STEROPES D7.1 - D7.3. Satellite-based (Sentinel-2) soil sampling strategies for SOC estimation at field- and farm-scale

Deliverable 7.3 1 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Towards climate-smart sustainable management of agricultural soils STEROPES WP7: additional activities regarding spectral models and soil sampling strategy T7.3: SATELLITE-BASED SAMPLING STRATEGIES D 7.1 and D7.3 Satellite-based (Sentinel-2) soil sampling strategies for SOC estimation at fieldand farm-scale Due date of deliverable: M54 (D7.3), M56 (D7.1) Actual submission date: 31.07.2024 Deliverable 7.3 2 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Deliverable 7.3 3 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 GENERAL DATA Grant Agreement: 862695 Project acronym: EJP SOIL Project title: Towards climate-smart sustainable management of agricultural soils Project website: www.ejpsoil.eu Start date of the project: February 1st, 2021 Project duration: 60 months Name of lead contractor: INRAE Funding source: H2020-SFS-2018-2020 / H2020-SFS-2019-1 Type of action: European Joint Project COFUND DELIVERABLE NUMBER: D7.1 - D7.3 DELIVERABLE TITLE: Satellite-based (Sentinel-2) soil sampling strategies for SOC estimation at fieldand farm-scale DELIVERABLE TYPE: Report AUTHORS: Adriaan Vanderhasselt (ILVO), Fabio Castaldi (CNR), Mehdi Saberioon (ILVO), Bert Callens (ILVO), Nick Berckvens (ILVO), Panos Ilias (ILVO), Paul Quataert (ILVO), Greet Ruysschaert (ILVO), Aldo Dal Prà (CNR) Deliverable 7.3 4 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Table of Contents 1 Introduction ................................................................................................................................................ 5 2 Research questions ................................................................................................................................... 12 3 Determination of a cost-effective plot-scale sample size ......................................................................... 13 3.1 Precision .......................................................................................................................................... 13 3.2 Comparison of data on plot-scale precision from scientific literature to experimental Flemish conditions ..................................................................................................................................................... 19 3.2.1 Experiment ................................................................................................................................. 19 3.2.2 Results......................................................................................................................................... 20 3.2.3 Costs (ILVO test case – Flanders) ................................................................................................ 21 4 Satellite-based sampling strategies .......................................................................................................... 24 4.1 Materials and Methods ................................................................................................................... 24 4.1.1 Test sites ..................................................................................................................................... 24 4.1.2 Sentinel-2 composite layer – Mosaicking ................................................................................... 32 4.1.3 Soil sampling strategies .............................................................................................................. 33 4.1.4 Sampling, processing and analysis protocol ............................................................................... 40 4.1.5 Local (farm-scale) calibration of regional SOC map.................................................................... 42 4.2 Results ............................................................................................................................................. 43 4.2.1 Spectral comparison ................................................................................................................... 43 4.2.2 SOC content comparison ............................................................................................................ 45 4.2.3 Local calibration .......................................................................................................................... 55 5 References ................................................................................................................................................ 61 Deliverable 7.3 5 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 1 Introduction The sense of urgency that emanated from the 2015 United Nations Climate Change Conference in Paris (COP 21) has spurred on the search for effective climate change mitigation measures in the past decade. Farming practices have received a lot of attention in this regard, because agricultural soils have seen a massive loss in their soil organic carbon (SOC) stocks ever since the onset of agriculture (Sanderman et al., 2017) and therefore possess a large potential for restoration (Lal et al., 2018; Minasny et al., 2017). Various farming practices could help with this restoration, (Batjes, 2019; Bossio et al., 2020; Paustian et al., 2016; Spencer et al., 2011) and this is driving an upsurge in new ‘carbon farming‘ initiatives, both from the public and private sector, that aim to incentivize farmers to adopt these practices (Amelung et al., 2020; Black et al., 2020; Commission, 2021; Elkerbout et al., 2020; Guan et al., 2023). However, the lack of robust and cost-effective monitoring systems has proven a major barrier to the upscaling of these carbon farming initiatives (Commission, 2021; Oldfield et al., 2021; Smith et al., 2020). With the (currently still provisional) Carbon Removal and Carbon Farming (CRCF) Regulation, and relatedly the Soil Monitoring Law, the European Commission aims to tackle this issue and provide a statistically sound and cost-effective SOC monitoring or estimation framework (Commission, 2021; Günther et al., 2024; Kotschik et al., 2024). Most existing SOC estimation methods make use of direct measurement through soil sampling and laboratory analyses (FAO, 2019; Smith et al., 2020). These methods require the development of a sampling strategy, which entails making decisions on sample size, sample pattern and sample locations (Brus & De Gruijter, 1997). It is possible to achieve very high accuracy with direct measurement, however, there are various complicating factors that can put a serious constraint on the cost-effectiveness of these estimation methods. All steps of direct measurement (i.e. sampling, processing and analysis) have an intrinsically high cost (Batjes et al., 2023; Smith et al., 2020). Furthermore, the large background carbon stocks, the large spatial variability (even at smaller operational management units, like field or farm scale) and the inherently slow carbon gains all make it more difficult to achieve the, often very high, desired accuracy with a limited budget, and thus sample size (Biswas & Zhang, 2018; Garten & Wullschleger, 1999; Goidts et al., 2009; Post et al., 2001; Smith et al., 2020; Vanguelova et al., 2016). An alternative for direct measurement could be found with spectral methods that rely on the content-specific absorbance/reflectance of light on the soil surface in the visible and infrared (Vis-NIR: 400-1100 nm) and shortwave infrared (SWIR: 1100-2500 nm) regions of the electromagnetic spectrum (Ben-Dor et al., 2009; Chabrillat et al., 2019; Smith et al., 2020). These methods make use of prediction models, that are based on spectral libraries (comparison of traditional laboratory and reflectance measurements), to estimate the SOC content (Smith et al., 2020). Since these estimations don’t require any expensive and time-consuming Deliverable 7.3 6 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 standard laboratory analyses once the spectral library is created, the spectral measurements are very useful for collecting high resolution information on SOC content and thus also for mapping at many different scales (Chabrillat et al., 2019). Larger scale estimation and mapping has been greatly facilitated by the diverse set of platforms, like UAVs or satellites, that have been equipped with spectral sensors in recent years (BellonMaurel & McBratney, 2011; England & Rossel, 2018; Nayak et al., 2019). Several studies have been able to use satellite-derived spectral data for SOC content estimation (Köchy et al., 2015; Poggio et al., 2021; Vaudour et al., 2022). However, these estimations are often still of relatively low accuracy, especially at smaller operational management scales (i.e. field or farm), and they are therefore seldom used for direct carbon monitoring in carbon farming initiatives (Andries et al., 2021). These low accuracy levels can be attributed to a broad set of potential disrupting factors, such as soil moisture content, texture, salinity, surface roughness, dry or green vegetation cover and weather conditions (e.g. cloud cover, sunshine variation) (Castaldi et al., 2023; Dvorakova et al., 2021; Tziolas et al., 2021; Vaudour et al., 2022). Despite its relative inaccuracy as a direct estimation method, satellite-derived spectral data could still be a serious boon to SOC monitoring/estimation by providing a priori knowledge of the site-specific spatial variability for the development of optimized sampling strategies (Shao et al., 2021). Any sampling strategy that seeks to accurately estimate SOC content will need to take the entire range of SOC values within the zone of interest into account (Brus & Heuvelink, 2007). Sampling strategies that are based on the geographical space, like grid sampling, can offer a substantial improvement in cost-efficiency compared to simple random sampling and further improvements can be made by exploiting the relationship between SOC content and potential covariates (also referred to as ancillary data), which entails an optimization in the feature space (Brus & Heuvelink, 2007; Hengl et al., 2003; Minasny & McBratney, 2006). Multiple studies have used satellitederived spectral data in this way as covariate, both at field (e.g. Guan et al., 2023; Potash et al., 2022; Singh et al., 2012; van der Voort et al., 2023) and farm-scale (e.g. Bettigole et al., 2023; Castaldi, Hueni, et al., 2019; de Gruijter et al., 2016; Deluz et al., 2020; Malone et al., 2018). However, the strong site-specific nature of the relationship between SOC content and covariates, like the soil’s spectral patterns, still make it difficult to develop robust, scalable and easy to use methods for the optimization of sampling strategies (Bettigole et al., 2023; Kerry & Escola, 2021; Smith et al., 2020). Selection algorithms, like Kennard-Stone (e.g. Castaldi et al., 2019), Ospats (e.g. de Gruijter et al., 2016) and conditioned Latin hypercube sampling (e.g. van der Voort et al., 2023), have been successfully utilized to optimize sampling strategies for SOC content estimation based on satellite-derived spectral data. However, these algorithms do come with a certain requirement for computing power and programming skills. Many SOC monitoring initiatives therefore still rely on simpler random or regular grid sampling strategies (Bettigole et al., 2023; Smith et al., 2020). Deliverable 7.3 7 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 This study is carried out within the framework of the STEROPES project of the European Joint H2020 Program SOIL (https://ejpsoil.eu/soil-research/steropes/https://ejpsoil.eu/soil-research/steropes/). The overarching goal of this study is to develop cost-effective satellite-based sampling strategies for the determination of topsoil SOC content in cropland at fieldand farm-scale. In Flanders (Belgium), the usefulness of a regional SOC map, that was recently developed by the Horizon 2020 project of ENVISION (see Text box 1), as source of ancillary data to optimize soil sampling at farm-scale will be investigated (map-based sampling strategy). However, since SOC maps are not publicly available for all regions of Europe, we will also look at the direct use of the Sentinel-2 spectral bands with the Kennard-Stone selection algorithm to help guide sampling strategies at field-scale in Italy (satellite-based sampling strategy). During the development of optimized sampling strategies, extra care will be taken to ensure that the new approaches are robust and easy to use. The complexity and need of extended computing power and programming skills of many existing optimized sampling strategies still forms a major barrier to their implementation by carbon monitoring initiatives (Bettigole et al., 2023; Smith et al., 2020). As a second step, we will compare the results from the newly developed farm-scale sampling strategies in Flanders with the estimates from the regional ENVISION SOC map to calibrate the ENVISION SOC map accordingly, to the site-specific farm-scale conditions (ijken – see for example Veris scan (van der Schans & van der Berg, 2013)). We will subsequently assess if the adapted ENVISION SOC map manages to estimate the topsoil SOC content for every field of the farm of interest at enhanced accuracy, while keeping the required sample size to a minimum. To be able to link the soil sampling results with the Sentinel-2 data, we decided to keep the smallest resolution pixels (10mx10m) of these satellites as sampling support (i.e. plot) for our newly developed sampling strategies. The initial calibration sampling campaign for ENVISION took 16 subsamples in a Sentinel-2 pixel (i.e. plot) of 10mx10m, that were later bulked into one composite sample for lab analysis. This plot-scale sample size was calculated during the preparatory work for the Flemish Cmon monitoring network based on regional data on field-scale SOC variability (Lettens et al., 2021; Sleutel et al., 2021). Such an expanded sample size was deemed appropriate for the forest and grasslands soils that were included in the Cmon monitoring network, however, it seems highly likely that a smaller sample size could be used for just croplands with only limited impact on the precision of the SOC content estimation, while helping to reduce the overall costs. Text box 1: ENVISION SOC map for Flanders ENVISION is the acronym of the Horizon 2020 project “Monitoring of Environmental Practices for Sustainable Agriculture Supported by Earth Observation” (2020-2024) that focuses on offering tools for the continuous, large scale and uninterrupted monitoring of farm management activities with regards to Deliverable 7.3 8 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 sustainability (Tsardanidis et al., 2024). Five distinct business cases were developed within this project's framework that could use specific forms of earth observation data. Here our interest is limited to the business case developed by ILVO, on the monitoring of cropland topsoil SOC in Flanders by utilizing Sentinel-2 data. This source of satellite imagery from ESA (Sentinel missions) provides freely accessible, high-resolution data with both spatial and temporal variability. Sentinel-2 has a spatial resolution of 10 m for most bands and a revisit time of 5 days. The methodology that was followed to get to the desired cropland SOC maps can be divided into five phases, as is depicted in Figure 1. In Phase One, the goal was to develop a cloudless collection of bare soil pixels. As a first step, it was required to distinguish between cropland and the other land-uses. This could be done either by way of the European Space Agency (ESA) WorldCover 10 m 2020 product, which provides a global land cover map for 2020 at 10 m resolution based on Sentinel-1 and Sentinel-2 data, or with the Land Parcel Identification System (LPIS) information provided by the Flemish government. Besides that, disrupting factors, like cloud cover, soil moisture content and vegetation or plant residue cover, needed to be considered and were filtered out with a combination of thresholds on relevant spectral indices. Two, the modelling was performed in order to create a mapping between the extracted reflection signatures from the Bare Soil Collection and actual SOC measurements. Originally, the open-source, low-code machine learning library of PyCaret was used to perform data preparation, model training, hyperparameter tuning, analysis and interpretability, and model selection, but at a later stage (April 2024), this process was repeated with the TensorFlow machine learning platform. PyCaret tested a host of different machine learning algorithms in the PyCaret framework on Google Colab, to link Sentinel data from beginning 2018 to May 2022 and the Extra Trees Regression with standard options for PyCaret version 2 was chosen based on global statistics to generate the map, and some extra options such as sensitivity analysis, variable importance with Shapley values were integrated in the PyCaret network and applied. The TensorFlow approach was required to work in a new workflow which permitted API integration on the Google Cloud platform, and the TensorFlow map worked with more recent data from Sentinel-2, including data from 2022, 2023 until Januari 2024. The model itself was a basic ANN TensorFlow model. In the rest of the report the two distinct SOC maps will be referenced by their respective machine learning platform. The SOC measurements were carried out for 171 different sampling locations spread out over the different soil regions of Flanders (Tsardanidis et al., 2024). The selection of the exact sampling locations was carried out according to a stratified feature-based (i.e. 11 soil association regions based on the Soil Association Map of Flanders) approach with the Kennard-Stone algorithm (Kennard & Stone, 1969). The intent of this approach was to cover most SOC variability in croplands in Flanders. The sampling areas corresponded with Deliverable 7.3 9 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 one Sentinel-2 pixel of 10mx10m dimensions, and the middle was determined with RTK GPS precision. As is presented in Figure 2, within this area 16 randomly selected subsamples were collected into one composite sample to deal with the small-scale spatial variability. In Phase Three, the deployed model was applied to all pixels belonging to the larger area of interest (i.e. all Flemish cropland). In Phase Four, a technical validation of the model was performed for each soil association region separately, to get a better understanding of the accuracy of the deployed models. Figure 3 presents the results of the regional-scale validation for the ENVISION PyCaret SOC map. Phase Five consists of a broader evaluation and the making of improvements. Figure 1: Methodological phases followed in task 3.6 of the ENVISION project (Tsardanidis et al., 2024) Figure 2: SOC sampling area, corresponding with one Sentinel-2 pixel (10mx10m). The orange X’s depict the 16 randomly selected subsample locations. Deliverable 7.3 16 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 6 presents the precision that could be achieved for the plots and fields presented in Table 1 by varying the sample size. Precision is here represented by the (95%) margin of error (MOE) which equals half the length of the 95 % confidence interval. A 95% confidence interval around the sample mean is as follows (van Belle, 2008): 𝑌 ± 2 𝑠 √𝑛 ⁄= 𝑌 ± 𝑀𝑂𝐸 with 𝑌  the mean SOC content, 2 ∼ 1.96 the 97.5 % percentile of a normal distribution and n is the sample size. So, if we require that 𝑀𝑂𝐸 ≤ Δ, then the corresponding sample size should be: n ≥ 4 𝑠2 Δ2 This relationship makes clear that the sample size increases fast if a small MOE is aimed for. Aiming for a MOE half as small implies that the sample size should be four times as large. The values of the mean SOC content were gathered from the preliminary (2021-2022: first work year) results for the Flemish Cmon monitoring network (Oorts et al., 2023). For the 0-10 cm layer of cropland soils a mean SOC content of 1.62% (n=55) was found, for permanent grassland this was 3.81% (n=46) and for forests 5.17% (n=51). Despite its incompleteness (total sample size every 10 years: ncropland = 794; npermanent grassland = 406; nforests = 490), the ongoing Cmon monitoring network provides us with the most recent and comprehensive (good geographical spread) available for Flanders. The results presented in Figure 6 highlight the importance of a large sample size to achieve acceptable levels of precision, even for relatively small spatial scales, in forests and to a lesser extent also in grasslands. In cropland, on the other hand, smaller sample size seems to be acceptable. Deliverable 7.3 17 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 6: Relation between the margin of error for the topsoil SOC content (%) and the sample size for cropland (yellow), grassland (green) and forest (red) at plotup to field-scale. The full lines indicate the minimum and maximum variability found for each land-use. The measures of plot-scale variability (coefficient of variance – CV) were collected from existing scientific literature (see Table 1). The values of mean 0-10 cm layer SOC content (i.e. cropland: 1.62%; grassland: 3.81%; 5.17%) were gathered from the preliminary results of the Cmon monitoring network in Flanders (Oorts et al., 2023). From all studies listed in Table 1, the spatial scales investigated by Montagne et al. (2007) (400 m2 – #14), Saby et al. (2008) (1-400 m2 – #20), Goidts et al. (2009) (50.3 m2 – #24) and Poeplau et al. (2022) (400 m2 – # 33) came closest to the experimental conditions of our study (100 m2). The studies of Montagne et al. (2007) and Goidts et al. (2009) were thought less useful for our situation, because they had focused on a narrow range of soil types, while Saby et al. (2008) and Poeplau et al. (2022) assessed a broad range of European soils. Eventually, the value of spatial variability from Poeplau et al. (2022) was selected for our own calculations of expected precision, because it offered the most conservative estimation (8%). Figure 7 presents the expected precision with varying sample size for this level of spatial variability. At 16 the margin of error is found at 0.06%, only slightly above the detection limit of 0.04% at the laboratory of ILVO, and it stays below 0.10% until a sample size of seven. In Figure 8 the percentual change in the margin of error is given for each change in sample size by one. With the change in sample size from seven to eight there still is a considerable improvement to be obtained in terms of precision, but after that the gain starts to level off. At a sample size of eight, we could expect a margin of error of 0.09% (Figure 7), which would mean a 29% reduction from the sample size of 16. Deliverable 7.3 18 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 7: Variation in the margin of error for the topsoil SOC content (%) with the sample size expected for cropland at plot scale (i.e. 100 m2) by Poeplau et al. (2022) (CV: 8.0%). Figure 8: Change (%) in the margin of error by a stepwise (one) increase in sample size. Deliverable 7.3 19 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 3.2 Comparison of data on plot-scale precision from scientific literature to experimental Flemish conditions 3.2.1 Experiment To scrutinize the results for precision of SOC content determination that were presented in Section 3.1, we decided to replicate the sampling of three distinct plots of the experimental sampling campaign discussed in Section Errore. L'origine riferimento non è stata trovata. three times each. This replication of the composite samples (8 subsamples) allowed us to estimate a CV value (and derived margin of error value) that could be compared with the results from scientific literature presented in Table 1. One replicated sampling plot was selected randomly for each of the three soil types from the Belgian soil classification (i.e. silt (A – Leem), sandy loam (L – Zandleem) and light sandy loam (P – Licht zandleem) – (Van Ranst & Sys, 2000)) that were found on the two experimental farms (One in the Belgian municipality of Asse and the other in the municipality of Meise) that are characterized in more detail in Section 4.1.1.2. The three selected plots were: • Asse – sampling plot 19 (silt) • Meise – sampling plot 6 (sandy loam) • Meise – sampling plot 11 (light sandy loam) The exact locations of these replicated sampling plots are presented in Figure 9. Figure 9: Locations of the three replicated sampling plots (see numbers) in the farms in Asse (blue) and Meise (red). The soil types found within the boundaries of the indicated cropland fields are: dry silt ( ), moist silt ( ), wet silt ( ), dry sandy loam ( ), moist sandy loam ( ), wet sandy loam ( ) and moist light sandy loam ( ). Deliverable 7.3 20 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 3.2.2 Results Results of the replicated SOC content determination of the topsoil (0-10 cm) of the selected sampling plots are presented in Table 2. The measured SOC content varied between 1.16% at sampling plot 11 and 1.99% at sampling plot 6 both from the farm in Meise. The plot-scale variation (CV) ranged between 1.2% at sampling plot 6 of the farm in Meise and 7.1% at the sampling plot 19 of the farm in Asse. These CV values all fall within the range observed in Table 1, although rather at the lower section of this range. In Section 3.1, we reasoned that a CV value of 8.0%, as found by Poeplau et al. (2022), would be a prudent choice to use when calculating the desired plot-scale sample size. Since none of the CV values of the replicated sampling plots were found to be higher than 8.0%, we can safely assume that we will not have overestimated the expected precision of the SOC content determination. In fact, with the average CV of 4.3% derived at the farms, we would expect a precision (margin of error) of 0.045% at a sample size of eight (Figure 10), so 50% better than expected by Poeplau et al. (2022) (Figure 8). Table 2: Replicated (n=3) SOC content (%) of 0-10 cm depth interval at three experimental plots on distinct soil texture Farm Sample nr. Soil texture SOC content (%) – Replications Average (%) Standard deviation (%) CV (%) A B C Asse HvP_19 Silt 1.49 1.26 1.33 1.36 0.10 7.1 Meise DL_6 Sandy loam 1.99 1.93 1.96 1.96 0.02 1.2 Meise DL_11 Light sandy loam 1.17 1.28 1.16 1.20 0.05 4.5 Average: 4.3 Deliverable 7.3 21 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 10: Variation in the margin of error for the 0-10 cm SOC content (%) with the sample size expected for the experimental conditions at plot-scale (i.e. 100 m2) with a CV of 4.3%. 3.2.3 Costs (ILVO test case – Flanders) Sampling campaigns for the determination of SOC content by laboratory analyses are generally associated with relatively high costs (Batjes et al., 2023; Viscarra Rossel & Brus, 2018). The most important sources of cost for farm-scale sampling campaigns are sampling labor by field technicians, laboratory processing and analysis and fuel for transport. Some more detailed information on the contemporary (February 2024) costs associated with these factors at ILVO in Flanders (Belgium) is presented in Table 3. The cost of the sampling equipment, like RTK GPS and augur set, was not included in this calculation, since we can assume that most of the equipment can be used for multiple years. This is how the products are marketed. The cost for sample preparation and laboratory analysis does take personnel, overhead and machine/consumables into account. Table 3: Contemporary (Q1 2024) information on the costs associated with SOC sampling campaigns at the research institute of ILVO in Flanders (Belgium) Source Costs Personnel (2 Field technicians) 2 x € 366 workday-1 Sample preparation and laboratory analysis (dry combustion) – Research laboratory; no market prices € 41.30 sample-1 Fuel (diesel) € 1.867 l-1 Deliverable 7.3 22 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 A change in the plot-scale sample size (number of subsamples for one composite sample) is expected to have a large influence on the costs associated with personnel for sampling and potentially also on the costs associated with fuel consumption. However, the latter factor will mainly be impacted if the change in sample size alters the duration of the sampling campaign to such a degree that extra or less days are required. A change in sample size will not impact the costs associated with laboratory analysis, since it will not alter the total number of samples to be analyzed in a sampling campaign. In Table 4 we present the costs associated with the experimental sampling campaign at the farm in Asse, Flanders (see Section 4.1.1.2). In total 64 composite samples were taken on the farm’s 71.4 ha of cropland. All fields are geographically concentrated, the maximum distance between two fields is 2.8 km. The travel time between ILVO and the farm is around 30 minutes by car. The overall costs were calculated for both a plot-scale sample size of 16 and eight. We estimated that the sampling of one plot with 16 subsamples would take around 30 minutes (i.e. 10 minutes to locate sampling location with RTK GPS and 20 minutes for the actual sampling), while this could be reduced to 20 minutes (i.e. 10 minutes to locate sampling location with RTK GPS and 10 minutes for the actual sampling) when only eight subsamples were taken. The traveling time between sampling locations was kept constant at 15 minutes. According to these estimations, the sampling size of 16 would allow for 8.7 plots to be sampled by two field technicians per workday (= 7.5 h) and would therefore require 7.4 workdays to sample all 64 plots. The sampling size of eight would allow for 11.1 plots to be sampled per workday and require just 5.7 workdays to complete the sampling campaign. As a result, the sample size of 16 would cost € 5673.2, while the sample size of eight would cost € 4822.7. Consequently, a 15% reduction in overall costs could be achieved by halving the plot-scale sampling size from 16 to eight for this test case situation. Table 4: Costs associated with sampling 44 Sentinel-2 plots/pixels for topsoil SOC content determination with 16 and 8 subsamples at farm in Asse, Flanders Source of cost 16 subsamples 8 subsamples Personnel € 5405.5 € 4204.3 Sample preparation and laboratory analysis € 2643.2 € 2643.2 Fuel € 219.3 € 170.6 Total € 8268.0 € 7018.1 Research question 1: Is it possible to reduce the plot-scale (10mx10m) subsampling costs by reducing the number of subsamples currently used in the protocol of the Flemish Cmon monitoring network and used Deliverable 7.3 23 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 for the original ENVISION sampling campaign (i.e. 16 subsamples), without drastically reducing the precision of the SOC content estimation? Based on scientific literature, the change in the plot-scale (10mx10m) sample size from 16 to eight for the determination of cropland topsoil SOC content in Flanders was expected to worsen the precision (95% margin of error) by 30% from 0.06% to 0.09%, while reducing the overall costs in our experimental test case by 15%. The results derived from three replicated experimental sampling plots (0.045%) indicated that the values derived from scientific literature were not overestimating, but rather underestimating the precision of plot-scale topsoil SOC content determination at Flemish cropland conditions. Within our experimental conditions, we thought that this expected slight loss in precision would be an acceptable trade-off for the given reduction in sampling costs. However, this decision is very case dependent (e.g. increase in budget, precision/accuracy becomes limiting factor) and could relatively easily be reconsidered under new conditions or objectives with the results provided in this section. Deliverable 7.3 24 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 4 Satellite-based sampling strategies (D7.1) Research question 2 (Section 4): Can the Sentinel-2 spectral bands (satellite-based) or pre-existing SOC maps (map-based – e.g. ENVISION PyCaret SOC map for the region of Flanders) provide us with adequate knowledge on topsoil SOC variability at fieldand farm-scale to help optimize soil sampling strategies? Research question 3 (Section 4): Is it possible to meaningfully increase the accuracy of the ENVISION TensorFlow SOC map at field-scale by its local (i.e. farm-scale) calibration? 4.1 Materials and Methods 4.1.1 Test sites The newly developed satellite-based sampling strategies were assessed at three test sites: one in the province of Parma in Italy and two in the province of Vlaams-Brabant (Flanders, Belgium) in the municipalities of Asse and Meise (Figure 11). The Italian test site consists of two adjacent cropland fields on which a new sampling strategy, that was directly based on the Sentinel-2 spectral bands with the Kennard-Stone selection algorithm (satellite-based), was compared to a simple random sampling strategy. At the Belgian test site in Asse, the satellite-based sampling strategy was tested at farm-scale and compared to a sampling strategy based on a pre-existing SOC map (see Text box 1) for the region of Flanders (map-based). At the Belgian test site in Meise, the map-based sampling strategy was tested at farm-scale. We explicitly searched for one spatially concentrated farm with relatively homogeneous soil types (Asse) and another more spatially extended with a broader range of soil types (Meise). The sampling strategies will be discussed in more detail in Section 4.1.3. Deliverable 7.3 25 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 11: Location of the test sites in Parma, Italy (field-scale) and Asse and Meise, Flanders – Belgium (Farmscale) 4.1.1.1 Italian test site The Italian test site (Figure 12) consist of two fields (8 ha), located in the Po Valley North Italy, in a completely flat croplands area. The two fields are situated in the area adjacent to the main Barilla spa plant in the province of Parma. This area consists of well drained and deep alluvial soils classified as Fluvi-Eutric Cambisols (WRB, 2015); the surface soil layers are calcareous with clay loam texture; the deeper layers are more calcareous with loam texture. Fields were managed by crop rotation alternating cereals and leguminous; fields are crossed by drainage channels running parallels 30 m from each other. Tillage is usually carried out Deliverable 7.3 32 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 16 – Van Ranst & Sys, 2000)). In contrast to farm in Asse, which is characterized by homogeneous silt soils (Figure 14), the farm in Meise is spread out over three distinct soil types according to the Belgian soil map ( Figure 16): fields 3 and 4 are classified as silt, 1, 2, 5, 6, 8 and 9 as sandy loam and 7 as light sandy loam. Also, in terms of drainage class the fields can be considered quite varied: “dry” for fields 3, 4 and 6, “moist” for fields 1, 2, 7, 8 and 9 and “wet” for field 5. On the Soil Association Map of Flanders fields 1, 2 and 5 fall within region nr. 29 (wet sandy loam soils with texture B horizon or with crumbled texture B horizon), fields 3 and 4 in region nr. 31 (silty soils with texture B horizon: normal association), field 6 in region nr. 28 (dry sandy loam soils with texture B horizon or with crumbled texture B horizon) and fields 7, 8 and 9 in region nr. 19 (complex of associations 15 “wet sandy and loamy sand soils with humic and/or iron B horizon” and 17 “wet sandy to light sandy loam soils with color B horizon or with texture B horizon”). Table 6: Characteristics of the 9 cropland fields of the farm in Meise: area, soil cover at the time of soil sampling, primary crop(s) for the 2023 & 2024 growing seasons (excluding cover crops) and soil texture and drainage classes derived from the Belgian soil classification map. Field nr. Area (ha) Soil cover Crop 2023 Crop 2024 Texture class Drainage class Soil Association nr. 1 0.4 English ryegrass English ryegrass English ryegrass Sandy loam Moist 29 2 4.1 Winter barley Winter wheat Winter barley Sandy loam Moist & wet 29 Deliverable 7.3 33 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 3 0.8 Maize stubble Maize Potatoes Silt Dry 31 4 0.9 Maize stubble Maize Potatoes Silt Dry 31 5 2.3 Bare Potatoes Winter wheat Sandy loam Moist & wet 29 6 2.4 Maize stubble Maize Potatoes Sandy loam Dry & moist 28 7 1.0 Maize stubble Maize Maize Light sandy loam Moist 19 8 3.0 Maize stubble Maize Maize Sandy loam Moist 19 9 1.4 Maize stubble Maize Maize Sandy loam Moist 19 Total: 16.3 Figure 16: Soil types (texture and drainage classes of Belgian soil classification) present in the immediate vicinity of the farm in Meise. The soil types found within the boundaries of the indicated cropland fields are: dry silt ( ), dry sandy loam ( ), moist sandy loam ( ) and wet sandy loam ( ). 4.1.2 Sentinel-2 composite layer – Mosaicking A pixelwise temporal mosaicking approach was applied for the selection of bare soil images using Sentinel-2 imagery collections and obtaining composite bare soil layers for each test site from the time series 2020 – 2022 (Castaldi et al., 2023). The mosaicking approach was implemented by using the rgee R package (Aybar et al., 2020) to wrap the Earth Engine Python API for working on the Google Earth Engine environment (Gorelick et al., 2017) from within R (R Core Team, 2024). Google Earth Engine is a cloud-based platform Deliverable 7.3 34 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 which was here used to gain access to Copernicus Sentinel-2 MultiSpectral Instrument (MSI) level 2A dataset provided by European Space Agency (ESA). The approach consisted of two steps. In the first step, three filters were applied for each pixel and acquisition date to identify and select bare soil pixels: I. Atmospheric filter. This includes a cloud filter to remove dates affected by clouds and cloud shadow. The cs quality band from the combination between Cloud Score+ and S2_HARMONIZED dataset was used with a cs threshold of 0.7 to remove thin clouds, haze, and cirrus shadows. The cs value is the similarity between the observed pixel and theoretical clear reference pixel based on their spectral distance. A snow mask was also applied using the MSK_SNWPRB quality band provided by ESA, removing records with >10% probability of snow on the ground. II. Green vegetation filter. The Normalized Difference Vegetation Index (NDVI, calculated from band 4 (B4) and 8 (B8) as follows (B8-B4)/(B8+B4)) was computed and pixels with NDVI > 0.35 were filtered out because they were likely affected by green vegetation. III. Dry vegetation + soil moisture filter. Normalized Burn Ratio 2 (NBR2) index was computed and pixels having NBR2 > 0.125 (Castaldi et al. 2019b) were filtered out because they were likely affected by dry vegetation and/or by high soil moisture content. In the second step, if at least 3 bare soil dates were left after applying the filters for each pixel, the available data were combined to implement a mosaicking approach selecting the median (50th percentile) reflectance values for each band and pixel throughout the time series, with the aim to obtain spectral data representative of the average bare soil conditions not affected by extreme reflectance values. The outcome of the mosaicking approach is a composite bare soil layer of 20 m resolution for each test site. Pixels close to the field borders (20 m buffer) were removed to avoid including area mixed pixels affected by spectral information of other material than soil. The layers include reflectance information of the following bands of the MSI instrument (central wavelength): B1: (443 nm), B2 (492 nm), B3 (559 nm), B4 (664 nm), B5 (704 nm), B6 (740 nm), B7 (782 nm), B8 (842 nm), B8A (865 nm), B11 (1613 nm), and B12 (2202 nm) (ESA, 2015). 4.1.3 Soil sampling strategies 4.1.3.1 Satellite-based sampling strategy The composite bare soil layer obtained by Sentinel-2 time series was used to extract spectra from the bare soil pixels (20 m) within the boundaries of the test sites. After that the Kennard-Stone algorithm was used to Deliverable 7.3 35 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 select 50 spectra corresponding to 50 pixels within the test site using the prospectr package (Stevens & Ramirez-Lopez, 2024) in R. The Kennard-Stone algorithm optimizes the coverage of the spectral variability, selecting spectra uniformly distributed over the feature space from the entire spectral (Castaldi, Chabrillat, & van Wesemael, 2019; Kennard & Stone, 1969). The algorithm starts finding and selecting the two spectra that are furthest apart based on the Mahalanobis distance and removing them from the input matrix. Then, the procedure was repeated until the algorithm selected the desired number of spectra that correspond to 50 pixels and therefore 50 sampling locations for the field-scale application in Italy and 20 pixels for the farmscale experiment in Asse, Belgium. For the Italian site, we compared the satellite-based strategy and the Random selection, obtained randomly selecting 50 points within the test site in the QGIS environment. A composite bare soil layer was obtained both for the Italian (Figure 17) and the Belgian site (Figure 18). Figure 17: Composite bare soil layer obtained at the Italian test site in Parma using multi-temporal Sentinel2 collection (R: Sentinel-2 B4; G: Sentinel-2 B3; B: Sentinel-2 B2) and location of the sampling points according to the random and satellite-based sampling strategy. Deliverable 7.3 36 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 18: Composite bare soil layer obtained at the Belgian test site in Asse using multi-temporal Sentinel-2 collection (R: Sentinel-2 B4; G: Sentinel-2 B3; B: Sentinel-2 B2) and location of the sampling points according to the map-based and satellite-based sampling strategy. 4.1.3.2 Map-based sampling strategy This sampling strategy used the original PyCaret version of the original ENVISION SOC map as a source of ancillary information for the spatial variability of topsoil SOC content (see Text box 1). At the time of sampling (February-March 2024) the newer ENVISION TensorFlow SOC map had not yet become available. The later calibration stages of our study (see Section 4.1.5) did make use of the ENVISION TensorFlow SOC map, because of its more realistic regional range in topsoil SOC content. The strategy was implemented in the freely Deliverable 7.3 37 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 available and relatively user-friendly QGIS software. One of the main objectives during the development was to keep the procedure for the selection of sampling points as easily reproducible as possible for other farms throughout Flanders. The seven broad steps to this procedure are: 1) Division of the regional ENVISION PyCaret SOC map into ten even SOC content classes, each with a 0.1% range, between its approximate minimum (0.9%) and maximum (1.9%). To make computational power less of a limiting factor, this polygenization (i.e. change from raster layer with continuous output, to a vector layer consisting of discrete polygons) step was only done for a clipped (22 km x 16 km) portion of the ENVISION PyCaret SOC map immediately surrounding the farms (Figure 19). Most of the croplands surrounding the farms are characterized by, for Flanders, relatively low SOC contents, especially in the more silt-rich southern sections (see Figure 14) of this area. Figure 19: Clipped and polygenized section of the ENVISION PyCaret SOC map in the immediate surroundings of the farms in Asse and Meise. The experimental field boundaries are displayed in red. The polygenization divided the continuous ENVISION PyCaret SOC map into ten even classes, each of 0.10% range between 0.9% and 1.9% SOC content. 2) Clipping new SOC class vector layer on field boundaries (Figure 20). A buffer zone of 10 m within the field boundaries was additionally clipped from the SOC class layer to minimize the impact of boundary effects, such as the shadows of trees, hedges, houses, etc.. Deliverable 7.3 38 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 3) The smallest polygons (≤ 150 m2) were excluded from the SOC class layer to enhance the chances of selecting sampling points in larger, homogeneous sections of the field, where we expect the uncertainty of the SOC map to be lower. Figure 20: SOC classes found within the field boundaries + 10 m buffer of the farm in Asse. 4) To compile the farm-scale (linear) calibration dataset a pre-described number of polygons is randomly selected per SOC class (yellow frame – Figure 21). Five polygons were randomly selected for each of the six SOC classes present within field boundaries of the farm in Asse. However, since not all SOC classes have five polygons within the experimental boundaries, we ended up with 24 selected polygons in total: four for SOC class 1 and 4, five for 2, 3 and 5 and one for 6. 5) A similar procedure was followed to compile a separate field-scale validation dataset for field 1 of the farm in Asse (pink frame – Figure 21). This field was selected because it has the largest area (18.3 ha), while it also contains a large range of SOC classes (2-5). In total 20 additional polygons were selected: six of SOC classes 2, 3 and 4 and two of 5. Deliverable 7.3 39 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 21: polygons randomly selected per SOC class for farm-scale calibration (yellow frame – n = 24) and field-scale validation (pink frame – n = 20) at the farm in Asse. 6) A 4 m buffer zone inside the boundaries of the selected polygons was removed to try to keep the entire sampling plots (i.e. circle of 5 m radius around the selected sampling points) within the same polygon. 7) Random selection of one sampling point for each buffered polygon that was selected for both calibration (step 4) and validation (step 5) (red dots – Figure 22). Deliverable 7.3 40 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 22: sampling points (red dot) randomly selected within each highlighted polygon at the farm in Asse. The same selection procedure for sampling points was applied to the farm in Meise. Because of the smaller acreage of cropland (16.2 ha) for Danial Lemmens’ farm compared to the farm in Asse (71.4 ha), we decided to also limit the sample size for the local calibration dataset to twelve: three sampling points in SOC classes 2 and 4 and two in 3, 5 and 6. The field-scale validation dataset was compiled for field 2 (i.e. largest field, with relatively large range of SOC classes). To achieve an accurate estimation of the field-scale SOC content as possible, we decided to select 20 sampling points within the validation field: four for SOC classes 2 and 3, five for 4 and 5 and two for six. Figure 23 presents the selected polygons for calibration and validation of fields 2 (validation field) and 5 in the municipality of Meise. The polygons for calibration that were selected on other fields are not included on Figure 23 because of the large distance between the fields. Experimental sampling campaigns for both farms concerned were performed between February 19th and 24th by ILVO field technicians. Deliverable 7.3 41 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 23: sampling point (red dot) selected in fields 2 and 5 of the farm in Meise for farm-scale calibration (yellow frame - n = 12) and field-scale validation (pink frame - n = 20). 4.1.3.3 Statistical comparison between sampling strategies The sampling strategies tested both at fieldand farm-scale were compared by observing the main descriptive statistics of the SOC values and evaluating if significant differences exist in terms of variance, both for topsoil SOC content and each extracted Sentinel-2 band. The variance differences were evaluated running the Levene’s test after the Box-Cox transformation to comply with the normal assumption. Only for the random sampling strategy at the field experiment in Italy, the sampling selection was repeated 100 times to obtain more robust statistics for SOC content and spectral comparison. 4.1.4 Sampling, processing and analysis protocol 4.1.4.1 Sampling For the Belgian test sites, once on the field (Figure 24), the sampling point's location was identified by a CHC Navigation i90 RTK GPS (centimeter accurate). Afterwards an area with a radius of 5 m around the sampling point was marked by red sticks, within which 8 subsamples were collected randomly from two separate Deliverable 7.3 48 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 The variance trends according to the sampling size is quite different between the random and satellite-based sampling strategies: an almost constant trend for random, while for the satellite-based approach we observed a first sharp increase after 35 samples and a second after 40 (Figure 28). The variance of the random approach is higher than the satellite-based approach until 42 samples, after that the satellite-based approach was able to catch more SOC content variability. The Levene’s test for the equality of the variances provided most of the times significant differences between the tested approaches, except using 35-40 samples. Figure 28: Variance of the SOC content (%) for random and satellite-based sampling strategies according to the number of samples 4.2.2.2 Farm test sites The results for SOC content from the map-based sampling strategy at the farm in Asse are presented per SOC class (i.e. strata) in Table 8. The measured SOC content has a markedly higher range (0.71-2.47%) than the ENVISION PyCaret SOC map predicted (0.90-1.50%), but the general order of the SOC classes is still followed. This is especially true for the extreme values: the four lowest SOC contents are all found in SOC class 1 and the highest value is found in SOC class 6 and the second highest in SOC class 5. In the middle SOC classes (25) the differences become less discernable. The variability (CV) within several SOC classes comes very close to the total farm-scale variability (26.2%). Only SOC classes 1 (9.0%) and 3 (15.8%) and show a clear difference with the overall variability. This sampling strategy thus seems to be effective at covering the entire range of farm-scale SOC content, even at relatively low sampling sizes, but is less effective at limiting the variability within the SOC classes or strata (compared to the total variability), which would be required to help reduce the sample size compared to more simple random or geographical sampling strategies (Bettigole et al., 2023; De Gruijter et al., 2006; Peltoniemi et al., 2007). There appear to be environmental or management factors at play that could explain a sizeable portion of the SOC content spatial variability which is not covered by Deliverable 7.3 49 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 differences in the soil’s spectral patterns. The map-based sampling strategy might therefore win in costefficiency by adding new sources of ancillary data. However, it could also be that the range of the ENVISION PyCaret SOC estimates per SOC class was simply too small compared to the prediction models accuracy. It might have been more prudent to select an intra SOC class range that was higher than the RMSE of the ENVISION PyCaret prediction model for the Soil Association region of the farm in Asse (see Figure 3 - #32: 0.19%). Table 8: SOC content (% - 0-10 cm) from the map-based sampling strategy per SOC class observed on the ENVISION PyCaret SOC map at the farm in Asse (n=24) SOC class PyCaret SOC prediction interval (%) SOC content (% - 0-10 cm) Sample size Avg. (%) St. Dev. (%) CV (%) Min. (%) Max. (%) Area (m2) 1 0.9 - 1.0 4 0.78 0.07 9.0 0.71 0.86 1452 2 1.0 - 1.1 5 1.26 0.30 23.8 0.91 1.81 214951 3 1.1 - 1.2 5 1.33 0.21 15.8 1.01 1.59 342399 4 1.2 - 1.3 4 1.30 0.28 21.5 0.98 1.68 54814 5 1.3 - 1.4 5 1.51 0.41 27.2 1.16 2.30 6046 6 1.4 - 1.5 1 2.47 - - - - 524 Total 0.9 - 1.5 24 1.30* 0.34 26.2 0.71 2.47 620186 *Area adjusted for SOC classes within the farm’s field boundaries The measured SOC content is still relatively highly correlated to the ENVISION PyCaret SOC estimates. The correlation with the ENVISION TensorFlow SOC estimates could be considered rather moderate (Table 9). Another noteworthy result from this table is the low correlation between the two ENVISION SOC maps. Table 9: Correlation between three different sources of SOC content determination: measured with laboratory analysis and estimates from PyCaret and TensorFlow versions of the original ENVISION SOC maps at the farm in Asse (n=24) SOC content (0-10 cm) Original Pycaret Original TensorFlow SOC content (0-10 cm) 1.00 PyCaret SOC prediction 0.69 1.00 TensorFlow SOC prediction 0.51 0.32 1.00 Table 10 presents the average SOC content predictions per field of the farm in Asse determined from the samples of the map-based sampling strategy, using the average per SOC class (Table 8), and from the estimates of both the original (regional) ENVISION SOC maps. Both ENVISION SOC maps make a systematic underestimation of the SOC content, but this is slightly less the case for the ENVISION TensorFlow SOC map. For the only field with a relatively high intra-field samples size (field 1 – n=27), the overall average of all the samples was 1.18%, so actually closer to the estimations of the original ENVISION PyCaret (1.12%) and TensorFlow (1.15%) SOC maps than the 1.32% determined at farm-scale with the map-based sampling strategy. The within-field spatial variability (CV) was relatively low for both ENVISION SOC maps compared to the results from the literature studies discussed in Table 1. At farm-scale the spatial variability for these SOC Deliverable 7.3 50 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 maps remained just as low, much lower than the value determined for the map-based sampling strategy (26.2%). Table 10: Area-adjusted (for SOC classes) average SOC content (%) determined by laboratory analysis and average (%), standard deviation (%) and CV (%) of the SOC content derived from the ENVISION PyCaret SOC map and ENVISION TensorFlow SOC map for each field of the farm in Asse. Field nr. Area (ha) SOC content Original Pycaret Original TensorFlow Avg. (%) Avg. (%) St. Dev. (%) CV (%) Avg. (%) St. Dev. (%) CV (%) 1 18.3 1.32 1.12 0.06 5.4 1.15 0.05 4.3 2 7.2 1.31 1.07 0.03 2.8 1.13 0.02 1.8 3 2.9 1.29 1.10 0.03 2.7 1.18 0.04 3.4 4 1.1 1.30 1.10 0.03 2.7 1.17 0.04 3.4 5 4.6 1.29 1.15 0.06 5.2 1.12 0.02 1.8 6 2.0 1.31 1.13 0.04 3.5 1.26 0.10 7.9 7 0.4 1.40 1.16 0.05 4.3 1.22 0.03 2.5 8 0.7 1.31 1.15 0.03 2.6 1.27 0.07 5.5 9 0.9 1.32 1.11 0.03 2.7 1.19 0.03 2.5 10 1.8 1.32 1.08 0.06 5.6 1.18 0.05 4.2 11 3.9 1.26 1.06 0.04 3.8 1.07 0.01 0.9 12 2.1 1.31 1.11 0.03 2.7 1.14 0.02 1.8 13 1.0 1.27 1.10 0.04 3.6 1.14 0.02 1.8 14 4.6 1.33 1.29 0.06 4.7 1.16 0.03 2.6 15 5.5 1.28 1.14 0.04 3.5 1.15 0.03 2.6 16 2.4 1.29 1.12 0.04 3.6 1.19 0.04 3.4 17 3.7 1.32 1.13 0.04 3.5 1.09 0.03 2.8 18 8.0 1.29 1.13 0.06 5.3 1.13 0.02 1.8 19 0.4 1.32 1.11 0.04 3.6 1.12 0.03 2.7 Total 71.4 1.30 1.13 0.04 3.5 1.14 0.01 0.9 In Figure 29, the results for minimum, maximum and range of the measured SOC content from the map-based sampling strategy at the farm in Asse are compared with the corresponding results from the satellite-based sampling strategy. For the minimum value there is a clear congruence between both sampling strategies, but this is not the case for the maximum value, and therefore also range. Just as was noticed in Figure 27 at fieldscale in Italy, the satellite-based sampling strategy proved to be more efficient at selecting the left side of the SOC content distribution than the right side. However, this difference between the sampling strategies might also be explained by the extended buffer zone of 20 m for satellite-based compared to 10 m for map-based. It is in this diverging zone that most of the polygons for SOC classes 5 and 6 could be found, and so also almost all high SOC contents. The zone of between 10 m and 20 m from the field border often falls within the field’s headland, where traffic intensity is highest and most turning is done (Augustin et al., 2020), both contributing to a higher risk of soil compaction (Sunoj et al., 2021; Welch et al., 2016). Despite its lower crop yields (e.g. Sunoj et al., 2021), and thus lower potential carbon inputs, the headlands have been previously shown to have relatively high levels of SOC (e.g. Welch et al., 2016), which could be explained by the reduced carbon mineralization under higher bulk densities, observed by De Neve & Hofman (2000) for loamy sand soils, that Deliverable 7.3 51 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 would allow for carbon accumulation over time (Welch et al., 2016). It is striking how little the lines for both sampling strategies change across the tested sample sizes. This seems to highlight the ability of the sampling strategies to efficiently pick SOC extremes with their respective zone of interest. Figure 29: Minimum, maximum and range of SOC content (%) for the mapand satellite-based sampling strategies at the farm in Asse according to the sample size The variance trends according to the sampling size are remarkably different between the mapand satellitebased sampling strategies for the farm in Asse: an almost constant trend for the satellite-based approach with relatively low variance (~0.06%), while the map-based approach started with very high variance (0.50%), which decreased exponentially with the sample size to more steady state around 0.20% (Figure 30). Levene’s test did, however, not show any significant differences in variance between mapand satellite-based sampling strategies qua SOC content. This can probably be explained by the fact that much of the high variance of the map-based approach is caused by a limited number of samples with really high SOC content. Figure 30 also presents the trend of the variance according to the sample size for the map-based sampling strategy at the farm in Meise. At this farm the map-based approach does not show an exponential decline in variance with the sample size. The SOC variance makes an initial steep rise from sample size 3 to 6 and then starts to decline steadily but does not reach a steady state within the rather limited range of tested sample sizes. Deliverable 7.3 52 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 30: Variance of the SOC content (%) for the mapand satellite-based sampling strategies at the farm in Asse and the map-based sampling strategy at the farm in Meise, all according to the sample size The results for SOC content from the map-based sampling strategy at the farm in Meise are presented per SOC class in Table 11. Just as for the farm in Asse (Table 8), the SOC content has a markedly higher range (1.17-3.23%) than the ENVISION Pycaret SOC map predicted (1.00-1.50%). However, here the general order of the SOC classes is not really followed. The lowest SOC contents were found in the lowest (2) and highest (6) SOC classes, while the highest SOC contents were in the middle SOC classes (3-5). The intra SOC class variability for all classes, except for 2, is notably lower than the overall farm-scale variability. Although, we need to consider here the low number of samples per SOC class. The higher number of Soil Association regions that are found within the farm in Meise (Table 6) compared to the farm in Asse (Table 5) is represented by the higher farm-scale standard deviation in SOC content (0.57% vs. 0.34%), but the markedly higher measured SOC contents for the farm in Meise did result in similar CV values for both farms (28.5% vs. 26.2%). Table 11: SOC content (% - 0-10 cm) from the map-based sampling strategy per SOC class observed on the ENVISION PyCaret SOC map at the farm in Meise (n=12) SOC class PyCaret SOC prediction interval (%) SOC content (% - 0-10 cm) Sample size Avg. (%) St. Dev. (%) CV (%) Min. (%) Max. (%) Area (m2) 1 0.9 - 1.0 0 - - - - - 0 2 1.0 - 1.1 3 1.46 0.38 26.0 1.17 1.99 32621 3 1.1 - 1.2 2 2.42 0.06 2.5 2.36 2.48 61346 4 1.2 - 1.3 3 2.10 0.06 2.9 2.06 2.19 26153 5 1.3 - 1.4 2 2.77 0.47 17.0 2.30 3.23 12178 6 1.4 - 1.5 2 1.48 0.06 4.1 1.42 1.54 1363 Total 0.9 - 1.5 12 2.00 0.57 28.5 1.17 3.23 133661 *Area adjusted for SOC classes within the farm’s field boundaries Deliverable 7.3 53 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 We observed a very low correlation between the measured SOC content and the ENVISION PyCaret SOC estimates, while the correlation with the ENVISION TensorFlow was high (Table 12). The ENVISION TensorFlow SOC map would probably have been a much better source of ancillary data, compared to the PyCaret version, to provide knowledge on the farm-scale spatial variability in SOC content. Similarly, as at the farm in Asse (Table 9), the PyCaret and TensorFlow SOC maps showed a weak correlation to each other. Table 12: Correlation between three different sources of SOC content (% - 0-10 cm) determination: measured with laboratory analysis and estimates from PyCaret and TensorFlow versions of the original ENVISION SOC maps at the farm in Meise (n=32) SOC content (0-10 cm) Original Pycaret Original TensorFlow SOC content (0-10 cm) 1.00 PyCaret SOC prediction 0.09 1.00 TensorFlow SOC prediction 0.81 -0.32 1.00 Table 13 presents the average SOC content predictions per field of the farm in Meise determined from the samples of the map-based sampling strategy, using the average per SOC class (Table 11), and from the estimates of both the original (regional) ENVISION SOC maps. Both ENVISION SOC maps make a markedly high systematic underestimation of the SOC content. However, whereas the ENVISION PyCaret SOC map estimates the field-scale SOC content at both Belgian test sites to be of a similar order of magnitude (Table 10), the ENVISION TensorFlow SOC map did estimate relatively high field-scale SOC contents for some of the fields of the farm in Meise (fields 6, 7 and 8). This difference in Soil Association regions (Table 6) does seem to be better visible in the ENVISION TensorFlow SOC map. For the only field with a relatively high intra-field samples size (field 2 – n=22), the overall average of all the samples was 2.11%, so the determination at farmscale with the map-based sampling strategy (2.26% – n=12) is much more exact than the ENVISION SOC maps (PyCaret: 1.19%; TensorFlow: 1.19%), but it still does not have the accuracy we would expect for carbon farming initiatives (Andries et al., 2021). As expected from the geographical and feature space spread of the fields, the within-field variability at the farm in Meise was slightly higher than at the more spatially concentrated farm in Asse. However, the values were still lower than would be expected by the results discussed in Table 1. The ENVISION PyCaret SOC map predicted a farm-scale variability that was of a similar magnitude than the field-scale variability, much lower than the value determined for the map-based sampling strategy (28.5%). In contrast, the ENVISON TensorFlow SOC map predicted an approximately twice as high farm-scale variability compared to the field-scale. Goidts et al. (2009) did not observe large increases in spatial variability of SOC stocks from fieldto landscape-scale for croplands in Wallonia, Belgium. But Saby et al. (2008) did observe a steady increase in spatial variability of SOC content with increasing spatial scales in a review of European field studies. Deliverable 7.3 54 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Table 13: Area-adjusted (for SOC classes) average SOC content (%) determined by laboratory analysis and average (%), standard deviation (%) and CV (%) of the SOC content derived from the ENVISION PyCaret SOC map and ENVISION TensorFlow SOC map for each field of the farm in Meise. Field nr. Area (ha) SOC content Original Pycaret Original TensorFlow Avg. (%) Avg. (%) St. Dev. (%) CV (%) Avg. (%) St. Dev. (%) CV (%) 1 0.4 1.82 1.08 0.03 2.8 1.14 0.06 5.3 2 4.1 2.26 1.19 0.07 5.9 1.19 0.07 5.9 3 0.8 1.96 1.10 0.04 3.6 1.09 0.02 1.8 4 0.9 1.88 1.12 0.05 4.5 1.15 0.04 3.5 5 2.3 2.34 1.24 0.10 8.1 1.15 0.05 4.3 6 2.4 1.86 1.09 0.04 3.7 1.15 0.02 1.7 7 1.0 1.80 1.16 0.12 10.3 1.40 0.06 4.2 8 3.0 2.19 1.17 0.08 6.8 1.44 0.05 3.5 9 1.4 2.29 1.19 0.10 8.4 1.46 0.05 3.4 Total 16.3 2.14 1.17 0.09 7.7 1.25 0.14 11.2 In Figure 31 the different SOC contents determined by soil sampling and laboratory analysis and estimated by the original (regional) ENVISION PyCaret and TensorFlow SOC maps are presented per soil type at the two Belgian test sites (Table 5Table 6). While soil type had a significant impact (P < 0.01) on the measured SOC content, it’s impact on the estimates of the ENVISION SOC maps was much more limited. The mean measured SOC content for the silt soils (1.18%) was well estimated by the ENVISION PyCaret (1.15%) and TensorFlow (1.14%) SOC maps, but not for the sandy loam soils, where the mean measured SOC content (2.08%) was markedly underestimated by the ENVISION PyCaret (1.23%) and TensorFlow (1.21%) SOC maps. The measured SOC content was also characterized by high variation for both silt (standard deviation: 0.35%) and sandy loam (0.43%) soils, which was not translated by the ENVISION (silt: 0.10%; sandy loam: 0.12%) and TensorFlow (silt: 0.07%; sandy loam: 0.12%) SOC maps. It seems likely that the ENVISION prediction models could be greatly improved by adding data points from distinct soil types in Flanders to the calibration dataset. However, since our results are collected from just two farms, it could also be that differences in management practices play an important role in the differences observed in Figure 31. Deliverable 7.3 55 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 31: SOC content (% - 0-10 cm) at the two Belgian test sites, determined by soil sampling and laboratory analysis (blue) and predicted by the original ENVISION PyCaret (orange) and TensorFlow (green) SOC maps per soil type of the Belgian soil classification A similar story was observed for the drainage classes at the two Belgian test sites (Figure 32). Both ENVISION SOC maps had more difficulty predicting slower draining soils. While the ENVISION SOC maps made a slight overestimation of the “dry” soils, the “moist” and “wet” soils were considerably underestimated. Soil moisture disturbs the spectral signal of the SWIR bands and thereby negatively influences the estimation of SOC content (Castaldi, Chabrillat, Don, et al., 2019; Nocita et al., 2013). Inclusion of ancillary data on soil moisture content (or drainage class), like soil moisture maps that could be derived from Sentinel-1 and -2 spectral data (e.g. Bazzi et al., 2024; El Hajj et al., 2017; Vaudour et al., 2022), could help to increase the accuracy of remote sensing SOC content estimation methods. Texture class Silt Sandy loam Light sandy loam SOC content (%) 0 0,5 1 1,5 2 2,5 3 3,5 Deliverable 7.3 56 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Figure 32: SOC content (% - 0-10 cm) at the two Belgian test sites, determined by soil sampling and laboratory analysis (blue) and predicted by the original ENVISION PyCaret (orange) and TensorFlow (green) SOC maps per drainage class of the Belgian soil classification 4.2.3 Local calibration Figure 33 presents the relationships (calibration lines) between the estimates of the original (regional) ENVISION TensorFlow SOC map and the results from the mapand satellite-based sampling strategies at the farm in Asse and the map-based sampling strategy at the farm in Meise at different sample sizes. Almost all local calibration lines had a steeper slope than the ideal (1:1) calibration line, which highlights the much lower range of the ENVISION TensorFlow SOC map than the SOC content measurements at both Belgian test sites (Table 8Table 11). For the map-based sampling strategy at the farm in Asse, the slope of the calibration lines gradually decreased with increasing sample size. For the map-based sampling strategy at the farm in Meise, the slope of the calibration lines remained roughly constant with increasing sample size. At lower sample sizes (n < 10), there was substantial uncertainty (i.e. variation between replication calibration lines) for the local calibration lines from the map-based sampling strategies. The map-based sampling strategy at the farm in Asse started with a high correlation at a sample size of 3, but with increasing sample size the correlation fell back considerably. For the satellite-based sampling strategy at the farm in Asse, the opposite was true: it started with a low correlation at a sample size of 3, but with the addition of 3 extra sample points (and more) the correlation was doubled. The map-based sampling strategy at the farm in Meise resulted in a similar situation, although with a less striking increase in correlation for the increase in sample size from 3 to 5. At the farm in Meise, the local calibration lines indicated a systematic underestimation of the SOC content within Drainage class dry moist wet SOC content (%) 0 0,5 1 1,5 2 2,5 3 3,5 Deliverable 7.3 57 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 the local range of the ENVISION TensorFlow SOC map (1.0-1.5%). At the farm in Asse the calibration lines of both sampling strategies indicated an overestimation of the low SOC contents and underestimation of the high SOC contents. Figure 33: Local (farm-scale) calibration lines created with increasing sample sizes (n) from the mapand satellite-based calibration datasets for the farm in Asse and the map-based calibration dataset for the farm in Meise Deliverable 7.3 64 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Castaldi, F., Chabrillat, S., & van Wesemael, B. (2019). Sampling strategies for soil property mapping using multispectral Sentinel-2 and hyperspectral EnMAP satellite data. Remote Sensing, 11(3). https://doi.org/10.3390/rs11030309 Castaldi, F., Halil Koparan, M., Wetterlind, J., Žydelis, R., Vinci, I., Özge Savaş, A., Kıvrak, C., Tunçay, T., Volungevičius, J., Obber, S., Ragazzi, F., Malo, D., & Vaudour, E. (2023). Assessing the capability of Sentinel-2 time-series to estimate soil organic carbon and clay content at local scale in croplands. ISPRS Journal of Photogrammetry and Remote Sensing, 199(March), 40–60. https://doi.org/10.1016/j.isprsjprs.2023.03.016 Castaldi, F., Hueni, A., Chabrillat, S., Ward, K., Buttafuoco, G., Bomans, B., Vreys, K., Brell, M., & van Wesemael, B. (2019). Evaluating the capability of the Sentinel 2 data for soil organic carbon prediction in croplands. ISPRS Journal of Photogrammetry and Remote Sensing, 147(October 2018), 267–282. https://doi.org/10.1016/j.isprsjprs.2018.11.026 Chabrillat, S., Ben-Dor, E., Cierniewski, J., Gomez, C., Schmid, T., & van Wesemael, B. (2019). Imaging Spectroscopy for Soil Mapping and Monitoring. In Surveys in Geophysics (Vol. 40, Issue 3). Springer Netherlands. https://doi.org/10.1007/s10712-019-09524-0 Chang, C.-W., Laird, D. A., Mausbach, M. J., & Hurburgh, C. R. (2001). Near-Infrared Reflectance Spectroscopy–Principal Components Regression Analyses of Soil Properties. Soil Science Society of America Journal, 65(2), 480–490. https://doi.org/10.2136/sssaj2001.652480x Commission, E. (2021). Sustainable Carbon Cycling. Communication from the Commission to the European Parliament and the Council, 1–22. https://doi.org/10.1142/9789812791924_0004 Conant, R. T., Smith, G. R., & Paustian, K. (2003). Spatial Variability of Soil Carbon in Forested and Cultivated Sites. Journal of Environmental Quality, 32(1), 278–286. https://doi.org/10.2134/jeq2003.2780 D’Haene, K. (2008). The potential of reduced tillage agriculture in Flanders [Ghent University]. https://doi.org/https://doi.org/1854/10562 De Gruijter, J. J., Bierkens, M. F. P., Brus, D. J., & Knotters, M. (2006). Sampling for Natural Resource Monitoring. Springer. de Gruijter, J. J., McBratney, A. B., Minasny, B., Wheeler, I., Malone, B. P., & Stockmann, U. (2016). Farmscale soil carbon auditing. Geoderma, 265, 120–130. https://doi.org/10.1016/j.geoderma.2015.11.010 De Neve, S., & Hofman, G. (2000). Influence of soil compaction on carbon and nitrogen mineralization of soil Deliverable 7.3 65 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 organic matter and crop residues. Biology and Fertility of Soils, 30(5–6), 544–549. https://doi.org/10.1007/s003740050034 De Vos, B. (2009). Uncertainties of forest soil carbon stock assessment in Flanders [Ghent University]. https://doi.org/10.13140/RG.2.2.14480.56323 Deluz, C., Nussbaum, M., Sauzet, O., Gondret, K., & Boivin, P. (2020). Evaluation of the Potential for Soil Organic Carbon Content Monitoring With Farmers. Frontiers in Environmental Science, 8(July), 1–9. https://doi.org/10.3389/fenvs.2020.00113 Dondeyne, S., Vanierschot, L., Langohr, R., Van Ranst, E., & Deckers, J. A. (2014). The soil map of the Flemish region converted to the 3rd edition of the World Reference Base for soil resources (Issue November). https://doi.org/10.13140/2.1.4381.4089 Dvorakova, K., Heiden, U., & Van Wesemael, B. (2021). Sentinel-2 exposed soil composite for soil organic carbon prediction. Remote Sensing, 13(9), 1–20. https://doi.org/10.3390/rs13091791 El Hajj, M. El, Baghdadi, N., Zribi, M., & Bazzi, H. (2017). Synergic use of Sentinel-1 and Sentinel-2 images for operational soil moisture mapping at high spatial resolution over agricultural areas. Remote Sensing, 9(12), 1–28. https://doi.org/10.3390/rs9121292 Elkerbout, M., Egenhofer, C., Ferrer, J. N., Cătuţi, M., Kustova, I., & Rizos, V. (2020). The European Green Deal after Corona | CEPS. In Ceps - Policy Insights (Vols. 2020–06). https://www.ceps.eu/cepspublications/the-european-green-deal-after-corona/ England, J. R., & Rossel, R. A. V. (2018). Proximal sensing for soil carbon accounting. Soil, 4(2), 101–122. https://doi.org/10.5194/soil-4-101-2018 ESA. (2015). Sentinel-2 User Handbook. In Industrial and Engineering Chemistry (Vol. 1, Issue 2). https://doi.org/10.1021/ie51400a018 FAO. (2019). Measuring and modelling soil carbon stocks and stock changes in livestock production systems. http://www.fao.org/3/I9693EN/i9693en.pdf Garten, C. T., & Wullschleger, S. D. (1999). Soil Carbon Inventories under a Bioenergy Crop (Switchgrass): Measurement Limitations. Journal of Environmental Quality, 28(4), 1359–1365. https://doi.org/10.2134/jeq1999.00472425002800040041x Goidts, E., Van Wesemael, B., & Crucifix, M. (2009). Magnitude and sources of uncertainties in soil organic carbon (SOC) stock assessments at various scales. European Journal of Soil Science, 60(5), 723–739. Deliverable 7.3 66 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 https://doi.org/10.1111/j.1365-2389.2009.01157.x Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031 Guan, K., Jin, Z., Peng, B., Tang, J., DeLucia, E. H., West, P. C., Jiang, C., Wang, S., Kim, T., Zhou, W., Griffis, T., Liu, L., Yang, W. H., Qin, Z., Yang, Q., Margenot, A., Stuchiner, E. R., Kumar, V., Bernacchi, C., … Yang, S. J. (2023). A scalable framework for quantifying field-level agricultural carbon outcomes. Earth-Science Reviews, 243(November 2022), 104462. https://doi.org/10.1016/j.earscirev.2023.104462 Günther, P., Garske, B., Heyl, K., & Ekardt, F. (2024). Carbon farming, overestimated negative emissions and the limits to emissions trading in land-use governance: the EU carbon removal certification proposal. Environmental Sciences Europe, 36(1). https://doi.org/10.1186/s12302-024-00892-y Hengl, T., Rossiter, D. G., & Stein, A. (2003). Soil sampling strategies for spatial prediction by correlation with auxiliary maps. Australian Journal of Soil Research, 41(8), 1403–1422. https://doi.org/10.1071/SR03005 Kätterer, T., Andrén, O., & Persson, J. (2004). The impact of altered management on long-term agricultural soil carbon stocks - A Swedish case study. Nutrient Cycling in Agroecosystems, 70(2), 179–188. https://doi.org/10.1023/B:FRES.0000048481.34439.71 Kennard, R. W., & Stone, L. A. (1969). Computer Aided Design of Experiments. Technometrics, 11(1), 137. https://doi.org/10.2307/1266770 Kerry, R., & Escola, A. (2021). Sensing Approaches for Precision Agriculture. https://doi.org/10.1007/978-3030-78431-7 Köchy, M., Hiederer, R., & Freibauer, A. (2015). Global distribution of soil organic carbon – Part 1: Masses and frequency distributions of SOC stocks for the tropics, permafrost regions, wetlands, and the world. Soil, 1(1), 351–365. https://doi.org/10.5194/soil-1-351-2015 Kotschik, P., Princz, J., Silva, C. de L. e., Renaud, M., Marti-Roura, M., Brooks, B., Pieper, S., Rijk, I., Simini, M., Andres, S., Scholz-Starke, B., & Grenni, P. (2024). The upcoming European Soil Monitoring Law: An effective instrument for the protection of terrestrial ecosystems? Integrated Environmental Assessment and Management, 20(2), 316–321. https://doi.org/10.1002/ieam.4834 Kravchenko, A. N., Robertson, G. P., Hao, X., & Bullock, D. G. (2006). Management practice effects on surface total carbon: Differences in spatial variability patterns. Agronomy Journal, 98(6), 1559–1568. Deliverable 7.3 67 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 https://doi.org/10.2134/agronj2006.0066 Kumhálová, J., Kumhála, F., Kroulík, M., & Matějková, Š. (2011). The impact of topography on soil properties and yield and the effects of weather conditions. Precision Agriculture, 12(6), 813–830. https://doi.org/10.1007/s11119-011-9221-x Lafrance, P., & Banton, O. (1995). Implication of spatial variability of organic carbon on predicting pesticide mobility in soil. Geoderma, 65(3–4), 331–338. https://doi.org/10.1016/0016-7061(94)00051-B Lal, R., Smith, P., Jungkunst, H. F., Mitsch, W. J., Lehmann, J., Ramachandran Nair, P. K., McBratney, A. B., De Moraes Sá, J. C., Schneider, J., Zinn, Y. L., Skorupa, A. L. A., Zhang, H. L., Minasny, B., Srinivasrao, C., & Ravindranath, N. H. (2018). The carbon sequestration potential of terrestrial ecosystems. Journal of Soil and Water Conservation, 73(6), 145–152. https://doi.org/10.2489/jswc.73.6.145A Lettens, S., De Vos, B., & D’Hose, T. (2021). Verkennende analyse van binnen-perceels-variatie in bodem koolstof en implicaties voor (her)bemonstering. http://www.fedict.belgium.be/nl/over_fedict/nieuwsberichten/fedivers/fedivers_nov_2014/Item5 Malone, B., Hedley, C., Roudier, P., Minasny, B., Jones, E., & McBratney, A. (2018). Auditing on-farm soil carbon stocks using downscaled national mapping products: Examples from Australia and New Zealand. Geoderma Regional, 13(December 2017), 1–14. https://doi.org/10.1016/j.geodrs.2018.02.002 Manies, K. L., Harden, J. W., Kramer, L., & Parton, W. J. (2001). Carbon dynamics within agricultural and native sites in the loess region of Western lowa. Global Change Biology, 7(5), 545–555. https://doi.org/10.1046/j.1354-1013.2001.00427.x Minasny, B., Malone, B. P., McBratney, A. B., Angers, D. A., Arrouays, D., Chambers, A., Chaplot, V., Chen, Z. S., Cheng, K., Das, B. S., Field, D. J., Gimona, A., Hedley, C. B., Hong, S. Y., Mandal, B., Marchant, B. P., Martin, M., McConkey, B. G., Mulder, V. L., … Winowiecki, L. (2017). Soil carbon 4 per mille. Geoderma, 292, 59–86. https://doi.org/10.1016/j.geoderma.2017.01.002 Minasny, B., & McBratney, A. B. (2006). A conditioned Latin hypercube method for sampling in the presence of ancillary information. Computers and Geosciences, 32(9), 1378–1388. https://doi.org/10.1016/j.cageo.2005.12.009 Moeskops, B. (2010). Biochemical and microbial indicators of soil quality in contrasting agroecosystems. Ghent University. Montagne, D., Cornu, S., Bourennane, H., Baize, D., Ratié, C., & King, D. (2007). Effect of agricultural practices Deliverable 7.3 68 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 on trace-element distribution in soil. Communications in Soil Science and Plant Analysis, 38(3–4), 473– 491. https://doi.org/10.1080/00103620601174411 Mulla, D. J. (1993). Mapping and Managing Spatial Patterns In Soil Fertility and Crop Yield. Soil Specific Crop Management, 15, 15–26. https://doi.org/10.2134/1993.soilspecificcrop.c2 Nayak, A. K., Rahman, M. M., Naidu, R., Dhal, B., Swain, C. K., Nayak, A. D., Tripathi, R., Shahid, M., Islam, M. R., & Pathak, H. (2019). Current and emerging methodologies for estimating carbon sequestration in agricultural soils: A review. Science of the Total Environment, 665, 890–912. https://doi.org/10.1016/j.scitotenv.2019.02.125 Nocita, M., Stevens, A., Noon, C., & Van Wesemael, B. (2013). Prediction of soil organic carbon for different levels of soil moisture using Vis-NIR spectroscopy. Geoderma, 199, 37–42. https://doi.org/10.1016/j.geoderma.2012.07.020 Nolin, M. C., Guertin, S. P., & Wang, C. (1996). Within-Field Spatial Variability of Soil Nutrients and Corn Yield in a Montreal Lowlands Clay Soil (pp. 257–270). https://doi.org/10.2134/1996.precisionagproc3.c26 Oldfield, E. E., Eagle, A. J., Rublin, R. L., Rudek, J., Sanderman, J., & Gordon, D. R. (2021). Agricultural Soil Carbon Credits - Making sense of protocols for carbon sequestration and net greenhouse gas removals. https://www.science.org/doi/pdf/10.1126/science.abl7991?kasa_token Oorts, K. (2006). Effect of tillage systems on soil organic matter stocks and C and N fluxes in cereal cropping systems on a silt loam soil in Northern France. Katholieke Universiteit Leuven. Oorts, K., Amery, F., Lettens, S., D’Hose, T., De Vos, B., Luts, D., Salomez, J., & Swerts, M. (2023). Bodemkoolstofmonitoringnetwerk Cmon : Jaarrapport werkjaar 2 (1 juli 2022 - 30 juni 2023). 2, 54. https://archief.algemeen.omgeving.vlaanderen.be/xmlui/bitstream/handle/acd/915022/Cmon_jaarra pport_2_2022-2023.pdf Patzold, S., Mertens, F. M., Bornemann, L., Koleczek, B., Franke, J., Feilhauer, H., & Welp, G. (2008). Soil heterogeneity at the field scale: A challenge for precision crop protection. Precision Agriculture, 9(6), 367–390. https://doi.org/10.1007/s11119-008-9077-x Paustian, K., Lehmann, J., Ogle, S., Reay, D., Robertson, G. P., & Smith, P. (2016). Climate-smart soils. Nature, 532(7597), 49–57. https://doi.org/10.1038/nature17174 Peltoniemi, M., Heikkinen, J., & Mäkipää, R. (2007). Stratification of regional sampling by model-predicted changes of carbon stocks in forested mineral soils. Silva Fennica, 41(3), 527–539. Deliverable 7.3 69 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 https://doi.org/10.14214/sf.287 Piikki, K., Wetterlind, J., Söderström, M., & Stenberg, B. (2021). Perspectives on validation in digital soil mapping of continuous attributes—A review. Soil Use and Management, 37(1), 7–21. https://doi.org/10.1111/sum.12694 Piotrowska, A., Długosz, J., Namysłowska-Wilczyńska, B., & Zamorski, R. (2011). Field-scale variability of topsoil dehydrogenase and cellulase activities as affected by variability of some physico-chemical properties. Biology and Fertility of Soils, 47(1), 101–109. https://doi.org/10.1007/s00374-010-0507-3 Poeplau, C., Prietz, R., & Don, A. (2022). Plot-scale variability of organic carbon in temperate agricultural soils—Implications for soil monitoring#. Journal of Plant Nutrition and Soil Science, 185(3), 403–416. https://doi.org/10.1002/jpln.202100393 Poggio, L., De Sousa, L. M., Batjes, N. H., Heuvelink, G. B. M., Kempen, B., Ribeiro, E., & Rossiter, D. (2021). SoilGrids 2.0: Producing soil information for the globe with quantified spatial uncertainty. Soil, 7(1), 217–240. https://doi.org/10.5194/soil-7-217-2021 Post, W. M., Izaurralde, R. C., Mann, L. K., & Bliss, N. (2001). Monitoring and verifying changes of organic carbon in soil. In CDIAC Communications (Issue 27). https://doi.org/10.1023/A:1017514802028 Potash, E., Guan, K., Margenot, A., Lee, D. K., DeLucia, E., Wang, S., & Jang, C. (2022). How to estimate soil organic carbon stocks of agricultural fields? perspectives using ex-ante evaluation. Geoderma, 411(September 2021), 115693. https://doi.org/10.1016/j.geoderma.2021.115693 R Core Team. (2024). R: A language and Environment for Statistical Computing. R Foundation for Statistical Computing. www.R-projects.org Robertson, G. P., Crum, J. R., & Ellis, B. G. (1993). The spatial variability of soil resources following long-term disturbance. Oecologia, 96(4), 451–456. https://doi.org/10.1007/BF00320501 Röver, M., & Kaiser, E. A. (1999). Spatial heterogeneity within the plough layer: Low and moderate variability of soil properties. Soil Biology and Biochemistry, 31(2), 175–187. https://doi.org/10.1016/S00380717(97)00272-1 Saby, N. P. A., Bellamy, P. H., Morvan, X., Arrouays, D., Jones, R. J. A., Verheijen, F. G. A., Kibblewhite, M. G., Verdoodt, A., Üveges, J. B., Freudenschuß, A., & Simota, C. (2008). Will European soil-monitoring networks be able to detect changes in topsoil organic carbon content? Global Change Biology, 14(10), 2432–2442. https://doi.org/10.1111/j.1365-2486.2008.01658.x Deliverable 7.3 70 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Sanderman, J., Hengl, T., & Fiske, G. J. (2017). Soil carbon debt of 12,000 years of human land use. Proceedings of the National Academy of Sciences of the United States of America, 114(36), 9575–9580. https://doi.org/10.1073/pnas.1706103114 Schöning, I., Totsche, K. U., & Kögel-Knabner, I. (2006). Small scale spatial variability of organic carbon stocks in litter and solum of a forested Luvisol. Geoderma, 136(3–4), 631–642. https://doi.org/10.1016/j.geoderma.2006.04.023 Schrumpf, M., Schulze, E. D., Kaiser, K., & Schumacher, J. (2011). How accurately can soil organic carbon stocks and stock changes be quantified by soil inventories? Biogeosciences, 8(5), 1193–1212. https://doi.org/10.5194/bg-8-1193-2011 Shao, S., Zhang, H., Fan, M., Su, B., Wu, J., Zhang, M., Yang, L., & Gao, C. (2021). Spatial variability-based sample size allocation for stratified sampling. Catena, 206(May), 105509. https://doi.org/10.1016/j.catena.2021.105509 Shukla, M. K., Slater, B. K., Lal, R., & Cepuder, P. (2004). Spatial variability of soil properties and potential management classification of a chernozemic field in lower Austria. Soil Science, 169(12), 852–860. https://doi.org/10.1097/00010694-200412000-00004 Singh, K., Murphy, B. W., & Marchant, B. P. (2012). Towards cost-effective estimation of soil carbon stocks at the field scale. Soil Research, 50(8), 672–684. https://doi.org/10.1071/SR12119 Sleutel, S., D’Hose, T., Lettens, S., Ruysschaert, G., & De Vos, B. (2021). Actualisatie en verfijning van de onderbouwing van een methodiek voor de systematische monitoring van koolstofvoorraden in de bodem (opdracht OMG/VPO/BODEM/TWOL/2017/1) - Eindrapport. Sleutel, S., De Vijver, E., Moeskops, B., Bouckaert, L., Ameloot, N., De Bolle, S., Van Meirvenne, M., & De Neve, S. (2011). Onderbouwing van een methodiek voor de systematische monitorring van koolstofvoorraden in lanndbouwbodems (BOD/STUD/2010/05). http://hdl.handle.net/1854/LU2075996 Sleutel, S., Leinweber, P., Begum, S. A., Kader, M. A., Van Oostveldt, P., & De Neve, S. (2008). Composition of organic matter in sandy relict and cultivated heathlands as examined by pyrolysis-field ionization MS. Biogeochemistry, 89(2), 253–271. https://doi.org/10.1007/s10533-008-9217-4 Smith, P. (2004). How long before a change in soil organic carbon can be detected? Global Change Biology, 10(11), 1878–1883. https://doi.org/10.1111/j.1365-2486.2004.00854.x Deliverable 7.3 71 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Smith, P., Soussana, J. F., Angers, D., Schipper, L., Chenu, C., Rasse, D. P., Batjes, N. H., van Egmond, F., McNeill, S., Kuhnert, M., Arias-Navarro, C., Olesen, J. E., Chirinda, N., Fornara, D., Wollenberg, E., ÁlvaroFuentes, J., Sanz-Cobena, A., & Klumpp, K. (2020). How to measure, report and verify soil carbon change to realize the potential of soil carbon sequestration for atmospheric greenhouse gas removal. Global Change Biology, 26(1), 219–241. https://doi.org/10.1111/gcb.14815 Spencer, S., Ogle, S. M., Breidt, F. J., Goebel, J. J., & Paustian, K. (2011). Designing a national soil carbon monitoring network to support climate change policy: a case example for US agricultural lands. Greenhouse Gas Measurement and Management, 1(3–4), 167–178. https://doi.org/10.1080/20430779.2011.637696 Stanley, P., Spertus, J., Chiartas, J., Stark, P. B., & Bowles, T. (2023). Valid inferences about soil carbon in heterogeneous landscapes. Geoderma, 430(December 2022), 116323. https://doi.org/10.1016/j.geoderma.2022.116323 Stevens, A., & Ramirez-Lopez, L. (2024). Package ‘ prospectr .’ https://github.com/l-ramirez-lopez/prospectr Sunoj, S., Kharel, D., Kharel, T., Cho, J., Czymmek, K. J., & Ketterings, Q. M. (2021). Impact of headland area on whole field and farm corn silage and grain yield. Agronomy Journal, 113(1), 147–158. https://doi.org/10.1002/agj2.20489 Tits, M., Lorenz, W., Grootjans, A., P., J., A., E., S., D., J., B., & H., V. (2024). Bodemvruchtbaarheid van de Akkerbouwen Weilandpercelen in België en Noordelijk Frankrijk (2020-2023). Tsardanidis, I., Drivas, A., Sitokonstantinou, V., Tsouni, A., Tsioutsia, I., & Ilias, P. (2024). ENVISION - D3.7 data products final report. 1–134. https://envision-h2020.eu/wp-content/uploads/2024/03/D3.7-Dataproducts-final-report_F2.0.pdf Tziolas, N., Tsakiridis, N., Chabrillat, S., Demattê, J. A. M., Ben-Dor, E., Gholizadeh, A., Zalidis, G., & van Wesemael, B. (2021). Earth observation data-driven cropland soil monitoring: A review. Remote Sensing, 13(21). https://doi.org/10.3390/rs13214439 van Belle, G. (2008). Statistical Rules of Thumb, 2nd Edition. Wiley-Interscience. van der Schans, D., & van der Berg, W. (2013). Testen, Validatie en Toepassing van het Veris-sensorplatform: Veldanalyse van twee percelen op veenkoloniale grond (Issue 540). https://edepot.wur.nl/259239 van der Voort, T. S., Verweij, S., Fujita, Y., & Ros, G. H. (2023). Enabling soil carbon farming: presentation of a robust, affordable, and scalable method for soil carbon stock assessment. Agronomy for Sustainable Deliverable 7.3 72 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N° 862695 Development, 43(1). https://doi.org/10.1007/s13593-022-00856-7 Van Ranst, E., & Sys, C. (2000). Eenduidige legende voor de digitale bodemkaart van Vlaanderen (Schaal 1:20 000). In Laboratorium voor Bodemkunde (Issue April). Vanguelova, E. I., Bonifacio, E., De Vos, B., Hoosbeek, M. R., Berger, T. W., Vesterdal, L., Armolaitis, K., Celi, L., Dinca, L., Kjønaas, O. J., Pavlenda, P., Pumpanen, J., Püttsepp, Reidy, B., Simončič, P., Tobin, B., & Zhiyanski, M. (2016). Sources of errors and uncertainties in the assessment of forest soil carbon stocks at different scales—review and recommendations. Environmental Monitoring and Assessment, 188(11). https://doi.org/10.1007/s10661-016-5608-5 Vaudour, E., Gholizadeh, A., Castaldi, F., Saberioon, M., Borůvka, L., Urbina-Salazar, D., Fouad, Y., Arrouays, D., Richer-De-forges, A. C., Biney, J., Wetterlind, J., & Van Wesemael, B. (2022). Satellite Imagery to Map Topsoil Organic Carbon Content over Cultivated Areas: An Overview. Remote Sensing, 14(12), 1–22. https://doi.org/10.3390/rs14122917 Viscarra Rossel, R. A., & Brus, D. J. (2018). The cost-efficiency and reliability of two methods for soil organic C accounting. Land Degradation and Development, 29(3), 506–520. https://doi.org/10.1002/ldr.2887 Welch, R. Y., Behnke, G. D., Davis, A. S., Masiunas, J., & Villamil, M. B. (2016). Using cover crops in headlands of organic grain farms: Effects on soil properties, weeds and crop yields. Agriculture, Ecosystems and Environment, 216, 322–332. https://doi.org/10.1016/j.agee.2015.10.014 WRB, I. W. G. (2015). World Reference Base for Soil Resources 2014, update 2015 International soil classification system for naming soils and creating legends for soil maps. World Soil Resources Reports No. 106.