Geoalert platform
Published in

Geoalert platform

AI-assisted mapping of crop fields using free Sentinel imagery


Agriculture Digitalization is moving fast and there are more and more companies and developers providing solutions for crops monitoring, asset management, developing fertilizer optimization tools and doing other amazing digital stuff to make agriculture business more effective. Whenever it comes to agriculture management — the fundamental entity is the land parcel and the field that has its certain coordinates and borders and all remote sensing data is referred to the field.

That’s why many agriculture applications start from the Map of crop fields and many of them have such a Map implemented as a core product feature.

The only question is how to place all field boundaries you need on the map?

In the past the only viable option for such large scale mapping was:

  1. Hire a full team of professional cartographers for a season/project
  2. Hope that they will map everything fast enough

But today we can delegate some part of the routine mapping work to AI — to be more focused on the implementation phase and the application of the product to the end user needs.

We would like you to try Agro Mapflow AI for the extraction of crop fields masks from free Sentinel-2 imagery to drastically speed up your mapping work in the coverage of large scale agriculture applications.

Fig. 1 — Example output of field’s boundary extraction (polygon feature) performed by our AI model on Sentinel 2 imagery

So what if we could combine the speed of AI and human accuracy?

Let’s make some napkin math here using testing area in Belgorod region (Russia). (We took it as an example as Russia has quite large cropland parcels that are clear visible in Sentinel imagery)

This area of 2 422 sq. km has:

  • approx. 4 500 polygons of crop fields;
  • each polygon has about 100 vertices;

Handmade mapping from scratch would take by our estimates: 4500 (features) 100 (vertexes) / 350 (vertexes/hour) = 1285 work hours or 160 working days of the human cartographer


  • the Sentinel-2 AI-mapping pipeline processes a single image of this area in ~ 30 min;
  • usually several images are required to cover all fields in AOI (to avoid clouds, shadows, etc…). We used 6 images in our example which makes 3 hours of processing for the whole region.
Example area in Belgorod region (Russia).
Testing area in Belgorod region (Russia)

Did we come up with the final results for the crop fields map? Not yet

While AI overcomes human cartographers many times by speed performance, it makes errors due to instability in some cases, like cloudy areas, shadows from clouds, different seasons and fields patterns , etc.

To be more specific, let’s discuss 2 potential use cases, where the automatization of the mapping of crop fields will be greatly helpful, and discuss a cartographer’s workflow for iterative processing and merging of the resulting field masks.

Case 1: Count the number and the area size of all cultivated fields in the territory

If you are a bit experienced in free and popular QGIS ( — here you go to get started with Sentinel-2 and field masks using our super plugin Mapflow-QGIS. The plugin is powered by Mapflow API and all you need to start working is to create an account and get a token to login.

Here you will find the full user guide on how to search, preview and run the processing of Sentinel-2 imagery in Mapflow-QGIS.

Fig 2. Search results and preview of Sentinel-2 using Mapflow-QGIS plugin

If you get lucky to find the cloudless images that cover your area — it will take about 5 minutes to process the area of the 350 sq. km (Fig. 2)using a single Sentinel-2 image.

Fig. 3 — Segmentation mask of crop fields, extracted by AI from Sentinel-2 image

🤞 Voila! You’ve got the mask and the image in your QGIS layers.

Case 2: Inventorization of all fields in the territory

To get a precise map of all the crop fields cultivated in the territory within a specific period of time (let’s say the vegetation season) you have to do some preliminary work of the analysis of the Sentinel-2 images available. It should be done for 2 reasons:

  1. not to miss objects, that were cultivated at different times (see Fig. 4(a, b))
  2. to ensure the full coverage of the territory (in case of clouds, haze, snow etc. , see Fig. 4(c))
Fig. 4— Field masks extracted by our AI models in different conditions

For this purpose we suggest the iterative AI-mapping workflow based on the composition of field masks extractions from several Sentinel-2 images.

How it works given the testing territory mentioned above:

  • 6 images to process,
  • 1 cartographer*,
  • iterative AI-mapping guideline for the cartographer

As a result it took roughly 3 hours for the algorithm and 2 days for one cartographer to finalise the results (instead of 165 days)!

*Important — the cartographer was not familiar with this mapping workflow before.

A sneak peek to the iterative AI-mapping workflow 👇

Watch this tutorial for a brief visual explanation.

AI-mapping for crop land maps tutorial

The iterative mapping workflow is based on the following principles:

  1. Selection of the successful polygons produced by AI from multiple satellite images
  2. Minimizing the number of clicks for cartographer

This workflow allows speeding up by the orders of magnitude because cartographers are no longer needed to manually draw polygons of fields and to meditate trying to understand whether the field has to be mapped or not. They only need is to delete polygon features at the first step and iteratively fill the gaps.

❗️This workflow could be picked up in hours by any cartographer and we are already developing a series of tutorials explaining how it works.

For more details read the user guideline on our documentation website.

There is also one important thing to be taken into account when we speak about the effectiveness of the mapping of objects in satellite imagery — you need a handy tool for searching and processing images in one bottle.

And once more that’s why we are pleased to announce our special purpose tool that can be used by every cartographer for free.

Using our latest plugin Mapflow-QGIS (1.6.*) it is possible to search, view and process Sentinel-2 imagery with Mapflow AI.

In conclusion:

  • 🤞The new model of Sentinel-2 fields-mapping is successfully implemented into Mapflow and available for all users of the Mapflow-QGIS plugin. You can try it along with the high resolution model that we released before
  • We’ve released Mapflow-QGIS 1.6 🎉🎉🎉 — the main new feature is implementation of Catalog & Processing of Sentinel imagery
    (Sentinel search is powered by SkyWatch API 🙏— cool service and imagery provider, which we will tell much more about next time)
  • ❗️We designed the benchmark of an experimental agriculture territory showing that AI-mapping can speed up the whole process x80 times
  • ❗️❗️❗️We are looking for brave cartographers and developers who are ready to try iterative AI-mapping workflow — look at the guide and if you want to get more materials or suggest some edits, don’t hesitate to contact us
  • ✍️ 🤔 Don’t forget that apart from this service Mapflow provides Hi-res imagery model of fields mask extraction and roads mask extraction that can be useful for your overall agriculture mapping project — check related materials

Related articles:





About Geoalert platform for new applications of the Earth Observation data powered by Deep learning analysis

Recommended from Medium

Wait, Don’t Buy the Extended Warranty on Your Washing Machine

Clothing Retail Spend Time Series

Metaheuristic algorithms

Add cross-browser and cross-platform interactive charts, dashboards, graphs and maps to your…

Demystifying Time Series Concept in making a prediction

SQL for Data Scientist _Coursera

Python to detect your customers preferences

Text Preprocessing: How to handle Emoji ‘😄’ & Emoticon ‘ :-) ’?

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store


We apply Machine learning to automated analysis over Earth observation data

More from Medium

Crop Monitoring using Satellite Imagery — Part I — An Introduction.

Figure that depicts Satellite Imagery

Area Monitoring —  The Challenge of Small Parcels

Vegetation Index: How Is It Used in Precision Agriculture?

Satellite Insights at the Intersection of Megatrends