Area Monitoring — Agricultural Activity Throughout the Year
A look back at our Checks-by-Monitoring process in 2021
Written by Matic Lubej and Teo Cerovski.
2021 is over, summer of 2022 is here, and we are just launching a second run of the automated and continuous Checks-by-Monitoring (CbM) process with the Slovenian Paying Agency (PA). Last year we did this for the first time, where we set in place a process for monitoring over 800.000 parcels every two weeks using Sentinel-2 data and the Planet Fusion product by Planet.
By bringing the whole process to life, we have learned A LOT. It has many moving parts, from researching and producing various area-monitoring related markers, to setting up the full stack product for working with the results of the process and sharing them. As our knowledge grew, we shared it with the community in the form of a blog post series, and this blog post is the culmination of it all, showing how the rich information was processed by a well developed application in order to provide the PA with the most relevant information.
The Simulations
“When can we expect the majority of parcels to have a definite outcome?”, “How does the statistics develop through time?”, “How many months of data are required to obtain reliable results?” — These are some of the important questions asked by the PA already during the process, and they can be answered with the help of the continuous monitoring.
For example, meadows represent about half of the monitored parcels in Slovenia, so even a small fraction of non-compliant parcels can in absolute values quickly grow to unacceptable levels for the PA. Such a study is helpful to the PA in order to decide when to start bringing in the experts’ decisions.
In order to prepare for such questions and to generally better understand the situation through time, we performed a simulation of the process on limited time intervals for the year before, since we had the full time series. The study focuses on detecting mowing events on meadow-like parcels from 2019.
First we looked at the cumulative distribution of first mowing events, either labelled or detected. We see that the majority (> 80%) of mowing events occur already by August! This information is valuable because it means that it is not necessary to wait until October to start with the review work.
An additional interesting observation is that there is a consistent lag between the distributions for the labelled and detected mowing events, estimated to ~20 days. This is a consequence of how our mowing algorithm works, where mowing events are found only after the NDVI rises back above the threshold value, while they are sooner obvious to the reviewer.
It is also possible to estimate the overall fraction of meadows where at least one mowing event has been detected. The plot below shows that in 2019 the detection of mowing events started to rise mid June, with the total fraction of meadows also reaching ~80% around August.
Of course, mowing is not detected in all of the meadows even after the full span of the yearly time series is taken into account. This can be due to several reasons:
- meadow has not been mowed,
- mowing detector failure (false negative detection),
- not enough valid observations in the time-series.
While the first reason is completely valid, the remaining ones also have a significant impact on the final results. The second point described a natural property of models, which is “No model is perfect”, and there is not much you can do after optimizing the algorithm and deciding for a false positive (FP)/false negative (FN) rate the PA is comfortable with. If you’re interested in more details about the failure rates of our mowing detector, we encourage you to read the detailed blog post linked below.
The last point from the list above is connected to two problems we’ve encountered. The first is that there is a high fraction of very small parcels in Slovenia, which causes the obtained time-series to be contaminated with the surrounding area of the small parcel. The second problem is the Sentinel-2 orbit positioning, which causes the Central region of Slovenia having approximately half the valid observations compared to the East and the West region. We have decided to use the Planet Fusion product (from Planet) as an additional data source in order to tackle these issues. More info here:
Going Live
End of July 2021 we ran our production-ready application on all available observations of 2021 on 824 k parcels. Then we repeated this until end of November 2021 every two weeks. Each iteration required the following:
- ~ 20 h of signal processing time for two weeks time series of all parcels (including database related work, validation, etc.)
- ~3 h of GPU training time for training out crop- and land-classification models
- ~200 h of CPU time (usually ~3 h on a 60 core AWS machine) for computing the markers and the TLS
Already in the first iteration we had enough data to declare around 60 % of parcels as eligible. With each iteration more data was available and the number of inconclusive parcels dropped further.
The drop was more significant in the first few iterations, e.g. since more data had a significant impact on the model’s confidence, or due to the increased probability of a first mowing activity. The amplitude of this drop had reduced through time, converging to some stable value, like shown below.
While the impact of the markers and models was decreasing, the experts from the PA focused more and more on the remaining inconclusive parcels, utilizing their skills and know-how to further review them as eligible or not.
After incorporating the experts’ decisions into account, the number of inconclusive parcels towards the end of the year reduced by about a factor of two.
A large portion of inconclusive parcels are small as it is difficult to monitor them, they are either too small for monitoring (even with Planet Fusion) or relatively highly contaminated with neighboring pixels. The value of subsidies on such parcels is insignificant. When a sum of inconclusive claims on a single holding falls under a certain amount (set by the PA), a financial threshold is applied and such claims are considered eligible. This step greatly reduced the administrative burden by halving the amount of inconclusive parcels requiring expert review.
What Should Experts Focus On?
It is important to strategically select which parcels to review in the experts’ review process. In order to minimize redundant work, experts should focus on parcels which stay inconclusive until the end, or those that are less likely to be found eligible by the model.
In our study we tried to find patterns which define groups of parcels which tend to stay inconclusive, either based on their crop type, elevation, size, etc. No obvious patterns seemed to pop out during the analysis, or at least the groups weren’t significant enough to be considered. However, it is still possible to underline some guidelines which serve during the review process:
- Don’t review meadows until the very end.
Meadows have a long interval of allowed activity, so it’s more likely that a meadow parcel will be marked as eligible at some later point during the year. The remaining ones can be reviewed at the end of the season. - Wait with the review of very small parcels.
Satellite-based reviews of small parcels can be very inaccurate and much more prone to the reviewer’s bias. Most of these fall under the financial threshold anyway. - Focus on the parcels where the declared/predicted crops are not common.
Crop type information is often used in the TLS where the decision tree is split, which means it can have a big impact on the final decision. At the same time, since such crops are under-represented in the model training, the model is less confident about their prediction, which can lead to fluctuations of the path in the decision tree. These are good candidates for a review.
Across the Map
Visualizing the monitoring process in the form of a map can help to better understand the process and find potential biases in the models. Our methods of handling the data very much rely on the spatial splitting of the parcels into H3 hexagons from Uber. This enables us to quickly process the data from a specific location and provides various high-level information, such as the distribution of parcels across the country. The image below shows Slovenia covered with H3 hexagons, where each hexagon contains a circle with a radius proportional to the number of agricultural parcels inside it. One can quickly observe that the NW part of the country contains fewer parcels due to the mountainous nature of the region, and vice versa for the NE part.
We can observe how the fraction of inconclusive parcel (according to the model) changes through time across the full map if we plot it for each hexagons separately.
While the fraction of inconclusive parcels drops quickly for the majority of the country, we still have regions where the fraction of inconclusive parcels is high, likely due to areas problematic for remote sensing, such as mountains, locations near intermittent lakes, or virgin forests.
Going forward
With the experience and confidence gained over the last year we are now expanding the monitoring in Slovenia to winter & non-winter crops, and adding seven new payment schemes. This required developing new markers and improving the accuracy of existing ones. We have also improved the expert application for a more efficient review of inconclusive parcels. This includes using a geotagged-photo application in the process, which offers the capability to request and provide photographic evidence of eligibility to the experts and farmers, respectively.
We are very excited to see the results of this year’s monitoring campaign, to learn from the process, and to continue improving it in the coming years.
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
- Expert Judgement Application
- Agricultural Activity Throughout the Year (this post)
