Vegetation Masking of Remote Sensing Data Aids Machine Learning for Soil Fertility Prediction

Abstract

Soil nutrient content varies spatially across agricultural fields in hard-to-predict ways, particularly in floodplains with complex fluvial depositional history. Satellite reflectance data from the Sentinel-2 (S2) mission provides spatially continuous land reflectance data that can aid model development when used with point observations of nutrients. Reflectance from vegetation is assumed to obstruct land reflectance of bare soil, such that researchers have masked vegetation in models. We developed a routine for masking vegetation within Google Earth Engine (GEE) using Random Forest classification for iterative application to libraries of S2-images. Using gradient boosting, we then developed soil nutrient models for surface soils at a 250-ha agricultural site using S2 images. Soils were sampled at 2145 point locations to a 23-cm depth and analyzed for Ca, K, Mg, P, pH, S, and Zn. Results showed that masking vegetation improved model performance for models from subsets of the data (80% of samples used for model development, 20% validation), but full data sets did not require masking to achieve accuracy. Models of Ca, K, Mg, and S were successful (validation R² > 0.60 to 0.96), but models for pH, P, and Zn failed. Bare soil composite images from S2 data are helpful in predicting soil fertility in low-relief floodplains.

1. Introduction

Soil fertility is important for plants, including crops, and for supporting biological systems in general because crops and terrestrial organisms depend on nutrients that can be found in various quantities in the soil matrix [1]. Soil fertility varies spatially, even within homogenously managed agricultural fields, because of soil physical, chemical, and geological conditions [1,2,3,4]. Detailed soil fertility maps are desirable for effective soil management but are challenging to produce due to the high spatial and temporal variability of soil fertility factors (soil organic matter, mineralogy, plant-available nutrients, etc.), the high costs of soil sample extraction and laboratory analysis for control points, and the often unknown extent to which a limited set of samples can be relied upon to represent contiguous non-sampled soil areas [5,6,7].

Soils in floodplains are spatially complex owing to complex fluvial dynamics that can cause abrupt shifts in soil parent material, texture, and mineralogy even within topographically homogenous, flat landform profiles that show little apparent variability. Such landscapes are often rife with marks from archaic stream meanders and sporadic flood episodes that have deposited heterogeneous material over geologic time [5,8,9]. Land-surface reflectance of electromagnetic radiation is useful in the study of soil spatial and geographic variability [8,10,11,12,13,14,15,16,17,18] and is particularly salient in situations where variability may be unexpectedly high, such as in floodplains [5,9]. In such landscapes, where soil heterogeneity is high, spatial and temporal land-surface reflectance variability can indicate areas of soil variability that are not otherwise obvious [17]. Satellite reflectance data are sensitive both directly and indirectly to several properties related to soil fertility, including organic matter content, moisture content, soil mineralogy, soil depth, and soil texture [10,11,12,17,19,20]. For example, soil organic matter (SOM) is dark, causing soils with higher levels of SOM to have a lower albedo and reflect electromagnetic energy to a lesser extent than soils with lower SOM levels [13]. In turn, soils higher in SOM often have higher concentrations of nutrients. This and other relationships between soil properties and reflectance, such as the effect of soil moisture content in decreasing a soil’s albedo [21], may allow the use of reflectance data to explain the spatial distribution of nutrient contents in soils.

Complicating the interpretation of the signal measured by satellite reflectance from the soil are the variable effects of vegetation, both living and in the form of plant residues [13]. Importantly, plant communities cycle nutrients near the soil surface where they can be more readily available to growing plants, but they also block direct reflectance information from the soil surface [1,14,19]. In one sense, reflectance data from vegetation represents a response to environmental and soil conditions and can provide some indirect information about soil fertility. Where vegetation thrives, it can be inferred that conditions are adequate with respect to soil nutrition. On the other hand, vegetation blocks direct reflectance from bare soil to the satellite sensor and represents a modification of any direct signal from soil components. While the presence of vegetation in an image does not necessarily preclude its use in predictive models, relationships between land reflectance and soil properties are believed to be stronger when bare soil images are used [5,11]. This is because vegetation and crop residues can interfere with information gained about soil properties from reflectance images [13,14]. Conversely, analysis of bare soil may elucidate the most reliable relationships between soil fertility and reflectance because electromagnetic radiation reflected from bare soil gives information about the interaction of radiation with the components of soil with minimal interference. Vegetation obscures reflectance from the soil by covering it and can give a misleading uniformity of reflectance in a field of a single crop with uniform growth, obscuring soil variability. Vegetation can also show variability not pertaining to soil but rather to differences in planting rates, variations in crop choice, or other factors. Furthermore, different vegetation types may have different reflectance characteristics not directly related to soil characteristics. Also, differences in land use can be independent of soil characteristics but nevertheless display marked differences in reflectance.

For these reasons, several researchers have made attempts to remove the vegetation signal from reflectance data of serial imagery, preferentially retaining bare soil information when possible [8,15,16]. Given the relative scarcity of images depicting bare soil, particularly in humid locations where vegetative cover is consistent throughout growing seasons, many scenes may be needed for this approach. The scarcity of images results from humid areas having crop cover, crop residue, snow, or perennial vegetation cover for much of the year. Short windows of cultivation during planting seasons can sometimes offer glimpses of bare soil reflectance characteristics. Some authors used a W-Net approach to discriminate bare soil from other reflectance data, using thresholds of vegetation and wetness indices to mask features that are not bare soil [18]. Several authors mapped SOM during a bare soil period but only used one Sentinel-2 (S2) scene acquired when soils were sampled [22]. The use of time-series satellite data where satellite revisit times permit multiple collections of reflectance data within a single season or year has the advantage that they potentially show dynamic interactions among features such as precipitation, vegetation, soil moisture variability, and soil properties [8,16].

To our knowledge, no research has yet used machine learning supervised classification to automate vegetation masking applied to all available images in an S2 satellite collection to extract bare soil pixels for environmental correlation with soil properties. The objectives of the current study were to (1) determine the importance of remote sensing satellite reflectance data as an environmental covariate in predictive soil mapping of Ca, K, Mg, P, pH, S, and Zn in an agricultural floodplain, (2) develop a machine learning methodology using Google Earth Engine (GEE) for removing the potentially obstructive influence of vegetative pixels, thereby maximizing the information available from a suite of satellite images available for analysis, (3) determine whether pixel-based statistics from spectral data from satellites could be used to estimate soil fertility properties at the site, (4) use machine learning to train artificial intelligence to recognize and classify bare-soil portions of images that could then be analyzed for relationships between bare soil reflectance and soil fertility in the surface, and (5) make models of soil fertility properties for the site and assess them for accuracy. A further objective was to assess the accuracy of soil fertility predictions with and without the automated vegetative masking of images, thereby assessing the usefulness of the vegetation masking technique. To test our new vegetated pixel masking tool, we sought to determine the strength of relationships between soil fertility and reflectance scenes by generating a histogram-based gradient boosting interpolation routine to predict surface Mg, Ca, K, P, S, Zn, and pH in an agricultural area with extensive soil sample density (2145 point-locations). This research should be useful for efforts to generate high-detail, high-resolution geospatial maps of soil fertility properties in floodplains with intermittent vegetation cover.

2. Materials and Methods

2.1. Study Area

The study was conducted in Washington County, in Western Mississippi, USA, at the National Center for Alluvial Aquifer Research in the Mississippi Alluvial Plain. The site is approximately 250 hectares (2.5 km²) (Figure 1), with a mean elevation of approximately 37 m. Elevation is highest at the north end near Deer Creek and decreases only about three meters to the southern end of the site, nearly three kilometers away, in a slow, gradual declension with an average slope of approximately 0.1%. Following the pattern of gradual decrease in elevation, coarse sediments near Dear Creek overlie a thicker clay-rich layer of the broader floodplain, with the coarser levee deposit thinning out towards the south as the elevation decreases [9]. Texture grades from very fine sandy loam soils at the north (Bosket series, a fine-loamy, mixed, active, thermic Mollic Hapludalfs) to clay at the south (Downling, a very-fine, smectitic, nonacid, thermic Vertic Endoaquepts), and Sharkey (a very-fine, smectitic, thermic Chromic Epiaquerts) [23].

The site has a long history of row cropping with corn (Zea mays), cotton (Gossypium spp.), rice (Oryza sativa), and soybeans (Glycine max). Because the soils in the northern portions of the site are better drained, they have historically been cropped with alternating rotations of corn and cotton, while the southern portions of the site, marked by the higher clay soils of Sharkey and Downling, have been cropped with alternating rice and soybean rotations (Figure 1).

2.2. Soil Sampling and Laboratory Analysis

Soils were collected at 2145 point locations at the study site, approximately one sample for every 0.13 ha (i.e., approximately 9 samples per hectare). For each point location, soil from a sampling depth of 0–23 cm from the surface was aggregated, passed through a 2-mm sieve, and air-dried. To measure pH, a portion of each sample was then diluted with water to a 1:2 soil-to-water ratio and evaluated with a calibrated glass electrode. Soil nutrients Ca, K, Mg, P, S, and Zn were evaluated by extraction with the Mehlich-3 extract methods using a 1:10 soil mass:extractant solution volume ratio and analyzed by inductively coupled argon-plasma spectrometry (ICP, Agilent Technologies, Santa Clara, CA, USA) [24,25].

2.3. Data Analysis and Model Development

The analysis of soil, reflectance, and elevation data and development of soil nutrient models was conducted using R and packages tidyverse, caret, terra, sp, sf, clhs, ggpointdensity, and sen2r [26,27,28,29,30,31,32,33,34,35]. Development of the model to acquire S2 satellite reflectance data and to identify and mask vegetation from images was done using Python 3.12 and GEE [36].

2.3.1. Geospatial Data

We used a 1-m resolution digital elevation model made with laser altimetry for the site in Western Mississippi from the United States Geological Survey 3DEP program as a base map [37]. Elevation data were acquired between 24 January 2019 and 29 February 2019 using the Riegl 1560i system capable of acquiring 1.3 million measurements per second mounted on a Piper Navajo twin-engine aircraft [38]. The average point spacing was 2.04 points m⁻² [38]. Light-detection and ranging (LiDAR, or airborne laser altimetry) bare-earth returns were converted to 1-m elevation rasters through a proprietary linear interpolation algorithm applied to a triangulated irregular network with a required vertical accuracy of 0.196 m [37]. These DEM data were made freely available to the public through the USGS website [39].

Reflectance data collected by the MultiSpectral instrument aboard the Sentinel-2A and B satellites of the Copernicus Mission of the European Space Agency [40] were used as input data for the soil fertility models for several reasons. Features such as the wide availability of these data through GEE from 2017 to the present, free and public open access, high spatial resolution 10-m pixels for most bands of the sensor, and high temporal resolution with a 5-day revisit time, made this dataset practical and useful for geospatial modeling purposes [5].

Cloud removal was accomplished programmatically through the cloud probability dataset contained in the Earth Engine Data Catalog, which offers the probability of pixels being cloud. Probabilities in this dataset were calculated by the s2cloudless Python package, which uses a tree-based algorithm (LightGBM) [41] trained with S2 bands to infer cloud probabilities. Pixels with greater than 80% probability of being cloud were computed as non-data for all utilized S2 images.

Some authors developed a methodology for collecting a bare soil mosaic from Landsat images to produce an unfragmented spatially continuous surface, or composite, which could represent unobstructed soil reflectance [15]. Their hypothesis was that the soil surface of their area of interest had been exposed to satellite measurements at least once, and each occurrence from multiple images could be aggregated into a single representation. They used thresholds of univariate spectral indices calculated from mutable band reflectance, including the normalized difference vegetation index (NDVI) to distinguish vegetation from bare soil and the normalized burn ratio 2 (NBR2) to distinguish areas showing influences of straw lignin, burned vegetation, and cellulose from bare soil. All areas flagged as clouds, cloud shadows, vegetation, water, or 6 provisional thresholds of the NBR2 were classed as non-data. The remaining pixels for all available images were summarized by selecting the median value of each pixel from all images for each spatial grid cell, thereby producing the bare-soil composite image. This bare-soil composite image was then used as an environmental covariate to predict soil properties and compared to a set of measured ground truth data. A similar effort was made by other authors to create a bare-soil composite using Landsat-4 through Landsat-8 images from 1985–2017 [42]. These authors aggregated the data to provide a pixel-by-pixel analysis that returned, for each pixel, the most representative instance of a bare pixel according to thresholds established to characterize a bare soil index (BSI) value. Pixels with the best BSI scores were returned as values in the final output images. Models for soil properties of SOM, clay, silt, and sand were then created using the BSI pixel image and multiple linear regression.

The vegetation masking algorithm was developed for this project in GEE using 10,531 hand-labeled data points on nine scenes for training the model. The selection of the training set prioritized scenes distributed throughout the year and containing representative pixels of crops and bare soil. The training set comprised scenes from January, March, April, and October of 2019; February, March, August, and October of 2020; and November 2021. To train the machine learning algorithm to distinguish bare-soil parts of images from vegetated parts of images, it was necessary to choose images that contained pixels depicting both bare soil and vegetated areas. It was assumed that these 9 scenes would provide sufficient training and testing opportunities for developing a classifier and testing it with independent samples. About half of the points were classed by human operators as bare soil (n = 5448) and half were classed as vegetation (n = 5083). Further, 80% of these hand-labeled points were used to train the model, and 20% were retained as validation points for assessing the accuracy of the model.

The Random Forest algorithm implemented within the GEE API was then trained on the 80% training data sampled from the nine training scenes and was applied to the entire available catalog of S2 images for the site. The intent of the model was to predict vegetated pixels of any image within the S2 image catalog and use this classification as a mask for removing vegetated pixels from each available scene, thereby providing scenes with only bare soil pixels for further analysis. In our analysis, we only included scenes from the cropping seasons (between June and October, from 2019 to 2023). The intent was to examine the influence of growing vegetation on the performance of models predicting soil fertility and to compare vegetated scenes to bare soil scenes. Since the presence of vegetation is minimal for much of the year in the studied site, summary statistics of all the scenes for all months do not capture the effects of vegetation very well. For this reason and also to simulate the case where bare soil scenes are hard to obtain, we chose to limit our time span to the five cropping seasons from June to October, during the years 2019–2023.

Vegetation masking was applied to the time series of images in the following way. First, each pixel of each time-series image was classified as vegetated or not vegetated following the vegetation masking algorithm described above. Two datasets were then analyzed separately, those with vegetation pixels masked (masked imagery) and those without any modification to the images (not-masked). Every pixel classed as vegetation was flagged as non-data and removed from the ‘masked’ dataset, but retained in the ‘not-masked’ dataset (Figure 2). The first dataset consisted of composites representing the median and interquartile ranges (IQR) calculated without masking vegetation (not-masked) for the analyzed period (between June and October, from 2019 to 2023). Next, composites were created to represent the median and IQR using vegetation-free pixels (masked) for the analyzed period. Composites were created for bands 2, 3, 4, 5, 6, 7, 8, 11, and 12. Other Sentinel-2 bands, such as bands 1, 9, and 10, were not used because they were assumed not to be relevant to terrestrial analysis of vegetation and bare soil characteristics. Band 1, for instance, gives information relevant to coastal aerosols; band 9 gives information about water vapor; and band 10 gives information about cirrus clouds. Thus, 18 images were produced for each dataset, 9 representing median pixel values for each pixel, and 9 representing the IQR of values from the time series for each pixel. Thus, 36 images were produced without masking vegetation (n = 18) and with vegetation masking (n = 18), representing statistical summaries of the reflectance characteristics of the study site. The median composite images were calculated to represent the central tendency in the time series of the reflectance values, while the IQR composites were intended to capture the range of variability in reflectance over time. IQR was chosen over standard deviation because we believed it would be less susceptible to extreme values. The calculation of the median value was likewise preferred as a measure of central tendency over the mean because it would be less susceptible to extreme values. The difference in values of IQR and mean reflectance are discussed in greater detail in the results section, but briefly, it can be noted here, that pixels depicting vegetation have lower reflectance in the visible part of the spectrum and higher reflectance in the infrared and near-infrared parts of the spectrum. The intensity of reflectance values statistically summarized in the pixel statistics of the ‘masked’ vs. ‘not-masked’ datasets illustrate the differences due to vegetation removal.

2.3.2. Model Development and Assessment

We used a histogram-based gradient boosting model with soil fertility response variables and statistical summaries of S2 surface reflectance as predictor variables [43]. To build soil nutrient maps, point samples were split into training and test sets (80/20%). Nutrient models were tuned by 10-fold cross-validations (CV) repeated 1000 times and tested against the test sets. CV was performed in the training set while varying the learning rate parameter. The learning rate has the purpose of decreasing the rate with which the model fits the data. Gradient boosting is a tree-based algorithm that creates many shallow decision trees (weak learners) sequentially. Each tree in the sequence tries to minimize the residuals of previous trees. Because trees know of the errors committed by their predecessors, they fit the data quickly and easily overfit. The learning rate thus constrains the rate of residual correction so that trees learn meaningful rules instead of convenient, unrealistic ones.

The goodness of fit was assessed using the coefficient of determination (R²) and the root mean square error (RMSE). The best-fit models were then tested against the validation set to identify possible overfitting. Importance scores of band statistics were calculated using permutation importance. Performance scores from permutation tests were calculated using the test set (the 20% split unseen by models). Permutation consists of shuffling the values of one of the variables (thus making the variable meaningless) and then calculating the validation score again (R²). The metric in the x-axis is the decrease in R² after shuffling each variable 1000 times. Since several response variables were found to have similar spatial distributions, the most important variables selected by the algorithm for soil attributes were combined and used as explanatory variables for all response variables.

The effects of masking vegetation in prediction performance were also tested while varying sample density. The goal was to understand if the presence of vegetation affected prediction performance and if the effect was dependent on the number of samples used to train the spatial prediction model. Tests were performed for two sets of bands: visible (B2, B3, and B4), which were found less correlated to response variables, and infrared (B11 and B12), found to be more correlated to response variables, so to investigate the effects of vegetation masking in different parts of the spectrum. This evaluation was done by selecting n samples to train models and validating the model with all the remaining samples (2145 − n). One hundred repetitions were done for each model. First, results were compared for models trained based on masked and not-masked scenes using only the median of the tested bands. Next, IQR bands were added to test whether the results would improve.

3. Results

Soil fertility measurements of the 2145 soil samples taken from the soil surface of the 250-hectare study site are statistically summarized in

Table 1. Statistical summary of soil nutrient content extracted with Mehlich-3 extractant for surface soil samples (0–23 cm) at 2145 locations within the 250-hectare study site in Western Mississippi.

Soil Fertility Attributes	Mean	Std	Min	25%	50%	75%	Max
pH	6.97	0.37	5.66	6.76	7.01	7.22	8.07
	----------------------- ppm -----------------------
P	43.6	24.3	5.0	29	37	52	240
K	222	75.4	86	156	209	287	421
Mg	613	414	162	284	392	1039	1609
Ca	2821	1711	758	1440	2031	4758	6849
S	119	47.4	60	81	95	171	222
Zn	3.25	2.04	0.80	1.80	2.80	4.20	26.4

. With only a few exceptions among the samples, values for pH and concentrations of P, Mg, Ca, and Zn were well within the tolerable range for conditions for agronomic crops [1] (Table 1). Concentrations of S and K, on the other hand, showed potential for deficiencies in some sample locations and the need for nutrient management strategies to correct nutrient imbalances [1]. Histograms of the measured soil nutrients are given in the Supplemental Materials (Figure S1).

Data from the time-series imagery from the S2 mission for the study site is given in

Table 2. Reflectance characteristics of images prepared to represent the temporal dynamics of the site from S2 Satellite data before and after vegetation masking (referred to as not-masked and masked, respectively). Median images of S2 bands were calculated on a per-pixel basis and represent the median values over the time series of images available. IQR images represent the interquartile range of band reflectance values over time. The study site is in Western Mississippi.

	Masked Imagery
Site Image	Mean	Std	Min	25%	50%	75%	Max
Median B2	884	147	527	771	876	991	1445
Median B3	1207	217	738	1024	1216	1370	1844
Median B4	1476	232	900	1299	1484	1650	2188
Median B5	1857	321	1245	1578	1909	2098	2773
Median B6	2212	475	1303	1805	2213	2544	3541
Median B7	2416	510	1398	1974	2426	2781	3839
Median B8	2539	529	1495	2064	2562	2891	3912
Median B11	3499	515	2349	2930	3667	3895	4512
Median B12	2700	475	1779	2253	2755	3086	3946
IQR B2	461	157	189	353	437	520	1046
IQR B3	575	208	213	417	573	667	1343
IQR B4	741	247	194	560	724	867	1566
IQR B5	769	277	173	539	770	935	1716
IQR B6	921	400	268	592	837	1153	2185
IQR B7	976	415	301	648	872	1243	2485
IQR B8	976	383	242	653	918	1235	2297
IQR B11	1080	347	360	797	1067	1310	2158
IQR B12	1171	475	267	721	1181	1455	2765
	Not-masked imagery
Site Image	Mean	Std	Min	25%	50%	75%	Max
Median B2	569	96	364	502	565	620	986
Median B3	844	120	607	752	831	919	1398
Median B4	854	187	393	750	852	955	1594
Median B5	1368	190	994	1225	1347	1512	2300
Median B6	2731	475	1517	2387	2793	3054	3858
Median B7	3166	610	1681	2674	3260	3625	4528
Median B8	3257	621	1743	2736	3396	3731	4564
Median B11	2800	254	2248	2589	2795	2993	3854
Median B12	1799	275	1239	1590	1782	1985	3123
IQR B2	491	125	208	390	483	586	1096
IQR B3	553	165	224	417	543	672	1248
IQR B4	975	232	342	794	977	1147	1736
IQR B5	800	255	289	592	776	968	1541
IQR B6	1287	326	407	1049	1292	1502	2249
IQR B7	1865	486	726	1542	1804	2126	3491
IQR B8	1796	446	612	1510	1740	2021	3315
IQR B11	1165	374	314	813	1195	1429	2130
IQR B12	1405	411	467	1072	1409	1703	2738

for both original images and those that have had vegetation masking, removing pixels depicting vegetation by means of the masking algorithm. S2 reflectance imagery without vegetation masking had lower mean reflectance values for bands 2, 3, 4, 5, 11, and 12 than imagery with vegetation masked (Table 2). Bands 6, 7, and 8, however, showed stronger reflectance values in the images without vegetation masking, indicating the strong tendency of vegetation to reflect light in the near-infrared part of the spectrum while absorbing light in the visible part of the spectrum. In all bands except band 3, the IQR of the images that did not undergo vegetation masking was greater than for the images with masked vegetation (Table 2). Vegetation removal through the masking of pixels depicting vegetation after classification by the vegetation classification algorithm preferentially removes pixels depicting low reflectance in the visible portion of the spectrum while retaining pixels depicting high reflectance in the infrared and near-infrared portions of the spectrum because of the photon scattering characteristics of leaf mesophyll. This accounts for the differences in pixel statistics shown in Table 2. When vegetation is removed (‘masked’), values in the visible bands (bands 2, 3, and 4) are relatively increased and the infrared (bands 6, 7, and 8) are relatively decreased, while when vegetation pixels are retained (‘not-masked’) the effect is the opposite.

Reflectance among all bands showed a greater correlation to soil properties after the vegetation masking routine was applied (Figure 3). Phosphorus showed a particularly poor correlation to satellite reflectance before vegetation was masked but improved markedly after vegetation masking. Correlations to soil nutrient concentrations were not strong in either dataset within the B6 to B8 range, but they improved to significant levels (α = 5%) after the application of the vegetation masking routine. Correlations among S2 bands given in the Supplemental Materials (Figure S2) showed high degrees of correlation among adjacent bands, with greater significance of correlations among masked imagery than the not-masked imagery.

The vegetation masking algorithm implemented in GEE from the hand-selected training data was highly accurate (>99% accuracy, Figure 4). The y-axis indicates the number of observations giving the true value of the classification, while the x-axis indicates the number of observations giving the predicted value of the classification. For instance, 1352 observations were correctly predicted to be vegetation, and 1 observation was predicted to be soil when it was in fact vegetation. The remaining 1904 validation observations were correctly classified. No pixels predicted to be vegetation were found to be bare soil pixels. Seasonal changes associated with cropping patterns and changes after planting were evident after applying the vegetation masking model to the images in the Sentinel library within analyzed dates (between June and October, from 2019 to 2023) (Figure 5). The image from July 2021 shows similarities to the image from July 2023 in terms of patterns of where crops have been planted and their stages of growth and greening during the growing season, but some areas that have had bare soil are showing likely vegetation and vice versa. The vegetation masking model captures these trends. Likewise, the large change in greening from June 2023 to late July 2023 captures a quick change in vegetation cover associated with crop growth from the early to the middle parts of the growing season for the crops grown at the site.

The next two figures (Figure 6 and Figure 7) show the influence of sample size on the effectiveness of models built with S2 data that underwent vegetation masking compared to models developed with not-masked S2 data. Overall, as the sample size increased to about 40 samples, model performance improved greatly for K, Mg, Ca, and S. The model improvement was greatest for the masked datasets when using only median bands or when using both median and IQR bands as explanatory variables. Models trained with not-masked median data presented the lowest performance per sample density increase, but their performance noticeably improved after IQR data was added. Still, models that used not-masked data generated less accurate predictions than their masked counterparts.

Model performance for prediction of soil variables using spectral bands in the visible portion of the spectrum showed R² values approaching 0.75 for Mg, Ca, and S, the three analytes that were the most predictable as the sample size was increased to 100 samples (Figure 6). Potassium, while being less predictable than Mg, Ca, and S, nevertheless showed a strong increase in model strength when the median reflectance composites were masked. The variables pH, P, and Zn were not adequately predicted by spectral bands in the visible range using this modeling approach. Models built using the median pixel values of the time series of images after masking to remove pixels depicting vegetation (labeled masked median in Figure 6) were superior to their not-masked counterparts in all models. Like models in the visible range of the spectrum, prediction performance using spectral bands in the short-wave infrared portions of the spectrum showed R² values approaching or higher than 0.75 for Mg, Ca, and S, and values approaching 0.6 for K (Figure 7). Overall, not-masked median bands performed better in shortwave bands than in the visible portion of the spectrum. Here too, the masked S2 reflectance data supported better model performance for nutrient predictions than the not-masked S2 data, but the improvement from masking was more modest.

Models using the entire available soil point data with the S2 reflectance data (1716 point samples for training and 429 for independent validation) performed very well as expected for several soil nutrients but not well for others (

Table 3. Histogram-based gradient boosting model performance for predicting measured soil properties using all the sampled soil point data built from time-series analysis of S2 scenes to predict soil fertility characteristics using (a) S2 scenes after application of a vegetation masking algorithm, and (b) S2 scenes with no vegetation masking algorithm.

(a) Masked
	Training dataset (n = 1716)					Test dataset (n = 429)
Fertility attribute	R2	R2 std	RMSE	RMSE std	MAE	R2	RMSE	MAE
pH	0.34	0.08	0.30	0.02	0.04	−0.65	0.31	0.23
P	0.28	0.09	20.3	2.95	5.26	−1.05	20.1	13.2
K	0.70	0.05	40.8	3.9	3.58	0.62	40.1	29.6
Mg	0.95	0.01	95.4	10.26	0.57	0.95	86.5	59.1
Ca	0.93	0.01	445	40.62	67.15	0.94	404	303
S	0.95	0.01	10.4	0.89	0.89	0.95	9.71	7.35
Zn	0.35	0.09	1.66	0.33	0.57	−0.48	1.41	0.77
(b) Not-Masked
	Training dataset (n = 1716)					Test dataset (n = 429)
Fertility attribute	R2	R2 std	RMSE	RMSE std	MAE	R2	RMSE	MAE
pH	0.39	0.07	0.29	0.02	0.07	−0.28	0.29	0.21
P	0.34	0.10	19.3	3.18	5.14	−0.47	20.1	13.1
K	0.74	0.04	38.3	2.94	11.7	0.65	37.9	28.4
Mg	0.95	0.01	94.9	9.48	6.50	0.96	81.1	58.8
Ca	0.93	0.01	438	25.3	63.0	0.94	381	293
S	0.96	0.01	10.0	0.63	1.53	0.95	10.1	7.38
Zn	0.38	0.12	1.62	0.35	0.47	−0.20	1.41	0.82

).

Map visualizations of model output for predicted soil nutrients with high prediction efficacy showed strong spatial aggregation for Ca, K, Mg, and S (Figure 8). Models for P, Zn, and pH, however, showed little spatial aggregation and almost no prediction efficacy. This implies satellite reflectance data had very little, if any, ability to elucidate spatial patterns of P, Zn, and pH at the site. For Ca, K, Mg, and S, strong and accurate predictions showed bimodal distributions, with fields at the south of the study site having much higher levels of nutrients. Furthermore, the spatial patterns of these nutrients showed a high degree of aggregation coincident with the boundaries of fields, implying a strong relationship between field nutrient management and reflectance characteristics. Spatial visualizations of band reflectance data after vegetation masking are shown in the Supplemental Materials (Figure S3).

The relative importance in the models of the composite median and IQR images calculated from the time series of S2 data indicates that some bands are more useful for distinguishing soil fertility variability than others (Figure 9). The most useful bands for modeling were from the short-wave infrared range, which are also the bands that were least influenced by variability in vegetation characteristics. The models with the best fit were most strongly aided by median composite bands 11 and 8 in their ability to minimize the decrease in R² in nutrient predictions. Median values of bands 12, 6, and 7 were also among the most important, as was the IQR of band 12.

4. Discussion

Satellite imagery is a useful proxy for examining the variability of soil fertility in a floodplain agricultural soil because soil properties that relate to soil fertility variability influence the reflectance characteristics of the soil surface. The variables that were reliably predicted by models at the study site were Ca, Mg, K, and S. These four nutrients are bound within the cation exchange capacity (CEC) of soils and strongly associated with the organic matter complexes in relatively high numbers, and their concentrations often follow the trend of the variability of soil characteristics that dominate CEC, namely SOM and clay content [1]. In A horizons, in most temperate regions, a vast majority of the S in soils is in the organic form, meaning that S content should be strongly correlated with soil organic matter content [1]. Soils at the southern end of the study have greater clay content, greater SOM content, poorer drainage, and darker color. Increases in SOM, clay content, and moisture all decrease a soil’s albedo [14,21], thereby supporting a strong relationship between reflectance intensity as measured by satellite and measured soil properties dominated by CEC. This reliable relationship was useful for modeling soil nutrient spatial variability.

Prior to vegetation masking, the variability of reflectance due to vegetation was higher, as seen in the higher IQR of the reflectance data (Table 2). The IQR from bands 6, 7, and 8 were strongly influenced by the reflectance of infrared light from vegetation. The increased variability of these bands was not meaningfully related to soil fertility and likely reflected variability in vegetation characteristics not pertaining to soil fertility but perhaps to crop choice, variable planting dates, cropping intensity, or other factors (Figure 3). One exception to this observation is the IQR of the not-masked image representing band 5, which is better correlated to soil fertility variables than its masked counterpart. A similar effect was seen in band 11 and band 3.

Interestingly, the advantage conferred by masking vegetation (as seen in Figure 6 and Figure 7) did not translate to the full dataset. In fact, model performance for the models using not-masked S2 data was just as good as, or sometimes better than, models from the vegetation-masked dataset. This is likely because the density of soil samples was so high that nearly any spatial estimation method would likely produce relatively successful models.

The variables that were not as successfully modeled, pH, P, and Zn, have a more complex relationship to reflectance data and do not show reliable correlations at the site. Both pH and P were highly managed variables. P is a macronutrient that can have field-to-field variability relating to applications of P fertilizers related to management goals and cropping requirements. Soil pH is actively managed with lime and can vary greatly depending on management strategies based on crop tolerances.

The number of soil observations taken at this study site was intended to comprehensively characterize the soil nutrient variability. The sampling density was far greater than typical for such agricultural studies and was, in retrospect, excessive to the needs of the study. In a separate study, researchers looking at this dataset estimated that the sampling density required for adequate characterization of the soil variability for this site was on the order of 1 sample per 2.5 hectares, or about 100 samples for the site [5]. By exceeding this sample density by more than an order of magnitude, we were able to fully characterize the soil nutrient variability and determine the advantage conferred by using satellite reflectance data for continuous spatial estimations. Our data show that when developing models using fewer than 100 samples for the site, masking vegetation as non-data confers a distinct improvement in model performance for the respective nutrients investigated. This advantage is diminished as sampling density increases to the full dataset of 1716 samples used for model development and 429 samples used for model validation. At this high density, any confounding effects of including vegetated pixels in the model were outweighed by the greater density of data.

Masking vegetation can improve soil nutrient models built on satellite reflectance data and a limited set of soil observations. Reflectance data from the surface is documented to be different for different soil properties, so reflectance properties provide real information about the soils. Vegetation cover blocks the satellite view of the soil surface and degrades the effectiveness of the ability of the gradient boosting model to predict soil fertility response variables. Masking greatly improved models produced by datasets with point samples intentionally reduced well below the complete dataset. In all bands except Band 3, the IQR of images that did not undergo vegetation masking was greater than for the images with masked vegetation, implying greater consistency and homogeneity in the scenes with vegetation masked as non-data. This gives further evidence that S2 reflectance data provides useful covariate information for predicting soil properties of K, Mg, Ca, and S, especially when masking vegetation.

5. Conclusions

Composite images representing time-series statistics from spectral data from satellites were useful for predicting the spatial variability of soil fertility measurements for K, Mg, Ca, and S at a low-relief floodplain agricultural site. Further investigation is needed to determine if other methods can estimate pH, P, and Zn. The vegetation masking procedure described herein using GEE was a successful approach to isolating bare soil pixels that best-represented reflectance data for soils versus soils and vegetation combined. As such, the method described could be readily replicated and can be applied using S2 or other satellite-based remotely sensed data products at other site locations. The application scope of this method is directed toward floodplain landscapes with soil variability that can be related to reflectance characteristics where vegetation potentially hinders direct observation of soil reflectance. The masking technique allows for automated summary statistics of reflectance values from scenes during unvegetated intervals. Further application and analysis in other landscapes and land uses is warranted. Using time-series pixel-based statistics of satellite reflectance data over time as predictor variables also has a strong utility by providing covariates for predicting nutrient concentrations at the site, particularly soil nutrients associated with the CEC complex. When modeling pixel-based statistics of reflectance variables over time to predict soil nutrients, vegetation interferes with prediction, and masking it could improve nutrient concentration predictions.