1. Introduction
The European Copernicus program operates its own constellation of Earth observation satellites and produces an immense amount of satellite data. Together with the fast development of cloud processing platform services, it gave rise to an impulse to exploit these resources and technology advances within the EU’s Common Agriculture Policy (CAP).
A substantial element of the CAP is a system of farmer subsidies granted to ensure income stability and to remunerate them for environmentally friendly farming and delivering public goods not normally paid for by the markets, such as taking care of the countryside (for instance flowering buffer strips for pollination) [
1]. However, this farmers’ direct support is ruled by eligibility criteria, commitments, and other obligations under the CAP regulations. Compliance with these rules has so far been checked by administrative and on-the-spot checks (OTSC), a sample approach practice performed yearly by the EU Member States on at least 5% of aid applications [
2]. The OTSC are performed by either classical field-based visits or checks via satellite imagery, called the control with remote sensing (CwRS) method. The CwRS uses a sequence of satellite images acquired during the crop season over the same selected sample of parcels. The sequence is usually composed of 2–4 data captures, from which one is of sub-meter resolution (for area measurement), complemented with images of slightly lower spatial resolution (for detecting field conditions over the season). A separate photointerpretation analysis of each image is the result at the end of the crop season as the conclusion for all sampled holdings. Satellite images have been used for the purpose of CAP control since 1993, and until now the number of acquired images has grown from 100 to 1400 captures for one control year. Both the amount of data and the quality (radiometric, spectral, and spatial resolution) of the satellite imagery have improved considerably. The list of various satellites used during the years includes: European Remote Sensing (ERS) satellite, Formosat 2, Radarsat, Landsat 5, Disaster Monitoring Constellation 2 (DMC 2), ResourceSat-1, RapidEye, THEOS, Cartosat 2, Deimos 1, CosmoSkymed, EROS A/B, Ikonos, QuickBird, Pleiades 1A/1B, SPOT 4/5, WorldView-4, PlanetScope, Kompsat-3/3A, SPOT 6/7, GeoEye-1, WorldView-1/2/3, last nine are currently used.
In 2018, the European Commission (EC) provided a legal and technical framework to exploit the advantages of Copernicus satellite data in the context of CAP controls. An amendment to Commission Implementing Regulation (EU) No. 809/2014 [
3] gave the EU Member States an option to gradually replace the OTSC with checks by monitoring (CbM). This introduced a monitoring and prevention system that mainly relies on continuous provision of Sentinel satellite data. The free availability of global coverage and temporal resolution of the data enabled a shift from the sampling and sequence mode to the continuous assessment of the whole territory. Such constant assessment provides an individual time for the decision on the eligibility for each holding, which could make financial procedures more flexible in comparison to the OTSC approach. Furthermore, the checks mechanism is coupled with an alert component. Farmers receive early warnings and can correct their aid application to avoid noncompliant declarations, avoiding penalties at the end of the season. In practice, this involves automatic image analysis to determine the actual state of the field. This gives information either about the presence of a crop, its phenological stage, or about a specific activity conducted in a given time range. The accumulated information about the state of the land is continuously compared with the expected scenario of events implied by the obligations and commitments of the specific payment scheme [
4].
As the CbM approach is based upon the use of Copernicus Sentinel data (Sentinel-1/2), its application scope is limited by the radiometric and geometric characteristics of these satellites. While the revisit time of Sentinel satellites provides an unprecedented temporal resolution, the ground sampling distance (GSD) and geometric positional accuracy could be too coarse for smaller spatial features. The CbM faces a tradeoff between the large amount of free data and its medium spatial resolution. Current OTSC image data have a spatial resolution of no more than 2.2 m for a panchromatic band. As a result, one of the major concerns of the EU Member States regarding the operational implementation of monitoring is the perceived inadequacy of the Sentinel sources for small fields. Cases where conclusion cannot be made regarding the eligibility of Sentinel imagery due to the size of the field (or other observed features) require alternative data sources, such as a higher resolution data stack, geotagged photographs, or field inspections. Inevitably, these alternative sources and processes have a higher cost than the automated processing of free Sentinel data, so managing them is essential.
This study addresses the concern of EU Member States regarding the monitoring of the small fields. It specifically focuses on Sentinel-2, i.e., optical data. The goals of the research are: (i) to quantify the capability of Sentinel-2 to provide information on the activity or state of a field; (ii) to examine the impacts of various geospatial parameters of the field on the suitability of Sentinel-2 data; (iii) to estimate the limiting geospatial criteria above which the Sentinel-2 data stacks still provide reliable information on the state or activity of the field, all in the context of the CbM of small parcels. A similarity assessment of the two sets of NDVI time series (Sentinel-2 and PlanetScope) extracted over the same fields is used as the main tool to achieve the research goals.
5. Discussion
For the purpose of this study, a set of normalized difference vegetation index time series from Sentinel-2 calculated over a number of fields with various geospatial parameters was compared with a set of NDVI time series from a higher resolution satellite (i.e., PlanetScope) extracted for the same group of fields. A similarity assessment of the two sets of NDVI time series was used as the main tool to reach the research goals. The latest studies dealing with image data time series in the agriculture domain mainly address the issue of understanding and monitoring of the crop phenology [
30,
31,
32], focusing on cropland mask construction [
33,
34] and land cover classification [
11,
35]. The use of time series to assess cropping intensity became an actual topic due to its importance as a critical input data variable for many climate land surface and crop models [
36,
37]. The most recent research studies in the context of the CAP exploit the Sentinel-2 data time series to detect agriculture land use anomalies [
38] or map the grassland use intensity [
39]. The paramount topic in the agriculture sector is, however, crop identification. Previously, [
40] presented a first experience with crop type classification using one of the first acquired (preoperational) Sentinel-2 images. The benefits of the subsequently available time series were explored in [
41], using a random forest (RF) classifier for crop mapping purposes. This study confirmed that multitemporal information increases the crop type classification accuracy. The benchmark for various classification methods on simulated Sentinel-2 time series was described in [
42]. Further research on real multitemporal Sentinel-2 data [
43] compared the performance of supervised machine learning techniques (support vector machines (SVMs) and RF) with respect to their accuracy and execution time, highlighting the superior performance of Sentinel-2 for this type of application when compared to Lansat-8 OLI. Additionally, [
44] showed that the SVM and RF outperformed the traditional statistical maximum likelihood classification method. Since the temporal resolution of Sentinel-2 is constrained by cloud cover conditions, [
45,
46,
47,
48] explored multisource data (Sentinel-1–Landsat-8) use to complement Sentinel-2 time series and increase the crop type classification accuracy.
An operational solution for multitemporal crop classification approaches at a national scale was developed by European Space Agency's Sentinel-2 for Agriculture project (Sen2Agri). Sen2Agri delivered an open-source system that derives 4 basic products from Sentinel-2 and Landsat 8 time series: cloud-free composites, dynamic cropland masks, crop type maps, and vegetation status indicators. Additionally, [
49] described the overall approach and key principles embedded in the processing chain and presented the results from an independent validation of the monthly cropland masks and crop type maps. However, the CbM approach, as described in [
4], is primarily based on the detection of particular phenomena (“markers” and behavior) to continuously monitor a presumed agricultural business scenario, allowing early feedback to be issued to individual farmers. Crop mapping is not foreseen as a key component, but crop identification remains a valid and relevant ancillary instrument. Member States adopting CbM use the crop identification and classification algorithms for various reasons.
Regarding the topic of monitoring small fields, recent research [
50] dealt with assessments on the impact of sensor pixel sizes for EU CAP. However, the study limited its consideration to a simple assumption that for fields with no pure pixels inside their boundaries, no CAP services can be provided, mainly pointing to crop determination. Our study presents the application of the comparative spatiotemporal time series analysis to assess the limits of satellite data regarding the smallest traced entity in the context of CbM, taking into account its full complexity.
The potential of Sentinel-2 data for the state assessment and the event detection on small parcels, as presented in
Section 4.2, was found to be better than expected. The evaluation was performed over cloud-free pixels with a sufficient number of observations, and therefore the constraint for small parcels in the agricultural regions very frequently covered by clouds during the growing season, as mentioned for instance in [
49], was not considered.
Concerning the visual assessment, a difference (0.43 Cohen’s kappa) between the operators was observed. Even if according to [
26] this might still be assessed as a moderate agreement, other researchers studies [
51] considered this as a low degree of the agreement. While one operator (A) evaluated the NDVI time series similarly, and thus provided the same information content for 87% of parcels, the other operator (B) reached a level of 73% from the total amount. In comparison with the statistical approach, which resulted in 90% of agreements, there was visible conformity with operator A, whilst operator B most likely adopted stricter evaluation criteria. This confirms that in this case, the visual assessment was a subjective methodology strongly influenced by experience and personal perception. As such, we considered it as a supportive element for further analyses.
Regarding the PlanetScope NDVI time series (as a reference for our analysis), based on expert judgment it was concluded that for 2% there were doubts about the correctness due to the weak trend of the profiles. No dependency on the parcel size was found. For 80% of the parcels, the profiles of both sensors were assessed as not being similar. Additionally, 60% of these parcels were below our limiting criteria (<8 pixels and >60% pixels lost). After detailed analysis, we could observe that 65% of these parcels had elongated geometry, measuring 4–15 m wide. Parcels with such a shape are generally vulnerable, especially if the long side is perpendicular to the eventual shift. For the proper decision on the correctness, we would, however, need to have collected evidence during the crop season on the field.
From the investigation of the contextual parameters of the parcels, there is no clear evidence that the type of arable crop in the agriculture parcel would have an impact on the similarity of the NDVI time series pairs. The research was, however, restricted to arable crops only. This restriction is caused by the chosen statistical evaluation methodology of similarity using the correlation coefficient. The correlation coefficient uses averages as a pivot in order to measure the linearity between each NDVI time series pair. Thus, it assumes that the times series have marked trends over time. By contrast, parcels where slight or no trend is expected (e.g., permanent grazed grassland or permanent crops) exhibit time series with an almost constant value, and few fluctuations can be considered as noise. Hence, applying this methodology on these crops is not appropriate. Although the arable crop itself does not prove to have an impact on the similarity, the surface surrounding the evaluated parcel seems to be a significant factor. The results obtained from the study showed a higher rate of positive similarity status for parcels having the same crops in the surroundings. As a limitation of this result, we consider the fact that the algorithm used for the decision did not consider the amount of surrounding crops. In this respect, the resulting number of parcels with the same crop in the surrounding could be exaggerated (i.e., biased). We, therefore, assume that the positive influence of having the same crop in the surroundings could be even higher than the one shown in
Figure 8.
Further analyses on the relationship between the correlation coefficient and quantitative parameters identified two parameters that were more relevant than the others; a number of clean or full pixels and a share of pixels lost after applying half of the negative pixel buffer. The possible explanation for that is the importance of the shape and position of the parcel on the raster. As outlined in
Section 2.4, the parameter expressing the loss of pixels took into account both the shape and position, and emerged as a very suitable descriptor of the parcel. By contrast, the size of the parcel was revealed to be a vague description, with an apparent lack of a strong correlation.
Working further with the three selected parameters, i.e., surrounding the parcel, the ratio of Sentinel-2 pixels lost, and the number of full pixels after application of the buffer, we built a model displaying the possible impacts of their relationships, and combinations on the probability of the estimated correlation coefficient. From a priori defined thresholds, we derived the limiting parameters of the parcel geometry, according to which we could estimate and quantify the proportion of parcels (estimated group 2) with a higher risk of Sentinel-2 data failure to provide conclusive information about the parcel. Looking at
Figure 8, one might theoretically question the stability of the model, as the tested sample does not cover all possible combinations of parameters. However, some of these combinations or dimensions are simply not realistic in the field.
The results in
Table 6 show the proportions of parcels assessed as deviating in the similarity assessment exercise (group 1) for both groups of isolated parcels below (i.e., estimated group 2) and above the limiting parameters (estimated group 3). The proportions demonstrate the correctness and validity of the proposed limiting criteria.
As indicated earlier in
Section 2.2, the geolocation accuracy of Sentinel-2B was described as unstable during the tested period. We, therefore, investigated a representation of Sentinel-2B in the NDVI time series sample and the proportion of Sentinel-2B in the NDVI profiles of parcels that deviated (group 1,
Section 4.2). For both cases, the participation of Sentinel-2B amounted to approximately 40%. We did not observe any higher occurrence of Sentinel-2B values in those NDVI time series that did not pass the test of similarity. We concluded that Sentinel-2B instability did not impact the results of this study. A positive aspect of the applied approach might be the fact that the curve matching method applied to compare the similarity of NDVI profiles was not highly sensitive to a single date outlier, unless it was a periodical or longer duration phenomenon.
As mentioned in the introduction section, the quality of the produced time series has a direct impact on the similarity assessment results. The cloud cover mask and the co-registration accuracy leave much room for improvement and we assume that the potential of Sentinel-2 imagery will be even higher in the near future, for instance after the introduction of a global reference image [
7].
6. Conclusions
The introduction of the checks by monitoring approach in the framework of the European Common Agriculture Policy gave an impulse to explore the ability of Sentinel-2 data to detect a parcel’s behavior in terms of the geometry of that parcel. The research demonstrated that the size of the parcel is not an appropriate measure for the evaluation of Sentinel-2 data suitability for CbM. The number of full pixels and the ratio of pixels lost after the application of a half-pixel-wide negative buffer emerged as more appropriate parameters. Our findings also suggest that the surrounding land use or land cover is a relevant factor that needs to be duly taken into account.
For the majority of tested parcels, the Sentinel-2 data provided the same or similar information about the state and the activity on the field as PlanetScope imagery. Considering the deliberate small size of the parcels in the studied sample, the results proved the high potential of Sentinel-2 imagery. Thus, for the checks by monitoring, the study provided evidence of a good design strategy involving the trade-off between the temporal and spatial resolution of the Sentinel-2 satellite.
The limiting criteria expressed by geospatial parcel parameters serve to identify a group of “vulnerable” agricultural parcels for which there is a high probability that Sentinel-2 might not be a suitable satellite to assess their state and related agricultural activity. Member states, therefore, might adopt the definition of these spatial limits when designing their monitoring approaches.
In our research, we applied the comparison technique on the time series pairs retrieved from two image datasets of different spatial resolution to estimate the potential and spatial limitations of the coarser one. This innovative approach, as well as the whole evaluation strategy with thresholds adapted to the specific needs, could be used for a similar assessment of any other satellite data and application purpose.