Area Monitoring — The Challenge of Small Parcels

Anze Zupanc
Sentinel Hub Blog
Published in
12 min readJan 20, 2022

--

With a little help from our friends at Planet — From PlanetScope to Planet Fusion

The EU Common Agricultural Policy requires the control of subsidies claimed annually for millions of agricultural parcels over all European Member States. All parcels, whether small, large, elongated or rectangular, need to be monitored. Small and elongated parcels pose a challenge for the Checks by Monitoring system (as well as the future Area Monitoring System) that uses Copernicus Sentinel data as the main data source. We have stumbled upon the very same problem when implementing the system for The Slovenian National Paying Agency. In this blog post we describe our experiences and results addressing these chalenges using PlanetScope and Planet Fusion data to monitor parcels that are too small to be monitored with Sentinel.

Planet Fusion, May 1st 2021, Ptuj, Slovenia

Introduction

To ensure the quality of the parcels’ signal, Joint Research Centre (JRC), the research body of the European Commission, recommends a limiting criteria of at least eight full Sentinel-2 pixels inside the border of the specific FOI (feature of interest) and, if possible, introduction of 5-meter (1/2 Sentinel-2 pixel) internal buffer. This rule makes a lot of sense. Sentinel-2 data are not perfectly aligned (same for all other automatically processed satellite data), its multi-temporal geometrical accuracy being assessed as 0.3 pixels, which is why the spectral measurements at the border will be distorted with the signal of the neighbourhood objects. And taking statistics of numerous observations within the field will surely produce much better quality of the results. The reality, however, is often not too compliant with scientific best practices. In Slovenia, and several other member states, where agriculture parcels were often split during the inheritance proceedings, more than 40% of the parcels are lost by complying to this rule, even if we don’t take the internal buffer into account.

Due to this structure of the parcels, we have decided, together with a Paying Agency, to use Sentinel for monitoring of all parcels, which contain at least one full Sentinel-2 pixel. Based on our assessment, the signal is still good enough, albeit the impact of the size is noticeable in the marker results. We did not really have much choice, if we wanted to establish the monitoring system. Even then, we have a challenge represented by the leftmost bar in the picture. These are parcels that do not contain even a single full Sentinel-2 pixel. One might argue that even though there are indeed a large number of “problematic” parcels, these represent only a small percentage of the total area (and thus distributed funds), 3% to be precise, so this is not really a problem. However, the current set of Check by Monitoring and Area Monitoring System (AMS) rules require all parcels to be monitored and this makes sense to ensure fair and non-discriminatory handling of all farm holders.

Note: the final criteria for AMS QA are not yet set and it might be that small parcels will be reduced in importance.

Distribution of parcels without a single full Sentinel-2 pixels contained in terms of number (left) and area (right).

Addressing the challenge of small parcels

With Sentinel data being out of the picture for monitoring of small parcels, there are a handful of options remaining. The most commonly mentioned one is using Geotagged photo application, where the Paying Agency asks the farmer to take several pictures of the field using their mobile phones. This sounds simple, but introduces a new administrative burden for the farmer, the hated red tape, and might represent a significant technological obstacle for the older generation. It is also time consuming, both for the farmer as well as for the Paying Agency, whose staff then has to go through hundreds of thousands of good-or-bad photos. On the other side of the spectrum is the rapid field visit option, Paying Agency taking most of the burden, but this makes the overall process extremely expensive.

Using Earth observation data seems to be the most feasible way to perform wall-to-wall monitoring. And so far the only feasible option we’ve found, aside from Sentinel, is Planet’s PlanetScope monitoring product, which provides systematic, near-daily imaging of the entire Earth’s landmass. The alternative very high-resolution satellite providers require tasking and are not technically feasible. We have experimented with one of them and they were not able to provide consistent weekly products even for an area 10x10km due to competitive tasks in the area. But if they were, this option would probably be way too expensive. Even in the foreseeable future this will probably not change. Several startups have announced plans for (sub)meter-level resolution and daily cadence, but it will take several years for them to get there. If they actually do manage. And then costs will probably remain problematic. Sentinel-2 Next Generation will presumably do the job, but this is at least five to ten years away.

Luckily, PlanetScope does seem to do the job well!

PlanetScope ortho imagery has 3.125m pixel resolution, which means that an area covered by one Sentinel-2 pixel is covered with around 10 PlanetScope or Planet Fusion pixels.

Sentinel-2 L2A (left) and Planet Fusion (right) true color visualization of an area in Slovenia with plenty of narrow and elongated parcels.

Experience with PlanetScope

At the beginning of the project, we started with PlanetScope for two reasons: Planet Fusion was not available at that time, and we could easily get it through our Third Party Data Import API. The latter meant that to start using the PlanetScope data we only had to order it. As for the rest, we could reuse existing pipelines to build signals datasets, train models, and, at the end, execute marker computation and assign traffic light decisions. We were thus able to focus on more difficult and important tasks that deliver value to the client. We didn’t have to spend time building yet another ETL pipeline.

That being said, when it comes to any proprietary data the first stage of the project — getting the data — is challenging, due to contractual matters. The process starts with many questions, and the “What items should we order?” was our first one. Questions that are not technical and to which one would typically like to get an answer to in a data-driven way. Unfortunately, due to the price tag attached to the imagery, the amount of experimentation one can afford is limited. To keep the costs of data for all small parcels within our budget we had to limit the spatial and temporal components of the data that we’ve ordered:

Original parcel geometries in red were buffered and merged into a single multi-polygon request geometry in blue.
  • We request data only for the area covered by small parcels. In order to reduce the number of requests and orders we grouped several parcels into one multi-polygon and applied a 5m buffer. See image to the left with geometries of one example order.
  • For each requested geometry, we were able to order around 250 observations (our time interval of interest was between January 1 and October 31, 2019). We reduced this number significantly by rejecting all scenes using the information contained in the metadata, like cloud cover, percentage of clear, snowy, etc. that were accessible before making an order. On average, 115 observations per requested geometry remained, which was still high. We’ve noticed that often multiple items are available per request geometry on a particular date. This can happen because requested geometry is within the overlap area of multiple scenes from the same satellite, scenes from two different satellites, or a combination of the two. In such cases, we reject all but one observation. This reduces the number of observations to 70 per request geometry, which was still too high a number for us. We reduced it to around 45 observations per request by requiring that the gap between two consecutive items in the time series is at least 3 days long. Note that with the “Area under management” business model by Planet, these optimizations would no longer be as important.

An example Normalized Difference Vegetation Index (NDVI) time-series of a meadow is shown below. The observations show an overarching curve, but the individual data points don’t always align with that curve. Detecting the markers we seek depends on the temporal pattern of the signals, so such noise in the time series presented a problem.

NDVI time series of a FOI claimed as a permanent meadow. Drop in the beginning of June is most probably due to a mowing. In the second half of the season the time series is very jagged and is difficult to discern any other mowing events. Observations from Dove Classic and Dove-R are shown as squares and circles, respectively.

We dug deeper into the problem and concluded that PlanetScope data would need to be better harmonized for our modelling (in fairness to Planet — this is not an uncommon challenge, even the “gold standard” Sentinels have one). The PlanetScope satellite constellation consists of three different generations of instruments: Dove Classic, Dove-R, and SuperDove. The data points in the time series above drew from first two generations of PlanetScope instruments and were not harmonized. When we ordered the data and ingested it with our API we did apply radiometric harmonization, but the resulting data was still too varied. Our poor man’s solution to this problem was to select observations based on meta-data that lead to lower variance of the signal. In the end, we kept only Dove Classic observations with sensing time around 9 a.m. We had to keep data from lower-quality Dove Classic instrument because the majority (over 70%) of observations from 2019 are from this type of instrument. If we would repeat this in 2022, we would almost certainly get dense-enough time series with SuperDove instrument (with the latest addition of 44 satellites there are plenty in the sky).

NDVI time series of four permanent meadows using all available PlanetScope observations (in gray) or with selected observations only (in blue). Observations from Dove Classic and Dove-R are shown as squares and circles, respectively.

If you want to learn more about the topic of data harmonization, we recommend checking excellent presentations given by Planet at Analysis Ready Data Workshops.

Planet Fusion Solution for CAP

We decided to test Planet Fusion within our area monitoring activities. We got Planet Fusion for eight 24x24 km² large tiles that together cover around one-quarter of all parcels in Slovenia. We were excited to be able to test this capabilty that promises cleaner data for our modelling through radiometric harmonization and the removal of clouds, cloud shadows, and other kinds of noise and provides gap-filled surface reflectance values daily.

The plot below shows Planet Fusion NDVI time series for the same parcel as shown above. As you can see the time series is smooth, which is to some extent by design.

Planet Fusion NDVI time series of the same parcel as shown above using PlanetScope data.

Mowing marker performance

We evaluated the performance of all our markers on both PlanetScope and Planet Fusion data and compared it with performance on Sentinel-2 data. In general, markers perform better on Planet Fusion than on PlanetScope, but differences are, at least in our case, not significant, except in the case of mowing marker. The plot below shows the performance of mowing marker on parcels with manually labeled mowing events for which we have PlanetScope, Planet Fusion and Sentinel-2 data.

The Pareto front of false positive (FP; x-axis) and false negative (FN; y-axis) mowing event detections. Each point corresponds to a unique set of mowing algorithm parameters. We want to get as close as possible to the lower-left corner.

Performance of the mowing marker is comparable on Planet Fusion and Sentinel-2 data. Results on PlanetScope data were lower because of the need for harmonization articulated above. To compensate for the higher variance of PlanetScope NDVI signal we have to increase the thresholds in our mowing marker algorithm to reduce the number of false-positive detections. But increasing the threshold also reduced the number of true positives (correctly detected mowing event). The number of parcels with detected mowing events using PlanetScope is thus lower by almost a factor a two when compared to Planet Fusion. Since the number of meadows (including small ones) is large in Slovenia this leads to a significant reduction of the number of inconclusive parcels and consequently large efforts the Slovenian National Paying Agency would have to invest to resolve them. This fact alone justifies the usage of Planet Fusion over PlanetScope for this specific use case.

Another way to compare how marker works on different sources is to run it on all parcels for which we have data for all options and cross-check the results. We did this on around one hundred thousand parcels. The confusion matrix below shows how many parcels we observe at least one event by both of the sources, just one of them or neither of them.

Comparison of mowing marker results for over 100,000 parcels using Sentinel-2 or Planet Fusion data.

The results from both sources agree for around 90% of them. More interesting are parcels where results disagree. The algorithm finds at least one event in Planet Fusion data but not in Sentinel-2 for over 12 thousand parcels. Most often this happens when the number of valid Sentinel-2 observations is small as shown with an example parcel below. As you can see there are only a handful of valid Sentinel-2 observations for this parcel so the probability of detecting mowing events is low. This clearly shows the advantage of near-daily cadence of the PlanetScope constellation.

Mowing marker results for the same parcel using Sentinel-2 (left) and Planet Fusion (right) data. Mowing event around end of June is detected using Planet Fusion data but not using Sentinel-2. The true color image chips from Planet Fusion confirm that the detected event is a valid mowing event.

On the other hand, there are only around 4 thousand parcels where the algorithm finds an event with Sentinel-2 but not with Planet Fusion.

Map of the average number of valid Sentinel-2 observations per parcel.

So let’s take a closer look at these two classes of parcels where markers from different sources disagree. To better understand the first group, we need to take a look at the distribution of a number of valid Sentinel-2 observations per parcel shown on the right. As you can see it has a double peak structure. The East and West part of Slovenia have roughly twice as many Sentinel-2 observations as Central Slovenia as shown on the map below.

Number of detected mowing events per parcel with Sentinel-2 (in blue) and Planet Fusion (in orange) data. Left and right plots show distributions for parcels from Central and East and West Slovenia, respectively.

If we then take a look at the distribution of the number of detected mowing events per parcel separately for parcels from Central Slovenia or East/West Slovenia we see that: in East/West Slovenia where the number of valid Sentinel-2 observations is high we observe good agreement between results from Sentinel-2 and Planet Fusion. But in the central part where the number of valid Sentinel-2 observations is low, we see that we observe many more events with Planet Fusion thanks to its denser time series.

Mowing marker results for two parcels using Sentinel-2 (left column) and Planet Fusion (right column) data.

The second group of parcels where we find an event with Sentinel-2 but not with Planet Fusion is also interesting. This seems to happen when the amplitude of NDVI drop is below the threshold in the Planet Fusion time series. We’ve noticed that amplitudes of NDVI drops are in general smaller in Planet Fusion time-series than they are in Sentinel-2. We compensate for this by lowering the threshold for a minimal drop of NDVI by 40% when we use Planet Fusion data. Even with a much lower threshold, there are cases where we don’t detect a mowing event with Planet Fusion data although it is visible in Sentinel-2. It could be that in the Planet Fusion time series the abrupt change caused by a mowing event might be smoothed out too much.

Conclusions

We presented results from studies that we did together with the Slovenian Paying Agency using 2019 data. Studies showed that Planet Fusion is an excellent data source for monitoring small and elongated parcels. Beyond resolution, the value is also showcased in the case of mowing markers, where consistency of individual observations in a time series is of paramount importance. During the 2021 Checks by Monitoring campaign we greatly benefited from this — having the data (technically) available for the full country, there was only a small step required to extend the sample, where we used it — we just had to make sure that it is still within the contracted volumes. In addition to small parcels, we therefore used Planet Fusion also to get a second opinion about parcels where mowing events were expected, but not detected with Sentinel-2 data. By utilizing Planet Fusion data in this way we were able to further reduce the number of inconclusive parcels and thus significantly reduce PA’s administrative burden.

Want to know more?

Join us and Planet for a live joint webinar on Feb 17th to learn how the Slovenian National Paying Agency has benefited from using Planet Fusion. Register here.

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:

--

--