivator (Cierniewski et al., 2015;.

Soil Properties
The physical and physicochemical properties of the soil were determined in the ground samples that had been sieved through 2 mm mesh. The soil texture was characterized by the hydrometer method according to standard PN-R-04032 (Polish Committee for Standardization PKN, 1998). The soil pH was obtained in water (1:1), and SOC was determined using oxidation by K 2 Cr 2 O 7 with H 2 SO 4 on a digestion block at 150° for 30 min, followed by titration of excess oxidant using FeSO 4 (Nelson & Sommers, 1996). The calcium carbonate content was obtained by volumetry (International Standard (ISO) 10693, 2002).
The soils chosen for this experiment are characterized by different basic properties (Table 1) typical for cropland of central Poland. The most coarse textured soil type was LS, classified as loamy sand, with the highest sand and the lowest clay and silt contents of the tested soils. The fine-grained texture class was the loam (L), with the highest content of silt and clay and the lowest of sand; this soil also had the highest organic carbon content and pH. Finally, the soil classified as sandy loam (SL) had an intermediate content of sand, silt and clay, and the lowest organic carbon content. All soil types had similar colors, that is, brown to light brown, with little differences in chroma value.

Rainfall Simulations
Two 6MS 05C air-induction flat fan nozzles (Agroplast, Sawin ul. Lubelska 24, Poland) were mounted 0.50 m apart on a movable boom at a fixed height of 2.0 m. The nozzles were then passed over the soil samples, producing a rainfall intensity of 16 mm/hr. The intensity of the simulated rainfall was maintained by keeping the pressure regulator in a constant position. To ensure spatially uniform rainfall on all trays, the entire surface of the tray was sprinkled by collecting water into evenly distributed cups before each simulated sprinkling. Successive levels of rainfall accumulation were obtained with constant rainfall intensity by controlling the duration of rainfall. At the start of the rainfall simulation, the soil samples were placed at a 3° slope ( Figure 1). During the experiment, the trays were exposed to rain eight times every few days so that they were completely dry before the next simulated rainfall (

Soil Surface Roughness Measurements
The SSR was registered photogrammetrically before the first simulated rainfall and then after each subsequent sprinkling. The images of each tray were recorded using a Sony A7 III digital camera, with CMOS matrix of 35.8 × 23.9 mm and resolution of 24.2 megapixels. The images were saved in JPEG format with a very low compression ratio. The estimated spatial resolution of the digital images was approximately 2 mm. The photographed trays were placed on the turntable and after its rotations, an image was taken aimed at their center from an angle of 30° above horizontal ( Figure 2). The total number of images covering a single tray depended on the SSR, but this value ranged from 30 to 40 images. The distance between the camera and the center of the soil samples was about 70 cm. A millimeter paper was placed under each soil sample to calibrate the DSMs to a local coordinate system and check the accuracy of photogrammetric processing. The digital camera was geometrically calibrated before each set of images and the model of geometric distortion was computed.
The images were processed and digital surface models (DSMs) generated using Agisoft Metashape software (Agisoft LLC., 2019) and Structure-from-Motion method. The process consisted of the following steps: (a) selection and import of images; (b) input camera calibration and alignment of images; (c) designation of geometric calibration and checkpoints; (d) camera optimization; (e) a dense point cloud generation; (f) DSM generation (Hendrickx et al., 2019). The geometric calibration model of the camera lens was used for each set of images. In each set of images, the same reference points were highlighted over the millimeter paper to allow local system definition. All calculations were made at high accuracy settings. The DSMs were exported in TIFF format and subjected to height standard deviation (HSD) and area ratio (T 3D ) testing using TNTmips 2018 software (Micro-Images Inc. Nebraska, USA).

Soil Spectra Measurements
Spectral measurements were taken with an ASD spectroradiometer (FieldSpec Analytical Spectral Devices, Inc., Boulder, Colorado, USA), recording reflected radiance in the range 350-2,500 nm with 3 nm spectral resolution in the visible and near infrared (VNIR 350-1,000 nm) range and 10 nm in the shortwave infrared (SWIR 1,300-2,500 nm). The proximal measurements were made using a pistol grip with a 10° Field of View (FOV) fore optics; this was mounted on a tripod at a height of 40 cm above the tray with the soil sample. The tray was illuminated at an angle of 45° to the vertical with a 400W halogen lamp positioned 90 cm from the sample. Each soil sample was measured four times and the tray was rotated 90° after each measurement. For calibration, the reference Spectralon panel (Labsphere Inc., model SRT-99-170, 31 × 31 cm) was used. A reference diffuse spectral reflectance (DSR) value was taken from the samples of the three textural groups prepared for laboratory analysis (ground and sieved) using a MugLite adapter connected to the spectrophotometer.

Statistical Analysis
The relationship between the SSR parameters (T 3D and HSD) and soil spectra was determined using regression random forest models (Breiman, 2001). First, each raw spectrum (RAW) was transformed using six methods: Savitzky-Golay (SG) (Savitzky & Golay, 1964); the first derivative of the Savitzky-Golay transformation (SG1d); the second derivative of the Savitzky-Golay transformation (SG2d); baseline alignment (BA) ; the first derivative of the baseline alignment (BA1d); the second derivative of the baseline alignment (BA2d).  For each of the seven spectra groups, viz. the RAW and six transformation methods given above, the following modeling and validation process was applied. The data set of spectra was split into two sets: a training set (70% of the observations; 681 spectra) and a test set (30% of the observations; 291 spectra). Importantly, the training and test sets contained different independent groups of measurements. The obtained models were assessed based on the coefficient of determination (R 2 ) and symmetric mean absolute percentage error (SMAPE). In the former, R 2 represents the squared correlation coefficient between the observed and predicted value; this value ranges between 0 and 1, with a higher result being better. SMAPE also allows for comparison between models of different spectra groups but focuses on the differences between predicted and actual values; the value indicates the performance of models in relative terms, with a lower score being better.
The training set was then used for two purposes: determining the optimal value of the random forest parameter mtry (Huang & Boutros, 2016) and selecting the 20 most relevant wavelengths with the minimum Redundancy Maximum Relevance (mRMR) method (Ding & Peng, 2003). Both objectives were achieved by 10-fold cross-validation of the training set and evaluation of the models using the SMAPE values. The remaining parameters had constant values: number of trees (500), Minimal Node Size (value 5 for regression problems (i.e., those where we defined T 3D and HSD) and 10 for classification problems (i.e., those where we defined granulometric groups), maximal tree depth (unlimited depth), minimal terminal node size (1). The optimal mtry value and selected predictors were then used to create the final model. Based on the final model, the significance of variables was determined using the unbiased impurity importance method (Nembrini et al., 2018). Ultimately, the final model was applied to the test set, and its prediction was compared with the observed test set values using R 2 and SMAPE measures. The above procedure was applied 14 times: two roughness measures (T 3D and HSD) multiplied by seven groups of spectra.
A second aim of the study was to determine the effectiveness by which the spectra distinguish granulometric groups and whether this ability changes with successive rainfall simulations. A classification random forest model was created for each combination of spectral group and rainfall simulation. First, the spectra data set was split into two sets: a training set (2/3 of the observations) and a test set (1/3 of the observations). The training and test sets contained different repetitions. The training set was used to determine the final model parameters, while the test set was used to measure the accuracy of the obtained results, that is, the proportion of the data that was predicted correctly.

Variation of Soil Surface Roughness Indices
Before the simulated rainfall, the highest HSD and T 3D values at the R3 roughness state were observed for soil L (respectively 16.2 mm and 2.5). This resulted from the higher content of silt, which is considered an important aggregate binding agent and one that promotes the formation of larger aggregates (Barthès et al., 2008;Bronick & Lal, 2005). As the accumulated rainfall increased, the differences between the roughness of R1 and R2 decreased, especially in the coarser soils (SL and LS) ( Figure 3).
In almost all cases, lower soil roughness values were observed after the final rainfall than before the first rainfall, that is, baseline, which is consistent with previous observations (Bertol et al., 2006;Eltz & Norton, 1997;Magunda et al., 1997;Rosa et al., 2012;Stańczyk & Baryła, 2016). This could be attributed to a loss of cohesive forces between soil particles , and the smoothing of the surface due to aggregate breakdown and filling depressions with sediments (Vermang et al., 2013). However, in field conditions, accumulated rainfall may also lead to the development of erosive features, which resulting in an increase in SSR (Hauer et al., 2001;Huang & Bradford, 1992, 1993Wang et al., 2016). In the present study, no erosion rills or erosion gullies were developed due to the slight slope of the soil samples and relatively small area of the trays.
After the first rain, the SSR was found to increase on some soil sample trays with the greatest increase occurring on the coarsest soil (LS) with the highest roughness state (R3), although this was only for the HSD index. This increase after the first simulated rainfall would have been related to the breakdown of aggregates by the energy of the impacted water droplets, while the relatively small sealing of the soil surface. This is the most important factor influencing the changes in soil micro topography (Wesemael et al., 1996). Similar research results were obtained by Panachuki et al. (2010), Rosa et al. (2012) or Xingming et al. (2017), where an initial increase in soil roughness was observed.
After the last rainfall event, when cumulative rainfall reached 117.5 mm, the greatest decrease of both HSD and T 3D indices was recorded on the surfaces of the most coarsely textured soil (LS); decrease was found to be 34% for R3 roughness and 35% for R2 (Figure 4).
It is worth adding that very high similarity was observed between the two indications (Spearman's correlation: 0.92), however, the rainfall effect was more visible in the T 3D values than in the HSD values ( Figure 5). The T 3D value demonstrated a faster rate of decline than HSD; this is to be expected as the T 3D value refers to the "micro" scale, that is, the tortuosity of soil.
Different correlations were observed between HSD and T 3D for each roughness state. While a high correlation was noted between HSD and T 3D in the moderate roughness range (R2), indicating a convergent information load (0.84), this correlation is half the value in the low and high roughness ranges (0.41 for R1 and 0.43 for R3). This indicates that the information expressed by the indices differs, and hence using two complementary indices provides a comprehensive characterization of the SSR.

Variation of Soil Spectra
Before the rainfall simulations were begun, the spectral reflectance of each soil was measured. The analysis did not include wavelengths below 450 nm from all spectra due to noise. Reference DSR values, of roughness-free and dry soil, are presented in Figure 6. In the visible range (wavelengths below 700 nm) the spectra are concave due to relative low content of SOC. In addition, the LS soil demonstrated lower spectral reflectance because of the higher sand content; it also demonstrated two local minima, at around 1,400 nm and 1,900 nm, due to the hydroscopic water content (Ben-Dor & Banin, 1995). The peaks are deeper for the L and SL soils than LS due to higher clay content adsorbing water on its surface (Herodowicz, 2017).
Despite the differences, certain regularities were observed between the spectra of soils with different roughness states and after successive simulated rainfalls (Figure 7). The initial level of roughness appears to have a significant influence on the initial level of the spectra, insofar that a higher roughness state was correlated with a lower spectral value (Cierniewski et al., 2002;Kaźmierowski et al., 2019;Richter et al., 2005;Wu et al., 2009). In addition, the initially roughest soils (R3) demonstrated the greatest differences in spectra level between successive rainfall simulations. Similar changes were noted between successive rainfall simulations were noted for the L and SL soils, with the spectra increasing with successive rainfall doses. No such trend was observed in the case of the LS soil, especially with the roughness R1 and R3; in this case, the spectral level decreases after the first few rainfall doses and rises after the next.

Relationship Between Soil Surface Roughness Parameters and Soil Spectra
The relationships between spectra transformed by the seven methods and the two roughness indices, from baseline to final rainfall, were analyzed by two methods: coefficient of determination (R 2 ) and SMAPE (Table 3). At the calibration step performed before the rainfall, both the T 3D and HSD demonstrated the highest R 2 (0.93 and 0.92, respectively); however, validation found that the T 3D index had a higher determination than the HSD index.
In the test set, the highest R 2 values for T 3D were obtained with the RAW (R 2 = 0.71), BA1d (R 2 = 0.71) and the SG (R 2 = 0.69) spectra. For the HSD, the highest determination was observed for the RAW spectra (R 2 = 0.61), and the spectra transformed by BA and by BA1d (R 2 = 0.59 for both). For both roughness indices, the spectra preprocessed with second-derivative transformations, that is, SG and BA, resulted in clearly smaller R 2 . This suggests that these transformations remove the roughness information from the spectral data. Furthermore, the   Table 3 Coefficients of Determination for Assessing T 3D and HSD Roughness Indices According to Spectral Transformed Method choice of preprocessing method appears to have a more significant influence on the prediction performance of the T 3D index than HSD.
For both indices, the most significant predictors (VIPs-Variable Importance in Projection), that is, wavelengths, can be identified for each predictive model based on RAW and transformed reflectance data. The most significant wavelengths for predicting the roughness index, based on different spectral reflectance transformations, are given in Figure 8.
The VIP wavelength range differed depending on the choice of transformation method. Where both roughness indices were modeled on the basis of RAW spectral data, the VIP wavelengths were the longest, falling in the near infrared range (Figure 8a). The wavelengths fell in the 810-820 nm range for HSD, and from 725 to 803 nm for T 3D . Both spectral ranges have previously been found to encompass the largest differences in reflectance associated with changes in SSR following rainfall . A previous study examining rough soil surfaces from various directions also found wavelengths longer than 700 nm to be best for detecting changes in roughness after artificial rainfall (Croft et al., 2012).
The shorter wavelengths were found to be the most important predictors following spectral transformation. In general, slightly shorter wavelengths were more useful for T 3D prediction than HSD prediction. The shortest wavelengths, that is, in the VIS range below 470 nm, offered the best roughness prediction when using a model based on BA spectra (Figure 8d). For the T 3D prediction, three wavelengths in the 557-616 nm range were clearly the most significant when using the SG transformation (Figure 8b), and those in the 570-573 nm range for the SG1d transformation, which demonstrated much lower VIP score values (Figure 8c). For HSD prediction, the most significant wavelengths using SG transformed spectra were within the 764-772 nm range, with this shifting to 568-594 nm after using the SG1d.
The usefulness of VIS wavelengths to assess changes in soil roughness after rainfall was also previously noted by Anderson and Kuhn (2008) while examining the directional reflectance from the soil surface subjected to artificial rainfall; their findings indicate that changes in roughness were best described in the VIS range by backscattered radiation with a wavelength of 658 nm, recorded at +30° view zenith angle. Chappell et al. (2006) propose that the largest changes in soil spectral reflectance occurring in response to changes in SSR under the influence of rainfall may be related to soil texture. The greatest changes in reflectance occurred in wavebands around 600 nm in the case of fine-textured soil, and in wavelengths of around 470 nm while in coarse textured soils. In other studies, Chappell et al. (2007) indicate that a combination of heavy pounding rain with severe abrasion causes a decrease in roughness and the greatest change in reflectance in the VIS range. The increase in reflectance from these surfaces, compared to untreated surfaces, was most relevant for wavebands shorter than 790 nm, and was associated with the appearance of spectral features at 480 and 510 nm, which occur due to the presence of hematite and goethite minerals.
The SWIR wavelengths were clearly the least important predictors in all the developed models. Among the 20 most significant wavelengths, the most relatively numerous were indicated for predictions based on raw spectral data and subjected to BA transformation. In T 3D and HSD modeling based on RAW data, the significant SWIR wavelengths lay within the 1,778-1,962 nm and 2,209-2,500 nm bands (Figures 8a and 8b). These wavelengths were slightly longer than indicated in previous studies: Croft et al. (2009) found a clear relationship between soil roughness indices and reflectance in the range of 1,550-1,750 nm, while Anderson and Kuhn (2008) found soil surface roughness parameters from directional observations to be most strongly related to wavelengths of 1,700 nm recorded from the backscattering direction at an angle below +15°.
The optimal SWIR ranges obtained in previous studies can be attributed to the presence of OH, H-O-H and SiOH bonds; these diagnostic spectral features are associated with three SWIR wavebands with reflectance peaks near 1,400, 1,900 and 2,200 nm (Parish, 2016;Rice et al., 2013). Our identified SWIR ranges are consistent with the observations of Chappell et al. (2006), who also note greater changes in the reflectance of longer SWIR wavelengths (1,930 and 2,450 nm) in sandy soils experiencing rainfall and abrasion than fine soils.

Relationship Between the Granulometric Group of Soils and Soil Spectra
Simulated rainfall was found to influence changes in soil roughness and spectral reflectance in the three studied soils with different textures. It was possible to accurately determine the texture of these soils on the basis of the spectra recorded after subsequent rainfall treatments (Table 4); however, the accuracy of the results influenced by the choice of applied data processing method. The best results were achieved with the BA method followed by the SG method (Table 4). However, on the basis of the existing data, it cannot be concluded whether the rainfall influences the possibilities of distinguishing soil textural groups.

Conclusion
The study investigated changes in the soil surface roughness (SSR) of three soils with different textures and with various initial surface states following simulated cumulative rainfall. Prior to the start of simulated rainfall, the differences in SSR, expressed by HSD and T 3D indices, between the three roughness states (R1, R2 and R3) were clear. With successive rainfall events, and the gradual smoothing of the surface, the values of both indices decreased and only the HSD showed a separation between roughness states after the last rainfall. A strong correlation index was found between the two roughness indices (r = 0.92); however, this relationship was approximately two times lower for high surface roughness (R3), with HSD >9.3 and T 3D > 1.3. The greatest decrease in roughness after all rainfall events was observed on the most coarse (LS) soil.
The changes in the SSR during the experiment were closely related to the changes in spectral reflectance from the soils; however, the spectral data was more closely related to the T 3D index than the HSD index. These relationships gave the best results, that is, the highest R 2 (0.71) and the lowest SMAPE (9.97), when RAW spectral data was used for T 3D prediction; in this case, the regression random forest models were used. Models based on untransformed spectra indicated that the most useful wavelengths for SSR prediction lay in the NIR range: for T 3D 725-803 nm and for HSD 810-820 nm.
Applying spectral transformation yielded slightly worse results. BA1d conversion provided the same R 2 (0.71) as RAW spectra, but the SMAPE error was higher (10.89). Spectral transformation also resulted in visible wavelengths being the most important predictors for SSR estimation. The shortest wavelengths (450-475 nm-blue range) were found to be most significant in the BA transformed spectra model. Wavelengths in the SWIR range (1,300-2,500 nm) were of relatively little relevance in description of SSR changes after rainfall.
The obtained results could be useful in hydrology, agriculture and remote sensing. The indicated wavelengths will be a valuable part of further research on rough soils after rainfall using the tested data transformation methods.