The European Union’s (EU) Common Agricultural Policy (CAP) is arriving at a crossroads in the implementation of the integrated administration and control systems (IACS) by the EU Member States. On the one hand, the technical components that underlie the IACS, such as the Land Parcel Identification System (LPIS) and digital GeoSpatial Aid Application (GSAA) are reaching levels of maturity and quality that causes systemic errors to approach the so-called “materiality threshold” ( 2%) in a majority of Member States. At the same time, however, public demands for improved environmental and climate performance need to be reconciled with calls from administrations and farmers for simplified administration and control procedures. This is particularly important in the current economic conjuncture, where the extra efforts to implement ambitious environmental and climate targets are not readily compensated by a rise in income from agricultural production and where the overall CAP budget is decreasing. In this paper, we focus on the introduction of Copernicus Sentinel high resolution data streams that facilitate reduction of burden on Member States’ administrations in the management and control of area-based support schemes and help to increase compliance by introducing a new model in which farmers are alerted, rather than penalised.
The spatial or area dimension entered the CAP in the 1990’s, as a control instrument to combat over declarations in the premium system such as for olive oil or cereals. Soon this control based on satellite imagery gave rise to a prevention instrument on the farmer side: the LPIS. By 2009 the bulk of the CAP aid schemes was completely decoupled from production and aid amounts were directly proportional to area extends.
The on-the-spot-check (OTSC) control system was developed on, at that time, state of the art satellite imagery and GNSS data capture. The system dealt with the area parameter and ensured that the CAP direct payments had among the lowest error rates of any EU policy.
The modernization of the CAP in 2013 introduced the greening schemes, imposing the farmers to perform measures that benefit the climate and environment. The OTSC mechanism was adjusted in an attempt to provide similar assurance on eligibility and area as it previously did on the decoupled payment. The adjustment had two direct consequences: first, it stretched the system to its limit, increasing the operational burden for the Member States and second, greening measures were often tailored to fit ease of control rather than to pursue higher environmental ambition.
The low error rate, high burden and constraints to introduce performance made the control mechanism much less effective and less efficient than it had been in its early days. Additionally, from the farmer’s perspective, it could lead to one of two negative outcomes: either loss of time and privacy (in case of compliance) or, less often, penalties (in case of non-compliance). These negative perceptions from farmer and administration were aired towards the political hierarchy.
The arrival of the Sentinels coincided with this crisis in the OTSC domain. Their wall to wall coverage, free availability and high revisit frequency offered the potential to address those elements where OTSC fell short.
However, harvesting the Sentinel potential, requires more than substitution or extension of OTSC data sources, as this would neither address the identified shortfalls nor lead to a reduction of the burden. Instead, the control mechanism is subject to a complete revision, and that involves a paradigm shift in the concept of control, known as “checks by monitoring”.
The checks by monitoring no longer adhere to a linear application/control/payment(penalty) procedure but aim to improve compliance by prevention and interacting with the farmer before payments. As a positive side effect, this reduces the need for penalties.
Fortunately, the introduction of Sentinel data streams in the “checks by monitoring” also coincides with the emergence of other digital technologies in agriculture and the rural economy, such as enhanced GNSS (Galileo, other improved GNSSs), improved digital communication networks (5G) and the uptake of sensors and robotics in agricultural practices. The combined use of these additional data streams is expected to create synergies between these data collection and exchange processes and the monitoring of progress towards CAP performance targets. It is obvious that these innovations require appropriate flanking R&D which addresses the upscaling issues involved. These include, amongst others, Big Data Analytics required to infer relevant spatial, radiometric and temporal patterns in the (open access) Sentinel sensor data, smart validation approaches, expansion and uptake of open reference data and open source software for near-instantaneous decision support (to both farmer and administration) and information aggregation that supports selecting and managing CAP environmental and climate interventions (e.g. towards water quality improvement, higher carbon sequestration in soils, etc.). The new Horizon Europe “food and natural resources” pillar allocates a 10 Billion Euro budget in the period 2021-2027 to support various R&D instruments in this domain.
The paper tackles challenges of managing these changes and tries to identify some of the current knowledge and technology gaps to migrate towards the anticipated operational prevention system.

In May 2018, the European Commission has adopted new rules to allow a range of modern technologies to be used when carrying out checks for the area-based payments of the Common Agricultural Policy (CAP). The huge potential of accurate and timely data provided by the Sentinels satellites was recognized and at the basis on the new “CAP monitoring approach”. According to these new rules, the eligibility for the aid or support requested should no more be decided through “on-the-spot checks” which take place once a year and over a 5% sample of the subsidy applications but using Sentinels data to continuously, regularly and systematically monitor the whole country.

The Sentinels for CAP (Sen4CAP) project aims at providing to the European and national stakeholders of the CAP validated algorithms, products and best practices for agricultural monitoring relevant for this CAP reform. Special attention is given both to Sentinel-1 and Sentinel-2. The Sen4CAP project is conducted with the guidance of a Steering Committee made of DG-Agri, DG-JRC, DG-Grow. It also involves six Paying Agencies (Czech Republic, Italy, Lithuania, Netherlands, Spain and Romania) acting as pilot countries.

During the first months of the project, the six pilot Paying Agencies were consulted to define their needs. Four type of Earth Observation products were identified: (i) crop type maps, (ii) grassland mowing detection products, (iii) agricultural practices monitoring product focusing on the detection of harvests, catch crops, nitrogen-fixing crops and land-lying fallows and (iv) biophysical indicators. These products have been identified as relevant to support the compliance decision for the crop diversification, permanent grassland, Ecological Focus Area and land abandonment payments.

A benchmarking exercise was conducted to test five algorithms for each product and select the one that best meets the requirements expressed by the Paying Agencies. The algorithms make use of both Sentinel 1 and 2 time series, run continuously along the year, deliver information at the parcel-level (as defined by the subsidy declarations shared by the Paying Agencies) and provide confidence indicators. Nine months after the start of the project, first prototype products generated for the years 2016 and 2017 were presented to the Steering Committee, to the pilot Paying Agencies and to the Committee for Direct Payments of DG Agri.

National crop type map was obtained over the Netherlands with an overall accuracy of 96% and F-Score values for the 5 main crop classes higher than 0.95. In the other pilot countries, crop type maps were generated over smaller test sites (10.000 km²), with accuracy values ranging between 66% and 93%. The detection of grassland mowing events was successfully validated over test sites in Netherlands and Czech Republic, with 88% and 89% of correct detections respectively. Agricultural practices product was also run over test sites in all the countries, but limited data were available for validation.
Even if obtained with methods not yet fully mature, these first results raised a clear interest from the Paying Agencies and the Steering Committee as they provided first evidence of how the monitoring approach could be conducted with Sentinels data.

On this basis, the first version of the open-source Sen4CAP processing system has been designed, as an operational standalone processing system made of modules generating agricultural products from Sentinel-2 L1C and Sentinel-1 SLC time series along the growing season. It is based on machine learning and time series analysis algorithms. Tools are designed in way that Paying Agencies and their suppliers can take them and customize to fit their business practices and their ICT infrastructure, ideally keeping core processing part in the cloud and connect to own system using web services and data transfer

The project is now entering in a phase of national-scale demonstration, where products are being generated at national scale over the six pilot countries. The 2018 processing is ongoing on a cloud infrastructure and will be delivered to the Paying Agencies before the end of the year. Their relevance will then be tested through a set of concrete eligibility decision applications. The demonstration will continue throughout 2019, building on the first results to improve the system and the products and providing appropriate capacity building and training to ensure the system uptake by the Paying Agencies. The transfer to the European DIAS cloud infrastructure is already being prepared, with a selection procedure started in August involving the comparison and the test of the different platforms.

With CAP policy going towards Area Monitoring, Sen4CAP tools and knowhow will serve a significant purpose both in methodology design, using lessons learned, discovered limitations and capabilities of the approach, as well as implementation based on project open-source outputs.

The Common Agriculture Policy is based on the provision to the European farmers of subsidies, supplied through several National Paying Agencies. These entities operate a check on the compliance of the farmers applications based on the integrated usage of few remote sensing images collected during the growing season and by ground surveys. According to the new Common Agriculture Policy implementation (2020+), this approach will evolve from controlling to monitoring each farm. This relevant change can be boosted by the availability of dense open satellite time series, collected by Sentinel-1 and Sentinel-2 constellation, allowing a continuous monitor all along the agronomic year on each agricultural parcels.
All Paying Agencies are requested to be gradually compliant with this new approach, therefore each of them must implement workflows allowing the control of the payment according to the new schema. In this scenario AGEA, the Italian Agency for Payments in Agriculture, has operationally implemented a workflow for the control of payments for the Single Payments Schema based on the usage of time series of Sentinel-1 and Sentinel-2 data. The workflow provides a description of the activities over each parcel through markers, that are computed from spectral indices extracted from satellite time series
The computation of these markers are the first step of a workflow that analyzes the markers with a traffic-light approach based on specific regulations. Green-light and red-light parcels are marked as compliant or not compliant, while parcels with other traffic-light color are addressed, at a later stage, to additional controls based on different techniques such as higher resolution satellite data, geotagged photos or ground surveys.
The whole workflow, integrating satellite derived information with additional verifications, has been operationally implemented over Foggia area, one of the largest Italian administrative province (about 7000 sqkm), with over 600,000 parcels in the farmers applications.
The workflow provides continuous update of markers, tailored according to the crop groups, all along the agronomic year, and uses all available Sentinel-1 and Sentinel-2 data in support to the Single Payments Schema.
The adopted processing schema, obtained results in terms of different type of analytics and in terms of improvement of AGEA control workflow are presented and analyzed

The current legal framework of the Common Agricultural Policy (CAP) from 2013 will be reformed by 2020 in order to modernize and simplify the CAP. In this upcoming CAP reform, the use of Copernicus satellite imagery is seen to take an increasing role for improving the Integrated Administration and Control System (IACS) and making it more cost efficient. Taking into account the operational aspect of the monitoring process of the CAP, the use of Sentinel-1 SAR images is a key point, especially over areas as Belgium with a frequent cloud cover. Indeed, SAR systems are barely affected by the weather conditions as SAR microwaves are able to penetrate cloud coverage. Moreover, knowing that the SAR backscatter is very sensitive to the geometry and the water content of the target, the main goal of this study is to assess the impact of different ways of processing the Sentinel-1 acquisition time series on the quality of crop type identification. These different processing approach concern ascending and/or descending S1 tracks with different acquisition time and different geometry.
The study area is the Wallonia, Southern part of Belgium, covering about 16.5 thousand square kilometres. The Wallonia is covered by six different Sentinel-1 ascending and descending tracks with a temporal frequency of one acquisition every 2-3 days. The analyses are carried out for the 2017 and 2018 seasons, using the Land Parcel Identification System (LPIS) built, managed and updated by the Walloon Paying Agency and the On The Spot Checks (OTSC) results as reference data. The total number of considered parcels is 279,419 for 2017 and 286,288 for 2018 campaigns, with more than 100 crop types for each year.
A Random Forest object oriented classification algorithm is used for crop type identification using completely independent calibration and validation data sets. In term of Overall Accuracy (OA), the classification result using Sentinel-1 imagery only reaches 0.93. The quality of different classification results is more accurately quantified using F1 scores for different crop types.
The impact of different SAR data sets used for the classification is assessed comparing the results obtained:
- using SAR data acquired on a single track or combining images acquired on several tracks, only ascending or descending tracks or both, ascending and descending tracks;
- using different periods of SAR data time series for different crop types.
The parcel size impact on accuracy of crop type identification using Sentinel 1 data is also assessed allowing to better characterize the capacities of Sentinel 1 imagery on crop type identification according to the field size distribution.

Operational monitoring of agricultural activities is essential for ensuring global food security. This, however, requires thorough knowledge of the precise location of agricultural areas in general, and agricultural parcel boundaries more specifically. Not only does this information contribute directly to the reporting of agricultural statistics, but also indirectly it can help focusing remote sensing tasks for field-level applications such as crop type mapping, phenological monitoring and disease detection. While such information on the location of parcels is currently collected on the national level in the context of the Common Agricultural Policy (CAP) of the European Union, near real-time open-access to these data is currently still lacking for many of the countries involved, nor do such data exist in any way for many other agricultural regions around the world.

Here we present an innovative method that uses deep learning tools to fully automatically detect and delineate agricultural parcels based on Sentinel-2 imagery. The backbone of the algorithm is a fully convolutional neural network which was trained on the agricultural parcel boundaries of Belgium as registered in the CAP framework. The algorithm starts from all available Sentinel-2 acquisitions in a certain time frame for a specific tile. The idea behind the method is that individual agricultural parcels behave differently along the growing season, and hence their spectral signature will have a field-specific behavior. The algorithm works on small blocks and constructs for each block automatically cloud-free NDVI stacks along the growing season. The convolutional neural network then detects edges around parcels as strong gradients in the NDVI signature. The final output aligns the individual blocks and also copes with overlapping Sentinel-2 tiles, resulting in a vector-based dataset containing all detected parcels in the desired region as a convenient shapefile.

We show that the performance on Sentinel-2 tiles over Belgium is excellent , and a verification of the method over a completely independent Bulgarian Sentinel-2 tile also proved to be good . Implementation of the method on the Terrascope platform (https://terrascope.be/en) and the ability to use a powerful GPU infrastructure allow our method to be easily upscaled to other regions in the world.

The automated parcel detection forms the basis for many useful applications. For example, satellite signal averaging over an entire agricultural parcel produces much cleaner crop signatures (e.g. for Sentinel-1) compared to pixel-based methods , resulting in higher accuracy feature-based crop maps (e.g. https://app.watchitgrow.be/cropmap/), and an improved performance for field-specific crop monitoring. Knowledge on the outer boundaries of a field may also help for focused analyses of within-field heterogeneities. Our method can therefore be helpful as a basis for a wide range of agricultural applications based on remote sensing.

One of the aims of the Common Agricultural Policy (CAP) is ensuring that farmers within the European Union (EU) make a reasonable living through the provision of subsidies. This serves as a great incentive to the farmers as they ensure that consumers have a stable supply of food. In order to ensure that these subsidies are judiciously expended, at the various country levels, computer-aided interpretation of very high resolution imagery and field campaigns are used to check the aerial claims of farmers before the subsidies are given out. These validated claims are consolidated into the Land Parcel Identification System (LPIS). Even though the LPIS is becoming freely available in a number of EU countries, there is still restricted access to this data in many regions of Germany. Even where they are available, the dataset is published annually and sometimes outdated. This makes it difficult to identify crop types at any point in time before the release of the dataset. This research seeks to tackle these issues by making use of the freely available Sentinel-1 and Sentinel-2 data and existing LPIS dataset of Germany for 2016 and 2017.This research adopts a two-step OBIA workflow that consists of parcel-identification and crop type mapping. In the first step, various features including vegetation indexes from Sentinel-2 and radar polarization ratios from Sentinel-1 are calculated. The optimal shape, scale and compactness parameters for delineating agricultural parcels using the Multiresolution Segmentation (MRS) algorithm for every region of Germany is identified based on the extracted features. The parameter combinations whose segmentation result has the highest geometric match to the LPIS parcels are considered the optimal ones. In the process, the optimal features are also identified. In the second step, using the LPIS parcels of 2017 as the reference layer, the crop type of newly segmented agricultural parcels, regardless of which point in time and geolocation, would be identified using a Transfer Learning (TL) scheme, which is still being developed. Initial results from the first step show that Sentinel-1 is not suited for identifying agricultural parcels largely due to the speckled and non-homogeneous nature of the images even after applying a multi-temporal speckle filter. With respect to the mono-temporal Sentinel-2 datasets, the scenes acquired during the summer season have the highest geometric match to the LPIS parcels. However, the highest geometric match was only realized by using a multi-temporal dataset that covered five different dates of the 2017 growing season.