Fine-Scale Mapping of Soil Organic Matter in Agricultural Soils Using UAVs and Machine Learning

Abstract

The fine-scale mapping of soil organic matter (SOM) in croplands is vital for the sustainable management of soil. Traditionally, SOM mapping relies on laboratory methods that are labor-intensive and costly. Recent advances in unmanned aerial vehicles (UAVs) afford new opportunities for rapid and low-cost SOM mapping at the field scale. However, the conversion from UAV measurements to SOM maps requires specific transfer models that still rely on local sampling. This study aimed to develop a method for predicting topsoil SOM at a high resolution on the field scale based on soil color information gained from low-altitude UAV imagery and machine learning. For this, we performed a UAV survey in cropland within the German loess belt. We used two fields, one for training and one for validation of the model, to test the model transferability. We analyzed 91 soil samples for SOM in the laboratory for the model calibration and 8 additional samples for external model validation. A random forest model (RF) showed good performance for the prediction of SOM based on UAV-derived color information with an RMSE of 0.13% and with an RPIQ of 2.42. The RF model was used to predict SOM at a point-support of 1 × 1 m. The SOM map revealed spatial patterns within the fields with a uniform spread of the prediction uncertainty. The validation of the model performed similarly to the calibration with an RMSE of 0.12% and an RPIQ of 2.05, albeit with a slight bias of 0.05%. This validation using external data showed that prediction models are transferable to neighboring fields, thus permitting the prediction on larger scale farms or enabling carbon monitoring over time.

1. Introduction

Soil organic matter (SOM) content, alongside its quality and dynamics, is one of the major indicators for the assessment of soil functions and ecosystem services [1]. SOM is vital for the sustainable management of soil via its central role in biogeochemical cycling, its ability to purify and retain water and nutrients, and as a driver of soil productivity and carbon sequestration [1,2,3,4]. Soil carbon is recognized by the European Union as a key property for soil quality and as one of the major threats to soil degradation and food security [5]. Further, soil carbon sequestration matters in mitigating climate change [4]. However, land management practices have shown to alter soil carbon dynamics, ultimately bringing it to a new equilibrium over the timeframe of several decades [6]. It is therefore a necessity to understand and monitor the spatio-temporal variability of soil carbon for effective and sustainable management of agricultural resources and carbon sequestration strategies [7].

Traditionally, soil carbon mapping relies on conventional laboratory methods that are labor-intensive, time-consuming, and costly [7]. With the emersion of proximal and remote sensing, new techniques are available for more efficient detection of various soil physiochemical properties and their predictive mapping [8,9,10]. The combined development of new spectroscopic sensor technology, increases in computational power, and advances in machine learning widely increased the availability of georeferenced data, which led to the emersion and success of digital soil mapping (DSM) in recent decades [7,11].

Such proximal and remote sensing approaches have been successfully adopted for the evaluation of soil organic carbon (SOC) content on different scales [12]. For proximal sensing, a sensor is in physical contact or close range (within 2 m) to the soil [13] while for remote sensing, electromagnetic radiation is measured without physical contact with the soil, generally using airborne or spaceborne systems [14]. A broad range of soil spectroscopic methods are now routinely used in the laboratory as a substitute for conventional analysis. Those methods typically rely on diagnostic absorptions in the visible-near-shortwave infrared (Vis-NIR-SWIR: 400–2500 nm) spectral region [15]. Soil spectroscopy is a rapid and non-destructive method that correlates the reflected radiation of soils to soil properties by using multivariate statistical methods such as machine learning [16]. More recently, the advances in soil spectroscopy have been translated to proximal [17,18,19,20] and remote sensing [5,21,22] platforms for the mapping of SOC at a larger scale [2]. While proximal soil sensing is utilized to map soil properties at the profile to field scale, remote sensing is mainly used for the characterization across larger scales, from field to global [23]. With the recent advent of unmanned aerial vehicles (UAVs) coupled with multispectral camera systems, this new technology is opening up its possibilities as a platform for soil mapping and precision agriculture right at the intersection between proximal and remote sensing. The main advantage of UAVs is that they can be deployed flexibly, compared to an air or spaceborne system, and can cover larger areas at a high spatio-temporal resolution compared to on-the-go sensors [24]. Due to the high resolution of UAV-based imagery and small impact of atmospheric factors, UAVs show great potential to capture field-scale variability of soil parameters [2].

With high precision, high flexibility, and nondestructive remote sensing systems, UAVs have been widely used to monitor crop vitality or biomass [25]. This includes studies, e.g., for the estimation of aboveground biomass in crops [26,27] or forests [28] and also the acquisition of classical crop growth or health indices [29,30]. However, few studies used UAVs to assess parameters of bare soils. In one of the first studies, Aldana-Jague et al. [24] used UAV-based multispectral imagery to map soil carbon. In a later study, Biney et al. [14] compared UAV-based SOC mapping with proximal and spaceborne sensing approaches. Zhang et al. [2] lately explored the capability of a UAV-borne spectrometer for SOC mapping in bare cropland. Those studies revealed the potential of UAVs in DSM as accurate and high-resolution tools in precision agriculture.

While most of the recent UAV-based studies used spectroscopic or hyperspectral sensors over the full Vis-NIR-SWIR region, some authors have shown that the vis part of the light spectrum, i.e., soil color, recorded as tristimulus values with a digital camera can give SOC predictions as good as full spectrum sensors [31,32]. Viscarra Rossel et al. [33] and Levin et al. [34] made first efforts using digital cameras to predict soil properties. Digital photography offers a cost-effective way to record color information at a very high spatial resolution [32]. This resolution has been used to characterize soils at the microscale [35,36], the profile scale [37], or the field scale [38]. However, at the field scale, soil color can only be recorded for selective samples. Here, UAVs open up a whole new possibility of field-wide high-resolution images that can be utilized for DSM approaches solely on color information.

The specific objective of this study was, therefore, to develop a method to predict topsoil SOM at a high resolution at the field scale based on soil color information acquired from low-altitude UAV imagery. To achieve this, our aims were (i) to develop SOM prediction models based on different soil color models and indices, (ii) assess the accuracy of the spatial SOM estimation, and (iii) show the potential of UAV-based imagery in DSM. With this study, we built on previous research that used soil color for soil carbon prediction in the laboratory or proximal sensing approaches, taking it to a next level in a UAV-based remote sensing application for DSM or precision agriculture approaches.

2. Materials and Methods

2.1. Study Area

The study area is located within the German loess belt in the western part of the Lower Rhine Bay at the foothills of the Eifel low mountain range. The area is in the northern part of the Euskirchen district at 50.73N and 6.74E (Figure 1a). The dominant soil types in the area are (haplic) Luvisols on periglacial loess deposits characterized by a silt loam texture. The soils in the area are well-drained and have a very good nutrient supply. The climate is temperate oceanic with mean temperatures between 2.5 (January) and 18 °C (July) and a mean annual precipitation of 573 mm (DWD, German Meteorological Service).

2.2. Data Collection

The data collection took place between May (UAV survey) and November 2021 (soil sampling). For this study, we used two adjacent cropland fields (A: 8.5 ha and B: 4 ha) separated by a farm road (Figure 1b). Both fields are conventionally cultivated with typical winter wheat, winter barley, and sugar beet rotation. In April 2021, both fields were cultivated with sugar beets. The preceding crop was winter barley followed by white mustard as a cover crop. The fields were plowed in fall 2020. Before the sowing of the sugar beets, the fields were harrowed in spring 2021. As an anomaly, 1 ha in the SE part of the larger field was used as a 40 m wide green strip from 2010 until 2020 as part of a state nature protection measure. The first tillage on this plot after 10 years was carried out in the fall of 2020. The topography of the fields is characterized by slightly SE sloping terrain with an average slope of about 5%. The range in elevation is between 151 and 142 m ASL (Figure 1c).

2.2.1. UAV-Based Survey

The UAV data were collected on 31 May 2021 on a sunny and cloudless day. The month before the survey, there was light rainfall reported almost every day with a total of 60 mm in May, so the soil was moist but not wet nor saturated. Because the weather caused a delay, the sugar beets were already in their early leaf development period. Two flights were made (one for each field) between 12:00 and 13:00 close to solar noon with a Matrice 200 Series Quadcopter (DJI Technology, Shenzhen, China) carrying the five-band multispectral camera system RedEdge-M (MicaSense, Seattle, WA, USA). As a multispectral device, the camera simultaneously records five discrete spectral bands: a blue (465–485 nm), a green (550–570 nm), a red (663–673 nm), a red edge (713–723 nm), and an NIR band (820–840 nm). Each sensor has an image resolution of 1280 × 960 pixels and a field of view of 47.2°. The camera was coupled to a downwelling light sensor to correct for changes in illumination. Before and after each flight, a calibrated reflectance panel was recorded for radiometric calibration. The fields were captured parallel to their longest side at a flight height of 75 m AGL and a flight speed of 5 m/s. The overlap between images was set to 70% to the front and side. This resulted in a ground sampling distance of 5.5 cm/pixel. For an accurate geometric correction, 5 ground control points (GCPs) were deployed across the edges of the study area. The positions of GCPs were measured with a D-RTK 2 real-time kinematic GNSS mobile station (DJI Technology) with a precision of 2 cm.

2.2.2. Soil Sampling

The soil sampling was conducted in November 2021 right after harvesting since the sugar beets had already sprouted in May. A set of 99 soil samples were collected from the A-horizon (0–20 cm) using a soil probe: 91 samples from field A in a regular grid and 8 samples from field B in an east to west transect. The sampling was designed to train a model based on a dense grid (field A) and validate the model with few carefully chosen samples (field B). The exact sample positions were recorded by the D-RTK 2 real-time kinematic GNSS mobile station with a precision of 2 cm. The soil samples were dried at 30 °C, ground, sieved (<2 mm), and thoroughly mixed for further analysis. SOM was determined using the loss-on-ignition (LOI) technique. LOI was measured as the weight loss at 500 °C for 4 h after oven drying at 105 °C overnight. The soil samples were taken under the assumption that SOM changes between UAV flight and the sampling (6 months) were not significant [2]. This assumption is supported by the very slow rate of SOM change reported for the western European loess region [39].

2.3. Digital Image Processing

The aim of image post-processing is the translation of digital numbers (DNs) into georeferenced and calibrated reflectance images. The photogrammetric processing of the aerial images was carried out in Pix4Dmapper (Pix4D, Lausanne, Switzerland). As a proprietary software, the underlying algorithms are not known in detail. However, the workflow follows the common photogrammetric operations of a Structure from Motion (SfM) workflow. This workflow includes an approximation of camera positions and orientations, geometric image correction, and the generation of point clouds comprised of matching points between overlapping images. After densification of the point cloud to the original image resolution, the images were merged into single orthomosaics per spectral band. The DNs of the images (recorded at 16-bit resolution) were converted into radiometrically corrected reflectance values by the use of the images of the reflectance target and the irradiation sensor data. The final reflectance images had a root mean squared error (RMSE) of 1.5, 1.2, and 1.6 m for the geolocation errors, i.e., difference in geolocation between initial and processed images, in their X, Y, and Z dimensions, respectively. The final orthomosaicked images for each spectral band were exported to the GeoTiff format in the original resolution of 5.5 cm/pixel for further analysis.

2.4. Image Analysis

In a first step, the reflectance images were cropped to the region of interests (fields A and B). As mentioned above, the flights had to be postponed for several weeks due to weather conditions so the sugar beets had already sprouted. Thus, we had to remove the vegetation from the soil background. Since the plants were still small, this was achieved using a simple threshold method. The differences in the reflectance between the soil and leaves were highest in the red band and a reflectance of 0.21 was found to differentiate best between the two. This created mask was used to remove pixels containing vegetation information from the reflectance maps. Subsequently, the resulting gaps were interpolated using the gdal fill nodata function in QGIS with a maximum search radius of 2 m and no smoothing.

For this study, we chose a support grid of 1 × 1 m for the DSM. Although UAV images provide a much higher spatial resolution, we chose this resolution as a good tradeoff between informational gain, possible applicability for farmers, and computational cost. At very high resolutions, artifacts resulting from shading due to tractor track marks, soil aggregates, or stones can influence the interpretability of the results. The resampling of images to a lower spatial resolution can mask those artifacts by averaging and can also improve the signal-to-noise ratio (SNR) of the images [38]. Additionally, this support resolution is compatible with free products available from the state survey North Rhine-Westphalia, e.g., a digital terrain model.

In this study, we used soil color as a proxy for SOM. Therefore, we converted the reflectance maps into the RGB color space. The reflectance in the red, green, and blue bands were multiplied by 255 and an additional correction factor of 3 according to Wadoux et al. [40]. Those 3 maps (R, G, and B) were transformed into different color space models (e.g., CIELab, CIELCh, HSV, CIE XYZ). In addition to the tristimulus color spaces, several color indices were calculated. Those indices are mostly ratios of RGB color values, so they are expected to reduce shading and bidirectional reflectance effects [34]. All color spaces and indices used are shown in

Table 1. Details on the 24 unique color variables of the used color space models and color indices.

Name	Abbreviation	Details	Reference
Red	R	RGB red channel	Berns [43]
Green	G	RGB green channel	Berns [43]
Blue	B	RGB blue channel	Berns [43]
Hue	H	Cylindrical color coordinate	Persson [42]
Saturation	S	Color intensity range	Persson [42]
Value	V	brightness/luminance factor	Persson [42]
Lightness	L	CIE metric representing brightness	Berns [43]
a*	a	CIE chromatic coordinate red-green	Berns [43]
b*	b	CIE chromatic coordinate blue-yellow	Berns [43]
C*	C	CIE cylindrical color coordinate	Berns [43]
h	h	CIE color saturation variable	Berns [43]
X	X	CIE chromatic variable	Berns [43]
Y	Y	CIE brightness/luminance factor	Berns [43]
Z	Z	CIE chromatic variable	Berns [43]
Normalized red	Rn	R/(R + G + B)	Kawashima et al. [44]
Normalized green	Gn	G/(R + G + B)	Kawashima et al. [44]
Normalized blue	Bn	B/(R + G + B)	Kawashima et al. [44]
Brightness index	BI	sqrt [(R2 + G2 + B2)/3]	Levin et al. [34]
Coloration index	CI	(R − G)/(R + G)	Levin et al. [34]
Hue index	HI	(2 × R − G − B)/(G − B)	Levin et al. [34]
Redness Index	RI	R2 / (B × G3)	Levin et al. [34]
Saturation Index	SI	(R − B)/(R+ B)	Levin et al. [34]
Green red difference index	GRDI	(G − R)/(G + R)	Tucker [45]
Dark green color index	DGCR	[(H − 60)/60 + (1 − S) + (1 − B)]/3	Karcher et al. [46]

. A more detailed overview of the color variables used is given by Heil et al. [32] and Gholizadeh et al. [38]. All color space transformations were calculated in R with custom functions based on the equations given in Viscarra Rossel et al. [41] and Persson [42] for the HSV color space.

2.5. Predictive SOM Modeling

The modeling techniques used in this study include partial least square regression (PLSR), the ensemble algorithm random forest (RF), and artificial neural networks (ANNs). Field A served as the training field while field B was used for validation purposes only. For the training dataset of the modeling, color and topographic covariates were extracted from all created raster maps for the 91 sample locations of field A as an interpolated value from the four nearest raster cells.

PLSR is a multivariate method that is conceptually similar to principal component analysis (PCA). It combines PCA with a multiple linear regression by specifying the linear relationship between many dependent and independent variables, covering the maximum covariance between variables [47]. RF is one of the latest improvements in ensemble learning [48]. It is similar to a decision tree but builds on manifold regression trees generated on a combination of bootstrap aggregation and a random subset of input variables at each split. It is widely used in environmental sciences as a non-parametric and non-linear machine learning approach. The basic concept of ANNs resembles the data processing in biological nervous systems [49]. In an ANN, some cells are used for information reception and others for forwarding and storing information, organized in several layers. The weights between the cells are optimized in the learning phase based on the backpropagation of the learning error [50].

Before fitting the machine learning models between SOM and the covariates, all models were fine-tuned using the caret packet. For the PLSR, the hyperparameter ncomp, the number of components, was used for the model building. For RF, we tuned the mtry parameter, the number of randomly selected predictor variables. The number of trees grown (ntree) was held constant at 500, as it mainly influences the computational time only. The tuning of an ANN is more complex than the other two algorithms, thus various hyperparameters had to be considered. We used a multilayer perceptron from the Keras library in R. The size parameter of the number of hidden neurons, the dropout rate, the batch size, the learning rate with decay, the activation function, and the rho parameter were tuned. All model runs used 250 epochs, i.e., the number of times the ANN sees the entire dataset. A grid search for the PLSR and RF and a random search for the ANN were used to find the optimal hyperparameters validated using 10-fold cross-validation.

2.6. Model Validation

As we had only 91 samples available from field A and used field B for external validation in a later step, we did not split the data into training and validation data. We used 10-fold cross-validation to evaluate the point support predictions. This modeling was repeated 50 times for each algorithm to obtain a distribution of the accuracy of prediction. We averaged the 50 realizations to obtain the overall model performance. The model performance metrics used are the root mean square error (RMSE), the mean error or bias (ME), the coefficient of determination (R²), Lin’s concordance correlation coefficient (CCC; [51]), and the ratio of performance to the interquartile range (RPIQ; [52]), according to the following formula:

$$\mathrm{ME}=\frac{1}{\mathrm{n}}\sum _{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{obs}}_{\mathrm{i}}-{\mathrm{pred}}_{\mathrm{i}})$$

(1)

$$\mathrm{RMSE}=\sqrt{\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{obs}}_{\mathrm{i}}-{\mathrm{pred}}_{\mathrm{i}}\right)}^2}{\mathrm{n}}}$$

(2)

$${\mathrm{R}}^2=1-\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{obs}}_{\mathrm{i}}-{\mathrm{pred}}_{\mathrm{i}})}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{obs}}_{\mathrm{i}}-\overline{\mathrm{obs}})}^2}$$

(3)

$$\mathrm{CCC}=\frac{2{\mathrm{r}\mathrm{s}}_{\mathrm{obs}}{\mathrm{s}}_{\mathrm{pred}}}{{\mathrm{s}}_{\mathrm{obs}}^2+{\mathrm{s}}_{\mathrm{pred}}^2+\left(\overline{\mathrm{obs}}-\overline{\mathrm{pred}}\right)}$$

(4)

$$\mathrm{RPIQ}=\frac{\mathrm{IQR}}{\mathrm{RMSE}}$$

(5)

where n is the number of samples, obs_i and pred_i are the observed and predicted values at position i; ($\overline{\mathrm{obs}}$) and ($\overline{\mathrm{pred}}$) are the mean of the observed and predicted values; s_obs and s_pred are the variance of the observed and predicted values; r the Pearson correlation coefficient; IQR is the interquartile of the observed values calculated as the difference between the 75th and the 25th percentiles.

2.7. Spatial Modeling

We used the best performing model to predict SOM point to point to the 1 × 1 m raster of field A. The uncertainty of the prediction was quantified using a non-parametric bootstrapping approach [53]. By this, we obtained the probability distribution of 50 prediction realizations at each pixel. The final robust SOM map was determined by taking the average of all predictions. We used the bootstrap to determine the deterministic and random components of the model. To quantify the model uncertainty, we computed the 90% prediction interval. This prediction interval gives the range of values within which the true values fall with a probability of 90%. The prediction interval was identified by the 0.05 and 0.95 quantiles of the realizations. The spatial modeling was validated using field B. We predicted SOM for field B in the same way as field A. Subsequently, we extracted the SOM values from the final map and estimated the goodness-of-fit parameters of the validation using the 8 laboratory measurements.

3. Results

3.1. Descriptive Statistics of SOM

The overall SOM contents (n = 99) showed a range from 2.79–3.94% with a mean of 3.37 ± 0.23% (Figure 2). There was no significant difference between the SOM content of fields A (n = 91) and B (n = 8), with mean values of 3.37 ± 0.24% and 3.33 ± 0.20%, respectively. The range of SOM is relatively narrow, as indicated by the low coefficient of variation (CV). The distribution of SOM was close to a normal distribution so that no transformation of the input variable was necessary.

3.2. Correlations between SOM and Soil Color Covariates

To overcome the problem of highly collinear color variables in the RGB color space, we transformed the RGB values into several color space models and indices. The linear relationship between all color variables and SOM is shown in Figure 3. The correlation matrix helped to identify color features with a strong linear correlation among themselves and with SOM. Strong negative correlations between −0.67 and −0.68 were found between the SOM and RGB values, the X and Y values, the lightness variables L* and V, and the brightness index. The only variable with a strong positive correlation towards SOM was the redness index with r = 0.62. The chromatic variables a and C showed no linear relationship with SOM.

3.3. Machine Learning Model Performance

The model performance parameters of the PLSR, RF, and ANN models are shown in

Table 2. Ten-fold cross-validation results of the SOM prediction performance for the partial least square (PLSR), random forest (RF), and artificial neural network (ANN) models. Values given are the mean of 50 model realizations. Abbreviations: ME = mean error, RMSE = root mean square error, R² = coefficient of determination, CCC = Lin’s concordance correlation coefficient, and RPIQ = ratio of performance to interquartile distance.

	ME [%]	RMSE [%]	R2	CCC	RPIQ
PLSR	0.00	0.18	0.38	0.61	1.73
RF	0.00	0.13	0.68	0.80	2.42
ANN	0.00	0.17	0.47	0.62	1.85

. There were vast differences in the performance between the algorithms used. The RF model showed the best performance for all error metrics, with an RMSE of 0.13% and with an RPIQ of 2.42. The PLSR, as the simplest model, performed worst, but also the ANN with the highest model complexity could not outreach the RF model’s performance. The ensemble approach of the RF model could significantly improve the predictive accuracy of the PLSR model. The ANN model, however, might not be trained in its full complexity because of the limited number of samples. Based on those goodness-of-fit parameters, the RF model was used for all further spatial modeling.

The fine-tuning of the RF training showed that a mtry value of 5 was optimal. The importance of the color covariates in the RF model is shown in Figure 4. In this study, the most important variables were the same as the variables showing the highest linear correlation with SOM, which are the color covariates associated with soil brightness. This is no surprise, as it is accepted that soil brightness is a function of the organic carbon and organic matter content. It also indicates linear dependencies between the underlying variables.

3.4. SOM Mapping

The final RF model was used to predict SOM spatially at a 1 × 1 m resolution. Figure 5 presents the SOM point-support predictions. It also shows the uncertainty of the prediction, as the lower and upper limit of the 90% prediction interval and the prediction interval range. The SOM predictions showed a spatial pattern with the lowest SOM contents towards the dip and the slightly steeper parts of the field in the southern part of the field. This area is adjacent to areas with higher SOM contents in parts of the former green strip. The prediction uncertainty showed the same spatial patterns and a uniform spread across the field so that the model shows no apparent bias towards the upper or lower limits of the prediction. The prediction map has a marginally smaller range in SOM content with 2.91–3.78%, compared to the 91 soil samples from that field (2.79–3.94%). There is no difference in the mean SOM contents, however, with 3.38 and 3.37%, respectively.

As a validation of the RF model, we predicted SOM at the same resolution on field B (Figure 6). The uncertainty of the prediction showed the same spatial pattern as the SOM prediction map. Field B was overall more homogeneous than field A, which also showed a narrower SOM range. The range for the predicted map was 3.14–3.70% while the range for the validation samples was 3.06–3.58%. The mean for the prediction was 3.46% compared to 3.33% for the samples. However, it has to be considered that the samples from field B do not cover the complete field. Plotting the extracted SOM values from the prediction map against the laboratory measurements showed the model is slightly biased with 0.05%, but the error metrics with an RMSE of 0.12% and an RPIQ of 2.05 are comparable to the cross-validated training performance (Figure 7).

4. Discussion

4.1. UAV-Borne Soil Color Measurements for SOM Assessment

Until now, most agricultural studies have used UAV data mainly to estimate crop growth parameters or to detect weed problems [54]. However, our results showed that UAV-based measurements of soil color can be an effective way to predict the SOM content of bare cropland. Several soil color features showed medium to high negative correlations with SOM. With a maximum of −0.68, the correlation in this study was smaller than −0.79 found by Heil et al. [32] or −0.86 found by Viscarra Rossel et al. [31]. Those studies, however, correlated SOC to color information recorded in the laboratory. Gholizadeh et al. [38] compared the correlation of SOC and color features extracted from digital images taken in the laboratory and from satellite images. The authors found significant but weaker correlations when using satellite images. This can be expected, as lighting conditions are not standardized for remote sensing data, and atmospheric disturbances can weaken the SNR. In this study, UAVs recorded images via close-range sensing, yet there are potential sources of error introduced, e.g., via the color correction procedure. Therefore, we expected the correlation to be lower compared to laboratory results, although the correlations were still in an acceptable range compared to the studies above. While UAV-based measurements are sometimes considered as an intermediate form between proximal and remote sensing, some important interfering factors still need to be considered [2]. Such interferences can be caused by (i) the soil characteristics, such as surface roughness [17], soil crusts [55], and moisture content [42]; (ii) variations in hemispherical illumination characteristics because of temporal solar zenith angle changes, atmospheric scattering, or cloud shadows [56]. These factors need to be considered when collecting and processing data. In this study, the interference of illumination was minimized as images were recorded in one day under optimal conditions (clear sky, solar noon) while the soil conditions were less optimal: the soil surface was dry, but there were crop residues on the field and sugar beets were emerging. However, pixels including seedlings were successfully removed from the high-resolution images, which might lower correlations.

We showed that the UAV-based measured color information is in line with laboratory measurements, i.e., color covariates with high correlation towards SOM were the same among studies. It is widely accepted that lightness decreases as the SOM content increases [57,58], so it was no surprise that the color variables with the highest correlation with SOM were all related to the brightness of the soil. This relationship is described as linearly or curvilinearly [59]. Therefore, soil brightness variables, such as L and V, decrease with increasing SOM content. Looking at the R, G, and B bands revealed that all tristimulus values showed high correlations between SOM but also between each other, as all three bands are influenced by the illumination intensity. This is the major disadvantage of this color system according to Viscarra Rossel et al. [33]. Therefore, several authors recommend using color space models with a designated lightness value, such as CIELab [31,36,38,60,61]. Opposed to this, Aitkenhead et al. [62] argued in favor of the RGB system for its ease of use. Levin et al. [34] recommended the use of soil color indices. As these indices are ratios of RGB color bands, the authors expected them to reduce interference caused by soil surface and bidirectional reflectance distribution function (BRDF) effects.

In conclusion, UAV-based color measurements of bare soil are possible. However, there are distracting factors such as soil moisture, crop residues, and soil surface roughness that have to be considered and have to be investigated in more detail.

4.2. UAV-Based SOM Prediction for Digital Soil Mapping

The modeling results demonstrated that assessing the topsoil SOM content of bare cropland based on UAV data can be used for the fine-scale mapping of SOM. Both the internal and the external model validation provided acceptable predictions of SOM at the field scale. The internal validation on the test field gave a high prediction accuracy (RMSE = 0.13%, RPIQ = 2.42). However, the high accuracy came at a high cost of analyzing 91 soil samples on the 8.5 ha field for training the local model. The chemical analysis of the soil samples is time-consuming and expensive for farmers. Therefore, some authors propose the use of a regional model [2,63] or spiking approaches to improve the accuracy of regional models by adding a small subset of local samples [64]. By this, the need for soil sampling could be decreased to a few control samples and costs could be split between several users. Yet, regional models are prone to bias, as UAV-based data collection and processing is not standardized and recording conditions can vary substantially, and thus introduce noise. However, the external validation using an adjacent test field gave the same accuracy as the internal validation (RMSE = 0.12%, RPIQ = 2.05). This is promising, as a local model could potentially be used at the farm scale and for monitoring SOC stocks over time. So, the initial high cost of a local model could redeem value over time at least for larger farms. Nevertheless, the SOM prediction maps made for both fields give a good representation of the variation in SOM at the field scale, which is valuable information for field management and monitoring strategies.

Another problem when working on a field scale, especially in highly homogenized cropland, is that the range in SOM or SOC is often small. This makes it challenging to train a good-performing model with great accuracy at a field or farm scale [65]. In this study, the SOM variation was small compared to other studies [14,24], but some studies successfully predicted SOC even at a field scale [2,38]. Here, we could also successfully calibrate a model that reflected the variety of SOM in the soil sample data with an uncertainty range distributed relatively uniformly across the field. However, the range of the concentration within the sample set is a determinant factor in the accuracy of a model. Kuang et al. [65] found that a wider range of the calibration data resulted in higher R² and RPIQ values but also larger RMSE values. Thus, one could assume that a multi-farm calibration dataset could produce a calibration model with higher accuracy, albeit at a higher RMSE. Despite this, the successful prediction of SOM using UAV-derived color information showed that UAV-based field scale SOM monitoring is possible at a fine resolution, even at small variations in SOM.

As the calibration of a field-scale model is challenging, we tested three different machine learning algorithms: a PLSR, an RF, and an ANN. The RF model outperformed the other two algorithms. RF is popular for soil predictions as it is suitable for datasets with few samples and many predictors [49]. RF is also robust to noise by fully using all input data and can handle irrelevant features such as the redundant color information in this dataset; thus, it was best suited for this UAV-derived data [27]. On the other end, PLSR performed worst. It implements data compressing to handle many predictor variables in highly collinear data, but as a linear model, it could potentially not learn the complex structure in the data as the other non-linear algorithms. Therefore, PLSR is still commonly used as a reference in modeling studies but is gradually being replaced by algorithms such as RF and ANN [25]. Those non-parametric and non-linear models can learn complex patterns in the data but are best when using many samples and features [49]. With small sample sizes, ANNs are prone to overfitting as ANNs’ structure aims to minimize the error in the training data [66]. Due to this weakness of ANNs, studies with smaller datasets often find RF that outperforms ANN, as ANNs can only showcase their full potential with very large datasets [67,68,69].

4.3. Potential of SOM Mapping in Agriculture

Few studies have examined the assessment of SOC or SOM using multispectral imagining sensors on UAV platforms [14,24]. This study confirmed the feasibility of UAV-based SOM mapping at a fine scale and corroborates the potential of this technology. Other studies used satellite imagery (mostly Sentinel-2) to predict SOC in cropland [5,21,22,70]. With the latest progress in satellite sensors, the spatial and spectral resolution of today’s multispectral satellite systems allows the quantitative assessment of soils [5]. The current generation of satellites is equipped with multispectral sensors that acquire images up to the SWIR region and the newest systems such as the Environmental Mapping and Analysis Program (EnMap) even reach into hyperspectral territory, expanding the spectral resolution considerably [71]. At a spatial resolution of 10 m, Sentinel-2 could be sufficient to detect spatial variability in SOM at a regional scale; yet, it is perhaps not enough to detect sudden changes in SOM at short distances [5]. Here, low-altitude UAVs provide more detailed information on field-scale variability, such as showing differences in SOM on field A at the depression or the area of the former green strip (Figure 5). Although a spatial resolution of up to a few centimeters is technically possible, it is practically not feasible. In this study, we decided to downsample the images to a 1 m resolution. This downscaling has the side effect that it removes noise from the image by smoothing out short-range artifacts, e.g., caused by BRDF effects due to surface roughness as discussed above [36]. Dornik et al. [72] found that in DSM, the optimal scale of predictor variables improved the accuracy of prediction models. Thus, resampling helped to improve the prediction accuracy at the cost of a centimeter-scale resolution. In a further comparison, the greatest limitation of UAVs is the spatial coverage that can be obtained by UAVs [24]. In addition, the amount of data produced demands a substantial amount of processing time. However, UAVs are more flexible in terms of acquisition time. Although many satellites have a short revisit time and data are freely available, acquisition dates are fixed, and cloud cover significantly reduces the number of available images. This is possibly the strongest advantage of UAVs over spaceborne sensors [14].

Considering all limitations, UAVs show a high potential for a rapid and accurate field-scale SOC or SOM prediction. Such a high-resolution method is very useful for the assessment of SOC stocks and their long-term monitoring. The future of UAVs is promising as a low-cost and versatile tool for DSM situated somewhere between remote and proximal soil sensing. Technology advances quickly and methodological technical challenges are faced in current studies so the use of drones in agricultural and soil science will presumably increase significantly. Soils and especially SOC vary across small scales that space-borne sensors cannot cover. This offers a large field of application that could, e.g., help to evaluate carbon sequestration efforts. Economically, UAVs become accessible to farmers and the initial high costs could pay off [24]. Here, we showed that no multi- or hyperspectral camera is necessary to produce accurate SOM maps and RGB color measurements are enough. This could further reduce the initial cost of a UAV system and make it even more accessible for farmers.

5. Conclusions

We investigated the capability of UAV-based bare soil images to predict and map the SOM content of the topsoil in cropland based on soil color information using an RF model and ground truth soil samples for model calibration. We trained the RF model on a field in the German loess belt and used an adjacent field to validate the model. The prediction accuracy was in the range of similar studies, albeit a small range of SOM within the field. However, the UAV-based color measurements correlated less with the SOM content compared to laboratory studies. This was expected because of atmospheric influences and BRDF effects. Despite this, measurements were adequate to produce SOM maps at a 1 m spatial resolution, exposing the within field organic matter variability. Furthermore, the validation using external data showed that one can transfer the prediction model to neighboring fields, thus permitting prediction on larger-scale farms or enabling carbon monitoring over time. With this study, we demonstrated a rapid, low-cost, but accurate method for mapping organic carbon or organic matter in agricultural soils. Except for the initial costs of a UAV system and the extensive soil sampling for the training dataset, farmers could produce accurate and fine-scale SOM maps at the cost of few validation samples. The area covered by a UAV flight and the level of detail that can be achieved make this technique ideally suited for precision agriculture applications. Here, we downscaled the resolution of the images to improve the SNR but submeter resolutions are achievable.

Based on these promising results, more research will need to address the atmospheric corrections applied to UAV imagery and investigate the influences of soil roughness, soil moisture, and crop residues on predictions. This also includes considerations about the transferability of models to different regions, which may necessitate the use of further environmental covariates to increase the prediction accuracy. The development of new multi- to even hyperspectral sensors is advancing rapidly. This will lead to higher spatial and spectral resolutions, which could help to characterize soil more accurately. However, in addition to all technological advances, there is an urge for standardization of the acquisition and the processing of UAV-borne imagery to make prediction models more comparable and transferable.