The decreasing soil organic carbon (SOC) content in agriculture soils is generally considered a major threat to the sustainability of soil cultivation. Its role is essential in many production and non-production soil functions as it controls the dynamics of various agrochemical processes in the soil. The natural equilibrium of the soil environment is endangered due to external, primarily anthropogenic effects, which lead to the development of several degradation processes. These can also affect the soil carbon stocks, especially in the topsoil layer. Soil is a vast carbon pool (the largest terrestrial) [1
], making it an essential component of the entire carbon cycle on Earth, especially in the context of expected climate and land-use changes [5
]. Therefore, recent research on SOC has received considerable attention. Monitoring, mapping, and describing the spatial variability of SOC (in landscape and also within-field scale) are the key prerequisites for understanding the effects of agricultural practices on SOC changes. Digital soil mapping methods are used to obtain this mapped variability using field sampling and additional environmental covariates [9
]. Remote sensing (RS) data represents one of the available data sources for such purposes within large areas. The data are also provided in a sufficient spatial resolution suitable even for local monitoring and applications [10
A number of studies [11
] have proven that aerial image hyperspectral data with many narrow spectral bands in VNIR-SWIR offer efficient input data to map the spatial variability of important soil properties. Compared to the costly and technically demanding processing of aerial hyperspectral data, multispectral and superspectral satellite data [23
], or multispectral and hyperspectral UAS data [29
] could be a readily available source of spectral data for regular application. Despite the increasing number of respective studies, the potential of this type of data has not been fully exploited. An important point to be addressed is the effect of different spectral and spatial resolutions on prediction ability. It is assumed that reduced spectral resolution, as in the case of multispectral data, results in a reduction in the model’s predictive capability. Despite this, a number of studies dealing with multispectral data [32
] have shown that results can be satisfactorily applied to the needs of precision farming, especially with regard to acquisition costs.
Nevertheless, potential inaccuracies in the outputs resulting from the use of remote RS data with lower spectral resolution and, in the case of satellite sensors, often with coarser spatial resolution, need to be considered. For example, sensors on Sentinel-2 and Landsat-8 satellites covering important organic matter absorption bands in both the visible and SWIR regions of the spectrum may have considerable potential for detailed mapping of SOC. However, only a few studies have confirmed this potential [35
]. Sensors with very high spatial resolution but only covering the VNIR spectrum (e.g., WorldView, Cartosat, Pléiades, and Deimos) are considered less applicable. However, as shown by Crucil et al. [30
] in a study comparing UAS-compatible multispectral and hyperspectral sensors operating only in the VNIR spectrum, similar results can be achieved with these sensors compared to reference hyperspectral data also using the SWIR spectrum. Unlike spectral resolution, spatial resolution is considered to have less of an effect on predictions relative to continuously changing soil properties. For example, Castaldi et al. [36
] showed that the spatial resolution of Sentinel-2 is adequate for SOC variability mapping both within the field and at a regional scale.
The signal-to-noise ratio (SNR) of sensors is another important issue that affects the prediction ability. Large noise interference in the acquired data associated with the short acquisition time at the investigated location is a disadvantage mainly for satellite data [38
], especially when data are scanned in narrow spectral bands and with high spatial resolution. For example, SOC prediction using the Hyperion hyperspectral satellite sensor [41
] may be affected by this phenomenon. A comparison of data from Hyperion and Advanced Land Imager (ALI) sensors [37
] showed that sensors with lower spectral resolution but higher SNR can provide better results for SOC prediction.
It follows from the above mentioned that there are still a number of uncertainties and unanswered questions about using the mentioned approaches in management practices. One reason is the difficulty of comparability and hence the possibility of evaluating results from individual studies using different sensors, preprocessing of spectral data, or statistical and numerical data processing techniques. Moreover, accuracy and prediction ability are often affected by other factors, such as different soil conditions, variability of the analyzed characteristics, the condition of the studied surface (moisture and surface roughness affecting vegetation and crop residues), various atmospheric conditions and light incidence geometry during image acquisition [48
]. This leads to reduced predictive ability compared to that obtained with soil laboratory spectroscopy [33
] and makes it difficult to map SOC at a large scale, especially in temperate regions, due to crop cover and various types of land parcel management.
There is only a portion of bare soil with a dry and non-rough surface in each RS image. Time series of individual images [51
] or multitemporal composites of spectral data can be used to reduce the influence of different surface conditions and eliminate vegetation. Exposed Soil Composite Mapping Processor (SCMaP) [55
], Geospatial Soil Sensing System (GEOS3) [56
], Bare Soil Composite Image [57
], and Barest Pixel Composite for Agricultural Areas [58
], all developed from Landsat time series, multitemporal bare soil image [59
] developed from RapidEye time series, or bare soil mosaic [60
] derived from Sentinel-2 data can serve as examples of such composites. However, only some of the composite products have been used to predict SOC [57
]. Promising results were achieved; however, the potential of these spectral composites has not yet been tested in a relevant number of studies, and further research is needed for its evaluation.
The objective of this study is to critically evaluate the capability of easily accessible data (and one commercially available source with very high spatial resolution) from different types of multispectral sensors to predict within-field variability of topsoil SOC concentration at a local scale. Real spectral image data, identical sampling and processing design, and similar surface conditions (dry conditions and minimal surface roughness) were ensured to achieve this goal. The data from currently operating sensors, including satellite data from Sentinel-2 and Landsat-8 with VNIR and SWIR bands, very-high-resolution data from CubeSat miniature Dove satellites from PlanetScope (VNIR), and data from the UAS-mounted Parrot Sequoia sensor (VNIR) were compared. Multitemporal bare soil composite of Sentinel-2 spectral data was also tested to evaluate the usability of this regional product for regular usage in local mapping. Mapping results from airborne hyperspectral data also used in preliminary studies [14
] were used as reference data for evaluating the spatial concordance among resulting maps and to analyze the importance of different spectral bands for SOC mapping. Although more datasets with wider variety of spectral and spatial parameters would be needed for a robust analysis and statistical testing, the study attempts to compare the SOC prediction models using real-world data from different sensors to evaluate the influence of spectral and spatial resolution and SNR on prediction accuracy. The hypothesis is that lower spectral and spatial resolution and SNR of image spectral data will lead to lower prediction accuracy.
The results showed a potential of different types of multispectral data for mapping SOC on a local scale and even for regional mapping using composite datasets. On the one hand, the predictive capability of models achieved poor or average results based on RPD evaluation (RPD from 1.16 to 1.65). This means that the error of these models is similar or only slightly lower than the standard deviation of the SOC samples. On the other hand, despite these rather inconclusive results, other findings (especially spatial correlation analysis) indicate a high correlation between the reference results obtained from the hyperspectral data and other maps, especially those derived from Sentinel-2 and PlanetScope (0.75 and 0.82). This was shown in the comparison of map outcomes. Thus, the final maps produced on the basis of multispectral data can, despite the low model accuracy metrics, precisely reflect the within-field variability of SOCs. The results (RPIQ values, scatterplots of predictions and observations, and maps) also show that the models predicted worse the values in the lower and upper tails of the distribution. This affects the model accuracy metrics. In this respect, it would be appropriate to address the issue of outliers in future research. It could help to increase the accuracy of the prediction in tailed values, which is generally a problem with machine learning methods. The prediction accuracy of models can also be improved by incorporating other environmental covariates (terrain, parent material maps, etc.) or incorporating covariate contextual information into the prediction models [88
The achieved results illustrate that multispectral data provide significantly worse SOC estimations than reference hyperspectral data regardless of the spatial resolution. This is due to the combination of higher SNR and spectral resolution. However, the availability of the hyperspectral data is, due to a lack of hyperspectral satellites in orbit, generally worse. Other drawbacks of the hyperspectral data are a high acquisition cost of aerial data, high demands on hardware, and know-how in the data processing. On the other side, the presented multispectral RS missions, especially Sentinel-2 and Landsat-8, provide large amounts of freely available data that can be suitable for SOC digital mapping. The results of our study show very similar prediction accuracy for all spaceborne sensors with only minor prediction variance, which could not be explained without a full factorial experiment design and consequent statistical testing of all variables. More data from different sensors would be needed for a robust analysis. Despite the limited number of sensors, interesting conclusions have been drawn.
Satellite multispectral sensors provided data only from a few broad spectral bands. This is the difference from the hyperspectral sensors which provide continuous reflectance curves in VNIR-SWIR spectra with high SNR and include more absorption features related to SOC [94
]. This allows for better results and higher accuracy of SOC prediction. Similar results were reported by Castaldi et al. [42
], who compared SOC estimation by the PLSR model using image data from the Advanced Land Imager (ALI) and Hyperion sensors on board the EO-1 satellite. Hyperion data provided better results than multispectral ALI data for clay, sand, and especially for SOC estimation. Cascaldi et al. [37
] estimated SOC and other soil properties using simulated data from soil spectral libraries and data from seven hyperspectral and multispectral sensors. Sentinel-2 MSI data showed prediction accuracy equal to simulated Hyperion data, which had very low SNR in the SWIR spectrum, but the Sentinel-2 data had significantly better results in terms of prediction accuracy (RPD = 1.55; RPIQ = 2.68) than Landsat-8 (RPD = 1.46; RPIQ = 2.51). The best results were achieved with EnMAP (RPD = 1.8; RPIQ = 3.11). According to their results, this was due to more bands in the SWIR region combined with narrower bands, which better reflect the spectral features of organic matter. Rosero-Vlasova et al. [96
] obtained similar results also using simulated satellite data. They achieved the best fit with models using simulated EnMAP reflectance (R2
= 0.93). The least reliable estimates (R2
= 0.4) came from the simulated Landsat model, while the Sentinel-2 model showed better performance (R2
= 0.63). In our study, we obtained slightly better results using non-simulated real satellite data from Landsat-8 (R2
= 0.65, RMSE = 0.28%) and Sentinel-2 (R2
= 0.68, RMSE = 0.26%). Moreover, the prediction accuracy of the Sentinel-2 model was slightly better than that of the Landsat-8 model, which is consistent with the aforementioned studies [37
Spectral absorption regions, which can be used to quantify soil organic carbon (SOC), are located mainly in broader bands in the visible region of the spectrum and in the narrower bands of the SWIR spectrum (between 1600 and 1900 nm and around 2100 and 2300 nm) [36
]. For this reason, the spectral resolution of the sensors significantly influences the quality of SOC predictions [34
]. It is, therefore, necessary to use data with appropriate spectral resolution taken across the VNIR-SWIR spectrum for accurate SOC estimates [101
]. This is also shown by the results of our study, where the importance of bands in individual prediction model was investigated. Red and NIR bands are the most important in the multispectral data use (except of PlanetSCOPE data). This suggests that not only the presence of spectral bands but also their constellation is very important. SWIR bands for Sentinel-2 and Landsat-8 were also very important prediction bands, especially SWIR 1 band around 1600 nm. Presence of these bands can be a significant advantage over the data that uses only bands in the visible and NIR spectrum. New-generation satellite hyperspectral sensors (e.g., EnMAP, PRISMA, HyspIRI, SHALOM) with relatively high SNR and high spatial resolution can make progress in this regard. Recent studies have demonstrated their potential for SOC prediction based on simulated data from point hyperspectral measurements [12
Unlike spectral resolution, the effect of spatial resolution is not so obvious. It can be assumed that higher spatial resolution can lead to slightly higher prediction accuracy, if other parameters of the sensors are identical. This was confirmed by the results of the study when upscaling of data led to a decrease in predictive ability—Sentinel-2 data from 20 m (RMSE = 0.26%) to 30 m (RMSE = 0.28%), PlanetScope data from 3 m (RMSE = 0.26%) to 30 m (RMSE = 0.27%), and Parrot Sequoia data from 1 m (RMSE = 0.31%) to 10 m (RMSE = 0.34%). The same trend was achieved by the model using the raw hyperspectral dataset with original spatial resolution of 3 m (RMSE = 0.20%). The hyperspectral datasets rescaled to 10 and 30 m resolution showed a decrease in the prediction accuracy (RMSE = 0.24%). These minor decreases in prediction accuracy could be caused by decreasing spatial resolution of the image, because the spectral resolution remains constant. However, we could not perform statistical testing of the decreasing prediction accuracy trend due to the lack of multiple instances of each rescaled model and its validation metrics. Steinberg et al. [12
] similarly investigated the influence of spatial resolution to SOC prediction by PLSR, comparing simulated spaceborne hyperspectral EnMAP and Airborne hyperspectral system (AHS) images with higher spatial resolution. Their results showed that EnMAP allowed prediction of iron oxide, clay, and SOC with an R2
between 0.53 and 0.67 compared to AHS imagery with an R2
between 0.64 and 0.74.
A great potential for local applications is linked to UAS spectral sensors, as shown by Crucil et al. [30
]. The main advantages of UAS data are low acquisition cost, high spatial resolution, and flying on demand. However, the prediction accuracy of the models using images from the Parrot Sequoia UAS camera (R2
= 0.72, RMSE = 0.31%) was lower compared to the spaceborne sensors. It should be noted that these sensors are much cheaper and built with inexpensive electronics parts, resulting in significantly lower SNR, which is not comparable with the SNR of agency-funded satellites [102
]. Although we do not have enough data to test this hypothesis, it can be assumed that the lower accuracy of Sequoia data is partially influenced by lower SNR. SNR has a proven effect on prediction accuracy [11
]. Gomez et al. [39
] concluded that the lower accuracy of SOC estimations using Hyperion spectra is because of lower SNR (~50:1) and spatial resolution compared to Agrispec field spectrometer data resampled to similar spectral resolution.
In this study, predictive ability was also evaluated using a time composite from Sentinel-2 data. The prediction of soil properties using RS data requires the presence of bare soil in the images. Thus, it is necessary to select images without the masking effect of vegetation. Mapping of larger areas accordingly is rather complicated, especially in temperate regions with different crop rotations throughout the year. The use of time composites is one of the few proven alternatives. The main challenges in composite development are cloud masking (and cloud shadows), definition of bare soils (vegetation masking, including non-photosynthetic vegetation, straw, and litter) on individual images and determination of the resulting reflectance values. Different approaches were used to derive the reflectance values of the final product in the previous studies. Thresholds of spectral indices are usually used for the vegetation masking: Mainly NDVI for green vegetation [56
], NBR2 [56
], or MID-infrared [57
] for non-photosynthetic vegetation or combined indexes, such as Bare soil index (BSI) [58
] or PV [55
]. Statistics from time series of masked data were used to derive final reflectance data—mean [55
], median [56
], or minimum [57
]. Other methods used to improve and obtain more stable values included, for example, exclusion of 5% quartile [58
], application of PCA components [54
], field-based standard deviation values [59
] or using low-pass filter [60
]. In this study, we used Sentinel-2 time series, PV index for masking vegetation and mean statistic for deriving reflectance data and Sentinel-2 composite. The prediction using this composite achieved only average results. However, these results were better than those achieved in a study by Diek et al. [58
], which was conducted at the national level using Landsat data, and comparable with results of Blasch et al. [59
] and Gallo et al. [57
]. The potential of these composite data is great; it can be used as an entry in DSM models. However, further development is necessary. It can be assumed that the weak predictive ability of these data is due to various factors. Above all, this product combines data that has been taken under different moisture conditions and surface roughness. Despite a progress in this area, it is necessary to develop new algorithms, not only for identifying bare soils, but also for removing the influence of moisture, surface roughness, or vegetation residues. Clouding and shading effects on input data is another aspect. Clouds and shadows can be masked using already developed algorithms, but the results are still not faultless. Therefore, further development is needed in this respect.
We must take into account that the results can be affected by the mismatch between the size of the sampling spot (composite sample in the area of 1 m2
) and the resolution of the sensor (1–30 m2
). However, evaluating this effect would require a very challenging experiment going beyond the scope of this research. Thus, further studies are needed to evaluate this effect. Another issue that could negatively affect the results of multispectral models in this study is the sampling design. As mentioned for example by Castaldi et al. [11
], one sampling for each remote sensing acquisition is required for a precise prediction. SOC concentration can have strong dynamics in both space and time, and differences in sampling and acquisition time can negatively affect the results. Although organic fertilizers were not applied and no significant erosion events occurred between the time of sampling and data acquisition, it is necessary to take into account the possible influence on the results. Unfortunately, this cannot be achieved without a time-differentiated and time-consuming and costly sampling. According to our best knowledge, any study assessing the effect of sampling and acquisition time on the SOC prediction has not been conducted yet, but is urgently needed.