Mapping Ecological Focus Areas within the EU CAP Controls Framework by Copernicus Sentinel-2 Data

Greening is a Common Agricultural Policy (CAP) subsidy that ensures that all EU farmers receiving income support produce climate and environmental benefits as part of their farming activities. To receive greening support, it is mandatory for the farmer to carry out three agricultural practices that are considered environmentally and climate friendly: (a) crop diversification; (b) maintenance of permanent meadows and pastures; and (c) presence of an Ecological Focus Area (EFA). Contributions are delivered and monitored by paying agencies (PP) that ordinarily perform administrative checks and spot checks. The latter are provided through photo-interpretation of high-resolution satellite or aerial images and, in specific cases, through local ground checks (GC) as well. In this work, stimulated by the Piemonte Regional Agency for Payments in Agriculture (ARPEA), a prototype service to support PPs’ controls within the greening CAP framework was proposed with special concern for EFA detection. The proposed approach is expected to represent a valid alternative or supporting tool for GC. It relies on the analysis of NDVI time series derived from Copernicus Sentinel-2 data. The study was conducted in the provinces of Turin, Asti and Vercelli within the Piedmont Region (NW Italy), and over 12,500 EFA fields were assessed. Since the recent National Report No. 5465 stipulates that mowing and any other soil management operation is prohibited on set-aside land designated as an EFA during the reference period (RP) between 1st March and 30th June, a time series (TS) of NDVI in the same period was generated. Once averaged at plot level, NDVI trends were modelled by a first-order polynomial, and the correspondent statistics (namely, R2, MAE and maximum residual) was computed. These were assumed to play the role of discriminants in EFA detection based on a thresholding approach (Otsu’s method), calibrated with reference to the training dataset. The threshold satisfaction was therefore tested, and, depending on the number of satisfied thresholds out of the possible three, EFA and non-EFA plots were detected with a different degree of reliability. The correspondent EFA map was generated for the area of interest and validated according to GCs as provided by the ARPEA. The results showed an overall accuracy of 84%, indicating that the approach is promising. The authors retain that this procedure represents a valid alternative (or integrating) tool for ground controls by PPs.


Ecological Focus Areas within CAP
Common Agricultural Policy (CAP) is one of the most important EU strategies, receiving around 35% of the European Union (EU) budget. CAP aims to support farmers with contributions to ensure safe and quality food at affordable prices; to protect the farming sector from excessive price volatility, market crises and imbalances affecting the food supply chain; to maintain rural communities across the EU by preserving the environment, animal welfare and biodiversity; to mitigate climate change effects in need of sustainable applications, aimed at verifying the eligibility of an application and the properness of the supplied documentation, as well as (b) spot controls (SC) that are required for a small subset (5%) of received applications to verify the truthfulness of the submitted information and to test the compliance of farming companies with CAP requirements. Controlled fields are selected according to risk (80%) and randomness (20%) criteria. Generally, SCs are achieved via photo-interpretation of very high-resolution satellite or aerial images; only in particular cases are ground controls (GC) operated.
The recent Italian National Report No. 5465, which applies EU Regulation No. 1307/2013 of the European Parliament and Council, references that mowing and any other soil management operation is prohibited on set-aside land designated as an EFA during RP. Currently, EFA controls can only be achieved through GCs. Being that GCs are highly expensive for PPs, they can only be performed on a small subset of the GSAA, leaving a high probability that false applications remain unknown. In this framework, S2 data are expected to play an important and improving role; they appear to be a valid and operational tool for supporting EFA controls.
It is worthwhile to remember that EFA definition does not depend on crop type; conversely, EFA fields can be, practically, assumed similar to fallow fields since, during the growing season, farmers cannot manage them. Consequently, EFA phenological behaviour is expected to show a progressively increasing NDVI profile with no significant variations due to possible agronomic actions, such as mowing, sowing or harvesting [42]. The period of the year in which this trend must be tested depends on both agronomic calendars and national regulations that, partially, already heed local agronomic calendars, themselves.
Presently, the literature appears to lack works that deal with EFA mapping by satellite in the CAP controls framework, making this work desirable.

Goals
According to the above-mentioned scenario, in this paper, a prototype procedure based on S2 imagery is proposed to support the EFA controls in Piemonte (NW Italy). It was developed within the framework of an institutional collaboration between the GEO4Agri Lab of the Department of Agricultural, Forest and Food Sciences (DISAFA) of the University of Torino and the Regional Agency for Payments in Agriculture (ARPEA).
The ARPEA explicitly required that the approach was compliant with actual operational scenarios and the available skills that an ordinary regional technician is expected to possess. In other words, it required that the eventual tool for EFA detection was simple enough and explicit in its agronomical meaning, in order to guarantee a conscious utilization and a critical approach to EFA controls. The problem of EFA detection is a typical land cover classification one that is expected to be faced when using satellite data [43]. The above-mentioned requirements from the ARPEA drove the authors to exclude adoption of  [44,45] and opt for a simpleR rule-based approach, where defined rules could be immediately agronomically interpreted. Moving towards MDL would make it difficult to interpret eventual classification criticalities and, consequently, to properly adjust the algorithm parameters that, in MDL, cannot be related to tangible factors close to the knowledge domain to which the ideal user of the system would belong. Consequently, the proposed approach aims to detect an EFA by applying simple rules, or tasks, to disaggregate a complex problem into a succession of simple and more controllable ones, whereby domain knowledge plays a crucial role in improving classification results.
In this paper, with reference to the study area (see next section), we propose a classification method for EFA detection based on the at-field level analysis of the local NDVI profile from Sentinel-2 imagery. With reference to auxiliary data and ground observations from the ARPEA, we defined simple rules to apply to proper statistics, obtainable from the local NDVI profile. The proposed classification process proceeds by testing the satisfaction of some conditions, which makes it possible to recognize EFAs with acceptable accuracy and, in the meantime, assign a degree of reliability of detection. After validating the results, a map of EFAs was generated for the 2021 year.

Study Area
The study area (AOI) includes 3 administrative provinces of the Piemonte region (NW -Italy): namely, Torino, Asti and the southern part of Vercelli ( Figure 1). The AOI is located at an altitude ranging between 200 and 300 m a.s.l., covering about 9636 km 2 . It is characterized by a highly diversified agricultural context, hosting many EFAs. This made it particularly attractive to the ARPEA, as it is called yearly to control the truthfulness of the declared EFAs.
criticalities and, consequently, to properly adjust the algorithm parameter cannot be related to tangible factors close to the knowledge domain to w user of the system would belong. Consequently, the proposed approach ai EFA by applying simple rules, or tasks, to disaggregate a complex problem sion of simple and more controllable ones, whereby domain knowledge role in improving classification results.
In this paper, with reference to the study area (see next section), we p fication method for EFA detection based on the at-field level analysis of profile from Sentinel-2 imagery. With reference to auxiliary data and groun from the ARPEA, we defined simple rules to apply to proper statistics, o the local NDVI profile. The proposed classification process proceeds by te faction of some conditions, which makes it possible to recognize EFAs w accuracy and, in the meantime, assign a degree of reliability of detection. A the results, a map of EFAs was generated for the 2021 year.

Study Area
The study area (AOI) includes 3 administrative provinces of the P (NW -Italy): namely, Torino, Asti and the southern part of Vercelli (Figure located at an altitude ranging between 200 and 300 m a.s.l., covering abou characterized by a highly diversified agricultural context, hosting many EF it particularly attractive to the ARPEA, as it is called yearly to control the the declared EFAs.

Available Data
In this work, the following data were used: (a) satellite multispectral images from the Copernicus Sentinel-2 Multi Spectral Instrument (MSI) sensor; (b) maps of the Ecological Focus Area as reported by farmers in the GSAA (year 2021); and (c) calibration and validation data from ground campaigns.

Satellite Data
A total of 43 Level 2A images were collected covering the period from 28 February to 30 June 2021. Level 2A images are BOA (at-the-bottom of the atmosphere) reflectancecalibrated, which is proper for land applications [46,47]. The images were obtained from the Copernicus Open Access Hub geoportal (scihub.copernicus.eu) and provided as 100 km × 100 km tiles orthoprojected in the WGS84 UTM 32N reference system. Red (band 4) and NIR (band 8) bands were used to compute the corresponding Normalized Difference Vegetation Index (NDVI) maps [48]. These have a geometric and temporal resolution of 10 m and 5 days, respectively. The S2 Level 2A dataset is supplied together with an additional band called the "scene classification layer" (SCL), which assigns a classification meaning to image pixels, including cloudy, shadowy and/or failed. This layer was used to automatically detect and remove bad observations corresponding to the three above-mentioned classes within the NDVI time series.

EFA Maps from Farmers
To apply for the CAP contribution, farmers must yearly provide the correspondent GSAA, reporting much information, including the crop type, size of plot, location, etc. EFAs can be part of farmers' GSAA, as they are compliant with grants for greening. For this work, the ARPEA provided a vector map locating 12,510 EFAs and sizing 6486.13 ha for the 2021 agronomic season. It is worth remarking that, presently, these are not public data but, instead, are the prerogative of paying agencies.

Ground Data
In every classification problem, ground truth plays a crucial role. The size of ground samples (GC) strictly depends on the accuracy expectation of users and on a-priori knowledge about the probability of occurrences of classes. The sample size for a classification problem involving solely 2 classes (EFA and non-EFA) can be computed according to Equation (1) [49]: where n is the required samples size, ε i is the expected error of the i-th class (in this work, this was assumed equal to 0.05), z is the z-statistic at the 95% confidence level (in this work, this was equal to 1.96) and P i is the expected population proportion of the i-th class. According to a previous work from the authors [40], the expected percentage of false/wrong declarations from farmers, as reported in the GSAA, was 20%. Therefore, Pi was preliminarily set as equal to 0.2. The resulting theoretical sample size was 246 plots. Consequently, a total of 255 GCs (sizing about 318 ha) were obtained from the ARPEA dataset surveyed in the summer of 2021. During ground campaigns, regional technicians reported on three different situations: (a) fields compliant with the correspondent GSAA; (b) fields showing obvious signs of management within RP and, therefore, non-compliant with the EFA requirements; and (c) non-vegetated fields that consequently cannot be assumed to be EFAs. The provided GCs were randomly split into training (TP) and validation (VP) sets, corresponding to 60% (153 fields) and 40% (102 fields) of the total, respectively. Splitting was performed at the class code level to guarantee similar representativeness for the abovementioned 3 situations. Table 1 reports the number and size of the GCs; Figure 2 shows their spatial distribution.

Composing NDVI Time Series
As previously mentioned, NDVI is one of the most widely used vegetation spectral indices in agriculture [50] to include applications related to the detection of agronomic practices [51]. Consequently, in this work, an approach based on the analysis of NDVI time series (TS) appeared to be suitable for detecting unauthorised practices within EFAs. TS was generated through a self-developed routine implemented in IDL v 8.0.1. [52], accounting for the following operational steps at pixel level: (a) local NDVI temporal profile extraction from TS; (b) removal of bad observations using SCL layer as driver; (c) profile smoothing by a symmetric Savitzky-Golay filter (kernel = 3, derivative = 0, degree = 1); and (d) profile regularization (time step = 5 days) by spline interpolation (tensor value = 10) [53,54]. It is worth noting that filtering, interpolation and regularization of NDVI profiles reduce the signal noise permitting highlighting of the most significant trend components over time [55]. Final TS resulted in a stack of 25 equally spaced NDVI time series maps covering RP.

Looking for Markers within NDVI Temporal Profiles
Within EFAs, favourable conditions are expected to be present for the development of organisms and, consequently, of biodiversity. Intensive agriculture where no EFA is present can greatly compromise this natural mechanism by degrading the biologic value of large areas. EFAs are supported by CAP to ensure that farmers move their practices in the direction of better preserving the ecological value of agricultural environments. For this purpose, during spring, i.e. RP, farmers must stop all of their actions in the part of their fields that they declared as EFAs. Nevertheless, farmers often violate this restriction

Composing NDVI Time Series
As previously mentioned, NDVI is one of the most widely used vegetation spectral indices in agriculture [50] to include applications related to the detection of agronomic practices [51]. Consequently, in this work, an approach based on the analysis of NDVI time series (TS) appeared to be suitable for detecting unauthorised practices within EFAs. TS was generated through a self-developed routine implemented in IDL v 8.0.1. [52], accounting for the following operational steps at pixel level: (a) local NDVI temporal profile extraction from TS; (b) removal of bad observations using SCL layer as driver; (c) profile smoothing by a symmetric Savitzky-Golay filter (kernel = 3, derivative = 0, degree = 1); and (d) profile regularization (time step = 5 days) by spline interpolation (tensor value = 10) [53,54]. It is worth noting that filtering, interpolation and regularization of NDVI profiles reduce the signal noise permitting highlighting of the most significant trend components over time [55]. Final TS resulted in a stack of 25 equally spaced NDVI time series maps covering RP.

Looking for Markers within NDVI Temporal Profiles
Within EFAs, favourable conditions are expected to be present for the development of organisms and, consequently, of biodiversity. Intensive agriculture where no EFA is present can greatly compromise this natural mechanism by degrading the biologic value of large areas. EFAs are supported by CAP to ensure that farmers move their practices in the direction of better preserving the ecological value of agricultural environments. For this purpose, during spring, i.e. RP, farmers must stop all of their actions in the part of their fields that they declared as EFAs. Nevertheless, farmers often violate this restriction in an attempt to maximize the economical profit, which can increase if an additional field is cultivated. This makes it desirable for PPs to control farmers' declarations concerning EFAs as widely as possible. NDVI TS can be used for this purpose, making it possible to detect (even retrospectively) the whole history of a farmer's activities. These (mowing, sowing or harvesting) are generally associated with significant positive or negative variations alongside the local NDVI temporal profile [42]. Figure 3 shows some examples of the NDVI temporal profiles (field averaged) corresponding to six fields from GCs that farmers declared as EFAs. Some of them can be easily assigned to EFAs, while some others highlight evident variations, possibly related to the violation of CAP restrictions. In Figure 3a, the reported NDVI profile is the one expected for the natural development of vegetation related to an EFA. In Figure 3b, some significant variations in the profile can be recognized, possibly related to mowing operations. Figure 3c shows a profile that suggests sowing and the consequent growing of a crop. Figure 3d,f show profiles characterized by crop harvesting at the beginning and at the end of the growing period, respectively. Figure 3f shows a profile where a significant operation can be recognized in the middle of the controlled period. These examples cannot be assumed as general because the date and intensity of variations can vary enormously within an analysed period. Nevertheless, they are useful to demonstrate that some markers can be easily recognized alongside the NDVI temporal profiles to alert the possible presence of unexpected agronomic actions by which EFAs should not be affected. in an attempt to maximize the economical profit, which can increase if an additional field is cultivated. This makes it desirable for PPs to control farmers' declarations concerning EFAs as widely as possible. NDVI TS can be used for this purpose, making it possible to detect (even retrospectively) the whole history of a farmer's activities. These (mowing, sowing or harvesting) are generally associated with significant positive or negative variations alongside the local NDVI temporal profile [42]. Figure 3 shows some examples of the NDVI temporal profiles (field averaged) corresponding to six fields from GCs that farmers declared as EFAs. Some of them can be easily assigned to EFAs, while some others highlight evident variations, possibly related to the violation of CAP restrictions. In Figure  3a, the reported NDVI profile is the one expected for the natural development of vegetation related to an EFA. In Figure 3b, some significant variations in the profile can be recognized, possibly related to mowing operations. Figure 3c shows a profile that suggests sowing and the consequent growing of a crop. Figures 3d and 3f show profiles characterized by crop harvesting at the beginning and at the end of the growing period, respectively. Figure 3f shows a profile where a significant operation can be recognized in the middle of the controlled period. These examples cannot be assumed as general because the date and intensity of variations can vary enormously within an analysed period. Nevertheless, they are useful to demonstrate that some markers can be easily recognized alongside the NDVI temporal profiles to alert the possible presence of unexpected agronomic actions by which EFAs should not be affected.

Detecting Markers
Detection of markers was based on the analysis of the average TS profile of fields that were preventively computed by zonal statistics by SAGA GIS 7.9 [56].
Markers' detection was achieved by composing a succession of rules that applied to the NDVI temporal profile of EFA-declared fields. The first condition we tested was that EFA-potential fields were vegetated during RP. Consequently, all GSAA plots never exceeding 0.4 NDVI value during RP were coded (code = 4) and excluded from the successive processing steps. The 0.4 NDVI value was set as the threshold, according to the literature, which reports that lower values correspond to not significant vegetation cover [57,58].
Concerning the vegetated fields, their expected phenological behavior during RP is consistent with natural vegetation. This is expected to show a continuous increasing trend in RP [59,60]. Consequently, if any agronomic practice occurred, it would generate an anomalous variation of the trend, determining significant differences that would possibly be detectable by exploring the residuals values between a hypothesized and modeled trend function and the actual NDVI temporal profile.
According to this idea, the authors modeled the phenological trend (hereafter called y i ) of the i-th EFA field (field averaged) by a first-order linear regression. Processing was achieved by R software [61,62]. The corresponding residuals were therefore calculated and assessed by computing some statistics to inform about these anomalies. In particular, the maximum residual absolute value (RES max ), the coefficient of determination (R 2 , [63,64]) and the mean absolute error (MAE, [65]) were computed and mapped for each GSAA field [65]. The RES max was assumed to correspond to the maximum residual absolute value resulting from the calibrated model.

From Residual Statistics to Markers
To transform residual-related statistics into markers for anomaly detection, the RES max , MAE and R 2 statistical distributions were analysed by Otsu's method. This method [66] proved to be an effective technique [67,68] for objectively defining a data-driven threshold by analysing the frequency distribution of the variable of interest. Otsu's method is a nonparametric and unsupervised method that maximizes both the separability and the internal similarity of two groups of values through the identification of a proper threshold value from the histogram.
Concerning the present work, Otsu's method was applied to the statistical distributions of the RES max , MAE and R 2 from the training set, thus obtaining three threshold values (T MAE , T R 2 , T RES ) useful for recognition of EFA fields at different reliability levels. The following conditions were tested according to Equation (2): Depending on the number of conditions that the i-th GSAA plot successfully satisfied, five different EFA classes were mapped. Class meaning is reported in Table 2. Classification was then validated with reference to the validation set VP (made of 102 GCs) by computing the correspondent confusion matrix at the fields level. The related Overall Accuracy (OA), User's Accuracy (UA) and Producer's Accuracy (PA) were derived according to Hay [69]. An EFA map was used, finally, to derive operational information for control purposes. In Figure 4, a scheme of the proposed workflow is shown. Classification was then validated with reference to the validation set VP (made of 102 GCs) by computing the correspondent confusion matrix at the fields level. The related Overall Accuracy (OA), User's Accuracy (UA) and Producer's Accuracy (PA) were derived according to Hay [69]. An EFA map was used, finally, to derive operational information for control purposes. In Figure 4, a scheme of the proposed workflow is shown.

Threshold Definition
Once the NDVI trend was modeled at the field level, and residuals were computed, statistical distributions of the R 2 , RESmax and MAE from the TP dataset were analyzed for differences between EFA and non-EFA parcels. Figure 5 shows the thresholds found by Otsu's method and boxplots summarizing the statistical distributions of the R 2 , RESmax and MAE. The obtained threshold values were 0.55; 0.125 and 0.05 for the R 2 , RESmax and MAE, respectively. The above-mentioned conditions were therefore tested at the field level for all EFA-declared plots in the area.

Threshold Definition
Once the NDVI trend was modeled at the field level, and residuals were computed, statistical distributions of the R 2 , RES max and MAE from the TP dataset were analyzed for differences between EFA and non-EFA parcels. Figure 5 shows the thresholds found by Otsu's method and boxplots summarizing the statistical distributions of the R 2 , RES max and MAE. The obtained threshold values were 0.55; 0.125 and 0.05 for the R 2 , RES max and MAE, respectively. The above-mentioned conditions were therefore tested at the field level for all EFA-declared plots in the area. Agronomy 2021, 11, x FOR PEER REVIEW 10 of 16 It is worth noting that threshold definition, based on Otsu's method, was specifically aimed at avoiding subjective choices. The method is adaptive and data-driven, making it possible to automatically detect different threshold values according to the specific area one is considering. Consequently, the method can be applied wherever just by properly calibrating threshold values depending on the local agronomic conditions.

Classification and Validation
Classification was achieved by assigning each plot with the number of satisfied conditions (see table 2). These were interpreted as classes of probability that the tested plot was actually an EFA. Classification was validated through computation of the correspondent confusion matrix (Table 3) obtained with reference to the VP dataset. Overall Accuracy (OA) 63% Table 3 shows that Classes 3, 0 and 4 are characterized by a very high UA: 77%, 91%, 100%, respectively. Conversely, Classes 2 and 1 show medium (57%) and low (4%) UA values, respectively. These situations make evident that the adoption of a single condition (Class 1), out of the three reported in Equation 2, is absolutely insufficient for the proper detection of EFA plots. This could be related to an intrinsic weakness of the modelled conditions or to causal elements, such as anomalies related to unfavourable climatic and environmental conditions within RP (frost, drought, precipitation or other natural interference) [70]. It is worth noting that threshold definition, based on Otsu's method, was specifically aimed at avoiding subjective choices. The method is adaptive and data-driven, making it possible to automatically detect different threshold values according to the specific area one is considering. Consequently, the method can be applied wherever just by properly calibrating threshold values depending on the local agronomic conditions.

Classification and Validation
Classification was achieved by assigning each plot with the number of satisfied conditions (see Table 2). These were interpreted as classes of probability that the tested plot was actually an EFA. Classification was validated through computation of the correspondent confusion matrix (Table 3) obtained with reference to the VP dataset.  Table 3 shows that Classes 3, 0 and 4 are characterized by a very high UA: 77%, 91%, 100%, respectively. Conversely, Classes 2 and 1 show medium (57%) and low (4%) UA values, respectively. These situations make evident that the adoption of a single condition (Class 1), out of the three reported in Equation (2), is absolutely insufficient for the proper detection of EFA plots. This could be related to an intrinsic weakness of the modelled conditions or to causal elements, such as anomalies related to unfavourable climatic and environmental conditions within RP (frost, drought, precipitation or other natural interference) [70].
Concerning PA, the results were high for both the EFA (91%) and non-vegetated (100%) classes. Alternatively, results appear to be low (47%) for the non-EFA class. In fact, high confusion can be observed between Classes 0 and 1, where 23 non-EFA fields (35% out of the total) of VP were assigned to Class 1. This element further suggests that the satisfaction of a single condition, out of the tested three, is too weak for classification purposes. Consequently, one can admit that plots may be reasonably recognized as an EFA if at least two conditions are satisfied. With this new point of view in mind, the authors interpreted Classes 0 and 1 as related to non-EFA plots and Classes 2 and 3 as related to EFA plots at a different level of probability of detection. A new confusion matrix was therefore generated, and the correspondent UA and PA statistics were computed (Table 4). After the new reorganisation of classes, more promising accuracy values were obtained. Specifically, the UA and PA of the non-EFA Classes 0 and 1 improved significantly, together with an OA that grew from 63% to 84%. These values are consistent with references from the literature. Zheng [71], using Landsat data for detecting changes in agricultural areas (Indiana, US), related agronomical practices through a multi-temporal analysis of vegetation indices (VIs) time series and found an OA of 90%. Dusseux [72] found that it was possible to detect agricultural practices (e.g. mowing of meadows) by satellite-derived VIs time series from SPOT (Satellite Pour l'Observation de la Terre) data in Brittany with an OA between 80 and 90%.

Interpreting EFA Map
In order to highlight the operational potentialities of an EFA map ( Figure 6) with the above-mentioned accuracies, some quantitative information were obtained and discussed in light of CAP controls. In particular, area computation was achieved for the mapped classes and possible interpretations of the resulting percentages were given. Results are reported in Table 5. Table 5 shows that, in the AOI, 6064 fields (about 48%) were classified as non-EFA, 378 (about 3%) were classified as non-vegetated and only 6068 (about 49% out of the total) were actually confirmed as an EFA (Classes 3 and 2). Specifically, for fields classified as an EFA, 3786 are in Class 3 (about 30%), and 2282 are in Class 2 (about 18%). Similar percentage values can be obtained with reference to their areas, in place of the number of fields.
To give an interpretation useful for CAP controls, it is worth noting that classified plots were only those that farmers declared as an EFA for the CAP campaign in 2021.
The first issue to be discussed is that the PA from classification in the control framework plays the most important role: to ensure the capability of the procedure to confirm or reject the hypothesis that the tested plot is actually an EFA.
Given the following conditions: (i) Classes 0-1 and Class 4 show a PA of 82 % and 100%, respectively; (ii) in the EFA map, about the 51% of controlled plots fall in these classes, therefore one can deduce that, with a probability of something more than 0.82, 51% of controlled 2021 EFA fields correspond to untruthful declarations from farmers.
Consequently, Classes 0-1 and 4 as mapped in the EFA map require a field control. On the contrary, being that the PA of Class 3 is equal to 88%, we can reasonably admit that no controls are required for this class from PP, and the farmers' declarations can be assumed compliant. For the 2021 campaign, class 3 corresponds to about 30% of controlled parcels that can avoid control by ground surveys.   Table 5 shows that, in the AOI, 6064 fields (about 48%) were classified as non-EFA, 378 (about 3%) were classified as non-vegetated and only 6068 (about 49% out of the total) were actually confirmed as an EFA (Classes 3 and 2). Specifically, for fields classified as an EFA, 3786 are in Class 3 (about 30%), and 2282 are in Class 2 (about 18%). Similar percentage values can be obtained with reference to their areas, in place of the number of fields.
To give an interpretation useful for CAP controls, it is worth noting that classified plots were only those that farmers declared as an EFA for the CAP campaign in 2021.
The first issue to be discussed is that the PA from classification in the control framework plays the most important role: to ensure the capability of the procedure to confirm or reject the hypothesis that the tested plot is actually an EFA.
Given the following conditions: (i) Classes 0-1 and Class 4 show a PA of 82 % and 100%, respectively; (ii) in the EFA map, about the 51% of controlled plots fall in these classes, therefore one can deduce that, with a probability of something more than 0.82,  To interpret the operational meaning of Class 2 for which no PA can be separately calculated, we can refer to the correspondent UA (57%). The UA defines the probability that only 57% of the Class 2 fields (about 18% of the total) are actually EFAs. This suggests that controls from PP are desirable in such a measure as they are obtainable via weighting programmed controls by this value. These preliminary results appear critical in terms of ordinary CAP controls strategy, highlighting a significant amount of possibly untruthful declarations detected with reasonably high accuracy via satellite data. These findings, obviously, raise intriguing questions regarding the CAP criteria for monitoring EFAs in order to ensure the properness of given economical grants. From this point of view, it is worth noting that ordinarily the ARPEA controls 5% of the received GSAA. A satellite-based approach, such as the one here presented, would guarantee a complete overview and the possibility of better focusing the GCs. Since controls are currently completely based on GCs, they cannot supply timely continuous information, which limits the controls to selected fields with a time that may not be exactly proper for the correct detection of the field management. Moreover, the time of the controls is strictly dependent on administrative constraints and, therefore, highly compromised in terms of synchronization with crop phenology. Depending on the time of the GC, some practices can be easily identified, such as the presence of other crops (i.e. corn, wheat), ploughing or harvesting. Others, however, such as mowing, may be almost impossible to detect if run too many days after the practice. These objective weaknesses of the current control process can be overcome by the utilization of satellite data, as demonstrated in this work.

Conclusions
In this work, a prototype service aimed at supporting PPs controls within the greening CAP framework, with special concern about Ecological Focus Area (EFA), is presented. Since the recent Italian National Report No. 5465 stipulates that mowing and any other soil management operations are prohibited on set-aside land designated as an EFA during RP, controls are required to test the respect for these conditions. The work was stimulated by the ARPEA for facing the operational task of EFA controls.
Since the ARPEA explicitly required that the classification approach was simple and explicit in its agronomical meaning, machine/deep learning-based (MDL) approaches were excluded. Conversely, a simpler, rule-based approach was developed for EFA detection based on the application of simple rules, or tasks, to disaggregate a complex problem into a succession of simple and more controllable ones, where agronomic knowledge plays a crucial role.
The method relies on the analysis of NDVI profiles from the Sentinel-2 mission, which is used for extracting three local statistics useful for EFA detection: the R 2 , RES max and MAE. They were computed by approximating the local NDVI temporal profile in the reference period with a first-order polynomial. The EFA fields were recognized by testing the satisfaction of three threshold-based conditions. The threshold values were defined according to Otsu's method and applied to a statistical distribution of the R 2 , RES max and MAE from GCs. The correspondent 2021 EFA map was generated and validated.
The methodology was applied to a study area that included the administrative provinces of Torino, Asti and Vercelli (Piemonte Region, NW Italy). Specifically, over 12,510 EFA fields (CAP campaign 2021) were assessed to test the truthfulness of farmers' GSAA declarations. Results showed an overall accuracy of 84%, indicating that the approach is promising. The authors retain that this procedure represents a valid alternative (or integrating) tool for ground controls by PPs. These preliminary results appear to be operationally critical in terms of CAP controls, highlighting a significant amount of possibly untruthful declarations detected with reasonably high accuracy via satellite data. These findings, obviously, raise intriguing questions regarding the present CAP criteria for monitoring EFAs in order to ensure the properness of given economical grants.
The procedure can be assumed applicable in every territory within the EU because EFAs do not depend on local crop types and behave similarly to natural meadows.
The procedure simply needs to be fine-tuned depending on the local agronomic calendar. In Italy, the reference period for testing conditions occurs between 1st March and 30th June (No. 5465, applying EU Regulation No. 1307/2013 of the European Parliament and Council).
Some limitations, however, still persist concerning the proposed procedure. One is possibly related to the presence of clouds over the AOI. In fact, the reference period useful for controls corresponds to the cloudiest period of th year. In unfavourable years, it is possible that detection will fail, since it relies entirely on optical data. Consequently, future development of the methodology could include a possible integration of SAR data from the S1 mission, as well as investigation into new phenological markers able to improve classification accuracy. Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.