Area Monitoring — Traffic Light System
Tailoring scenarios for evaluating specific land-use or crop type claims
The EU Common Agricultural Policy requires the control of subsidies claimed annually for millions of agricultural parcels over all European Member States. The advancement and operational availability of Copernicus Sentinel and other satellite data provides means to monitor and control large areas remotely. The overall objective of the Area Monitoring System is to convert the satellite-derived signals and markers into decisions about compliance with claims.
The system evaluates logical rules that compare observed conditions with criteria for compliance to a particular scheme/measure. It assigns one of the following codes to each Field of Interest (FOI):
- Green: FOI assessed and confirmed as compliant
- Yellow: FOI assessed but there is insufficient evidence neither to confirm nor reject explicitly the declared scheme / measure
- Red: FOI assessed and confirmed as definitely non-compliant
The Traffic Light System (TLS) for the basic payment scheme for Slovenia covers eleven different scenarios, where each scenario incorporates the use of markers and rules and is specially tailored for a specific land-use or crop type.
Within the basic payment scheme, it is more important to distinguish whether or not parcels are actually managed with some definite agricultural activity, rather than provide evidence that the crop claimed is actually the one grown. Hence, if evidence of any farming activity is found among markers and signals for a FOI, then this FOI is assigned a Green, even if, for example, corn is cultivated on an FOI claimed as winter wheat. Thus, the TLS searches for evidence of presence or absence of farming activity using the following markers (links to detailed description of each marker are given at the bottom of this blog post):
- Mowing marker searches for evidence that a FOI has been mowed.
- Bare-soil marker searches for evidence that a FOI has been ploughed.
- Consistency of a claim is checked with crop-group classification marker, similarity marker, and Euclidian distance marker. If signals of a FOI are found to be consistent with signals of other FOIs with the same claim, then this serves as indirect evidence of farming activity. Corn, winter wheat, or vegetables for example do not grow by themselves without any farming activity, such as ploughing, sowing, and harvest.
- Consistency of a claim of permanent crop, for example hops, fruit trees, etc., is checked with land-use marker, similarity marker, and Euclidian distance marker.
- Homogeneity marker identifies FOIs with heterogeneous land use that produces noisy signals which cannot confidently be assessed by the above markers.
- Mean-NDVI marker identifies FOIs with very little vegetation cover. A FOI can lose vegetation cover for example when it is turned into a built-up area.
The figure below gives a high-level overview of the TLS for the basic payment scheme claims.
FOIs are divided into two groups based on their size:
- large FOIs that contain at least one Sentinel-2 pixel,
- and small FOIs that don’t contain at least one full Sentinel-2 pixel.
We use Sentinel-2 signals to monitor the former and PlanetFusion signals for the latter. The TLS decision logic is exactly the same for both groups of FOIs and it is applied to both. In this blog we will present results for the group of FOIs monitored with Sentinel-2.
The functional TLS has been used to process all FOIs that contain at least one Sentinel-2 pixel (over 618 000) for 2019 in Slovenia and fed into the Expert App system for manual evaluation and assignment of Green (compliant) or follow-up with claimant. Red (non-compliant) is not automatically assigned to any FOI — leaving it for an expert operator to judge and communicate with the claimant before any penalties are imposed.
Scenarios are adjustable to business rules of each country’s Paying Agency. The marker system and the TLS are configurable and the Expert App, as well as supporting manual evaluation, enables and supports detailed review phases to check and optimise algorithms and decision thresholds.
The eleven developed and fine-tuned scenarios were applied to all Slovenian claims in 2019 and produced encouraging results, especially for arable land (S05) and meadows (S06).
Scenarios 5 (Arable Land/Annual crops) and 6 (Meadows and Grasses) are probably the most important ones in most countries. In Slovenia they represent over 90% of all FOIs. In the following sections we will go through the decision logic of these two scenarios with detailed examples for some of possible decisions.
Scenario 6: Meadows, grasses, clover, grass/clover mixtures, and fallow land
This scenario is applied to all FOIs claimed as meadows, grasses, clover, grass/clover mixtures and fallow land. The main farming activity on FOIs claimed in this group is expected to be mowing. Grasses, clover, grass/clover mixtures are grown on arable land and therefore FOIs with these claims can also be ploughed in the beginning or end of the season. This scenario therefore searches for evidence of farming activity using all markers except the land-use marker.
Condition 1: Low mean NDVI
If a FOI claimed as a meadow, grass, clover, or grass-clover mixture yields low mean NDVI throughout the season, then it is assigned Yellow, because a FOI growing these crop types should be vegetated for most of the season (and show high mean NDVI). Fallow land, however, is excluded from this conditional test because it can remain without vegetation through most of the season.
Below we show a NDVI time-series of a FOI from the top image in this blog post (annotated “partly built-up”) which is assigned Yellow due to this condition. Sentinel-2 imagery suggests that the northern part of the FOI is built-up with no farming activity ongoing.
Condition 2: Large and inhomogeneous
If a FOI has heterogenous land use, other markers might not be reliable, so they are assigned Yellow and set aside for manual evaluation.
Below is an example of a FOI claimed as meadow. However Sentinel-2 and digital ortho-photo imagery suggests that parts of this FOI are covered with at least three different land-use types: meadow, forest, and arable land.
Condition 3: Consistent with claimed group and at least one mowing event detected
In the case of FOIs claimed as meadows, grasses and the like, it is not sufficient for a FOI just to be found consistent with its claim, in addition evidence of mowing between May 1 and October 15 must be observed.
In the example FOI shown below, the crop-group classification marker has assigned over 99% probability that the FOI belongs to the claimed crop group. The mowing marker has found evidence for four mowing events between May and mid-October. Three of them can be easily recognized in Sentinel-2 RGB imagery shown below, and so the FOI can confidently be assigned Green.
Another interesting example FOI is shown in the Expert App animation below. In the case of this FOI three mowing events are also detected. During the second and third event, the FOI has only been partially mowed.
Condition 4: Consistency with arable land crop groups is high
As already mentioned, within the basic payment scheme it is irrelevant what is actually growing on claimed land as long as it is farmed. Therefore wrong claims should not necessarily be assigned Yellow. This condition checks if a FOI is highly consistent with actively farmed arable land crop groups and accordingly assigns Green.
Below we show an example of an Expert App screeenshot for a FOI claimed as clover-grass mixture. The crop-group classification marker provides a high confidence that corn is cultivated on this FOI, which is also corroborated by the similarity marker’s score. The NDVI profile of this FOI is much more similar to NDVI profiles of cornfields in its neighbourhood than to a NDVI profile of FOIs claimed as a meadow, grass, clover, and the like. In addition, the bare-soil marker identifies bare-soil in many of the observations from the start and end of the season, which is again consistent with farming practice for corn in Slovenia. Markers therefore provide strong evidence that the claim of this FOI is wrong, but since it is being farmed, it is confidently assigned Green in accordance with the scheme.
Condition 5: Bare soil detected between April and September
The aim of this condition is to assign Green to all FOIs that are being farmed, even if the claim is erroneous and any of the previous markers failed to provide evidence of farming activity. All FOIs reaching this point are assigned Green, if bare soil is detected between Julian days 90–275 (start of April to end of September).
Below an Expert App screenshot of an example similar to the one above is shown. The only difference here is that neither the crop-group classification nor the similarity marker provide high confidence about what is really growing on this FOI. They both agree that it is a type of summer cereal, but confidence is below 80%, which is required to assign Green automatically by this condition. However, regardless of what crop type is truly growing on this FOI, because the bare-soil marker makes conclusive bare-soil observations, there is sufficient evidence of farming activity to assign Green confidently anyway.
FOIs without evidence of farming activity
FOIs that end up at the very end node of this decision tree are those for which marker results provide no strong evidence of farming activity. These are typically FOIs that are consistent with the claimed group, but no conclusive evidence of mowing or ploughing is seen in satellite-derived signals and markers.
An example FOI claimed as meadow assigned Yellow is shown below. Crop-group classification, similarity and Euclidian distance markers suggest that this is really a meadow, however mowing marker does not detect any events. NDVI time-series and Sentinel-2 imagery shown in the screenshot of the Expert App below also do not provide any evidence that this FOI has been mowed or ploughed.
Scenario 5: Arable land, annual crops
This scenario is applied to all FOIs claimed as annual crops on arable land.
As can be seen in the above decision logic diagram, many of the conditions are similar or the same as the ones discussed above for Scenario 6. We’ll skip those in the overview below and focus only on the new ones.
Condition 4: Consistent with claimed group
As described in the introduction, consistency between FOI’s signals and its claim serves as a proxy for detection of farming activity. Annual crops are usually sown on a ploughed field and get harvested after senescence. Farming practice and phenology are different from crop type to crop type and together they shape the signals of each crop type differently. Such differences enable classification of different crops. Without farming activity, these signals would look very different.
Below we show NDVI time-series for a few FOIs, which were all found to be consistent with its claim based on crop-group classification, similarity and Euclidean distance markers.
Condition 5: Consistency is high with the predicted group
As discussed above, a FOI with a wrong claim that is farmed should still get assigned Green. Hence, FOIs for which crop-group classification, similarity and Euclidian distance markers provide high confidence about their prediction, get assigned Green even if the predicted crop group is different from the claimed one.
Below we show NDVI time-series for a few FOIs, which were all found to be not consistent with its claim based on crop-group classification, similarity and Euclidean distance markers.
Condition 6: Consistent with Scenario 6 groups (meadows, grasses) and mowed
If grass, clover, or grass-clover mixtures are growing on a FOI with a wrong claim of growing annual crops on arable land, then such FOI is assigned Green, if evidence of mowing is observed.
Below is an example of a FOI claimed to grow corn, but since the FOI closely resembles grass, clover, or grass-clover mixtures, and there is strong evidence of mowing, a Green can be assigned confidently.
Condition 7: Bare soil detected between April and September
If, on any of the FOIs reaching this point, bare soil is observed during the specified time interval (in this case the start April to the end of September), this is also interpreted as clear evidence of farming activity, and a Green may be confidently assigned, even if the wrong crop is claimed and classification markers do not yield high-enough confidence levels in their predictions.
Below is an example FOI without a high-enough confidence crop-group prediction. The crop-group classification marker is confused between groups containing buckwheat (claimed) and vegetables (predicted). The bare-soil marker has, however, detected multiple bare-soil observations, which provides strong evidence of farming activity. As a result this FOI is assigned Green.
FOIs without evidence of farming activity
FOIs that end up at the very end node of the decision tree are those for which marker results provide no evidence of farming activity.
An example FOI claimed as corn assigned Yellow is shown below. Crop-group classification, similarity and Euclidian distance markers suggest that this is really a meadow. However, the mowing marker does not detect any events nor does bare-soil marker find any clear bare-soil observations. The NDVI time-series and Sentinel-2 imagery shown in the screenshot of the Expert App below, also does not provide any evidence that this FOI has been mowed or ploughed, so it has to be left for manual assessment or communication with the claimant for more evidence.
Conclusion
The functional TLS has been developed, fine-tuned and optimised, with effective methods for crop and land use assignment, mowing and bare soil markers. It has been used to process over 618,000 FOIs for 2019 in Slovenia, yielding nearly 96% of Green assignments. All Yellow assignments have been fed into the Expert App system for manual evaluation and follow-up with claimant or reassignment to Green. A sample of FOIs assigned Green are also fed into the Expert App for random checking, verification and for use in improvement of the Machine Learning algorithms.
Scenarios, models, markers and thresholds for each are fully-adjustable to local conditions and the business rules of different country’s Paying Agencies. The marker system and TLS are fully configurable, as is the Expert App, which enables and supports further review phases to check and refine what works better.
Check the Area Monitoring documentation for more information.
Further reading
This post is one of the series of blogs related to our work in Area Monitoring. We have decided to openly share our knowledge on this subject as we believe that discussion and comparison of approaches are required among all the groups involved in it. We would welcome any kind of feedback, ideas and lessons learned. For those willing to do it publicly, we are happy to host them at this place.
The content:
- High-Level Concept
- Data Handling
- Outlier detection
- Identifying built-up areas
- Similarity Score
- Bare Soil Marker
- Mowing Marker
- Pixel-level Mowing Marker
- Crop Type Marker
- Homogeneity Marker
- Parcel Boundary Detection
- Land Cover Classification (still to come)
- Minimum Agriculture Activity (still to come)
- Combining the Markers into Decisions
- The Challenge of Small Parcels
- The value of super resolution — real world use case
- Traffic Light System (this post)
- Expert Judgement Application
- Agricultural Activity Throughout the Year
