In late 2019, the Ministry of Environment in Denmark urgently needed to calculate the total emission of ammonia. The best way of doing this was getting the total number of slurry tanks and information on the coverage of the tanks, as the emission factors are lower for covered tanks than from uncovered tanks. As there are no official statistics on slurry tanks, the ministry applied to SEGES, which is a part of the agricultural advisory service owned by the Danish farmers. However, even with broad knowledge and access to agricultural data, counting the number of slurry tanks located in 34,000 farms spread over a territory of 42,933 km2 was impossible to be done manually in a reasonable time. Let’s see how geospatial data helped in detecting 26,000 slurry tanks in just few hours.
SEGES needed a tool that will allow them to automate the whole process as much as possible. The method required was a combination of machine learning with existing agricultural GIS-databases. Their choice fell on Picterra, a cloud-based geospatial platform that automates the analysis of satellite and aerial imagery, enabling users to identify objects and patterns (road cracks, damaged roofs, etc.) at scale, anywhere on Earth. SEGES had two sources of data for this project:
- A WMS imagery server covering the whole of Denmark at
25 cm of ground resolution.
- The geolocation of the 34,000 farms to be investigated.
Using these two sources, SEGES worked with Picterra to build a slurry tanks custom detector and to run it over the whole country. Using this custom detector, SEGES was able to detect 26,000 slurry tanks over Denmark in a few hours with only 56 training annotations.
Bowled over by this amazing accomplishment using geospatial data, we engaged into an interesting interaction with Julien Rebetez, Lead Software and Machine Learning Engineer and Monika Ambrozowicz, Product Marketing Manager at Picterra to know more about the project.
1. What are slurry tanks?
Slurry tanks, also called slurry pits, slurry lagoons, or slurry stores, are circular concrete structures used by farmers to gather animal waste and other unusable organic matter that later can be turned into fertilizer. Because of the decomposition of this waste, gasses like nitrogen, potassium, phosphorus, and ammonia are emanating from the open tanks.
2. What was the problem that needed immediate attention?
The Ministry of Environment in Denmark needed to calculate the total emission of ammonia in the whole country. It is a colossal work and it’s hard to imagine to be done manually and without geospatial data.
3. Why was Picterra approached?
SEGES, a part of the agricultural advisory service owned by the Danish farmers that is responsible for that project, needed a tool that would allow them to automate the whole process as much as possible. The method required was a combination of machine learning with existing agricultural GIS-databases. Their choice fell on Picterra, because it is a unique geospatial platform that automates the analysis of satellite and aerial imagery, enabling users to identify objects and patterns (road cracks, damaged roofs, buildings, animals, crops, etc.) at scale, anywhere on Earth. Unlike other tools that are focused on specific applications, like buildings or agriculture, Picterra is flexible and its machine learning model can learn on different types of data. It also works very efficiently at scale and SEGES needed the results from the entire country. That’s why it was the perfect solution for detecting all slurry tanks in Denmark.
4. What role did machine learning and GIS play in resolving the situation?
Dealing with such a colossal amount of data would practically be impossible without remote sensing and machine learning. Manual on-site surveying at such a scale would take months and a massive amount of human and financial resources. At the same time, the specifics of slurry tanks make them suitable to be detected in imagery, at least with the right tool. They are round-shaped objects with only some variety in their appearance (they can be filled or empty and they can have a roof or not). They’re also located next to farms, that’s why the geolocation data of the 34,000 farms was perfect in this project to be used as detection areas restricting the areas where to run the detector.
5. Which data was used by Picterra to solve the problem?
As already mentioned, the geolocation data of the 34,000 farms in Denmark, and a WMS imagery server covering the whole of Denmark at 25 cm of ground resolution. Using these two sources, they then used Picterra to build a slurry tanks Custom Detector and run it over the whole country.
6. What solution did Picterra come up with?
Because slurry tanks are located next to farms, circles around each farm in Denmark were used as detection areas on Picterra. Thanks to Picterra’s iterative workflow that prefers the quality of annotations over their quantity, a very low amount of annotations was necessary: only 56 examples of slurry tanks across 48 training areas. Also, detection areas focused the detector on specific parts of the image (or the WMS) to save processing resources.
7. How much time did it take for the machine learning model to be able to recognize the slurry tanks?
Training the model is actually quite fast: it only takes a few minutes. This fast training time enables iterating over the model. In this particular project, the client started by spending 1-2 hours annotating images on Picterra and then trained the first version of the model. After assessing its performance, some more annotations were added to improve its performance in specific conditions. In total, the multiple iterations on the model took around 2 days, mostly spent verifying the model and improving it.
8. How was the solution implemented and how quickly did it bring results?
Geospatial data helped in detecting 26,000 slurry tanks in just few hours. SEGES had a short deadline and the whole project was delivered in two weeks. By using Picterra, SEGES was able to train a model in just a few days. Picterra provided guidance and support to improve the detector.
9. What’s the major unique selling point of this project?
The short turnaround time (2 weeks from purchase to delivery) and the scale (more than 1TB of imagery processed) would not have been possible with any other platform than Picterra. As already mentioned the very low amount of annotations that was required (56 annotations) for a project at the scale of an entire country is unique.
10. Anything else you would like to share.
Doing large scale projects with an interactive heatmap that can be shared with anyone is now possible and available to all Picterra users. What is also worth mentioning is that possible applications of Picterra are endless – detectors built on the platform have been used in city management, precision agriculture, forestry management, humanitarian and disaster risk management, farming, etc. Using the platform is 90% less expensive than data science services and it doesn’t require any computer science skills.
Click here to see an interactive map of slurry tanks in Denmark: http://cloud.picterra.ch/public/showcase/seges/index.html