Extraction of agricultural fields boundaries using hi-res satellite imagery

So, you finally got this contract for agricultural analytics. Or, you got this task in your department.

And the first thing that needs to be done is to mark up all those fields’ boundaries. But here’s the catch:

• There are tens of thousands of square kilometers of agricultural fields out there.
• Plenty of them have a complicated curved shape.
• Some areas contain a lot of small seedbeds, less than one hectare each.

Usually, professional map makers outline one field per 20–40 seconds.

Let’s take Rohtak district in India, for example:

• It has an area of 1 745 sq. km.
• A typical size of one field is 0.4 hectares. That means that there are about 350,000 fields out there!
• It will take one year of professional human labor to create a field map of just this one small state! If you need frequent updates — you have to keep an office of full-time cartographers.

Now there is another option — let the machine do this routine work just in one day!

Here you will see how to use this option!

The solution to making your life easier

Yes, we are developing solutions that can automate this routine manual mapping of agricultural fields. Using Machine Learning, we trained neural networks to delineate agricultural field boundaries using high-resolution imagery (the common method is called instance segmentation).

The solution is focused primarily on agricultural fields of small size. To `increase boundaries delineation accuracy, we use high-resolution imagery provided by Maxar with resolution up to 30cm per pixel! Below is a comparison of fields images in different resolutions for one of the regions of India:

Same fields viewed from different satellites.

Above are the images of the exact same territory. Freely available data (on the right) has resolution of 10m per pixel. On the left side is the data provided by the commercial remote sensing data provider. High-resolution allows us to detect agricultural fields that are in use and skip the abandoned farming fields. It’s possible to mark up boundaries of fields with accuracy up to 0.6 meters, excluding foreign objects like houses, power supply lines, windmills and trees.

What does the solution include?

The solution consists of:

1. Artificial neural networks that are pre-trained and can be quickly fine-tuned for agricultural field boundary delineation in a particular country. We use the same approach with buildings.
2. Post-processing algorithm that converts the output of the model to vector geojson
3. Geoalert platform that powers batch processing to scale up to a whole country

How well does the current model perform?

Let’s briefly talk about problem formulation:

• Semantic segmentation problem is about assigning a class to each individual pixel.
• Instance segmentation is about detecting and delineating each distinct object of interest appearing in an image.

Example of semantic segmentation approach (left) and instance segmentation (right) source

Here we will cover only the instance segmentation approach. As a result, we derive an individual mask for each field!

Example of ground truth markup (on the right), example of prediction of our model (left).

Work on the “Indian” model is still in progress. There is a lot of room for improvement. But yet here are some results that we’d like to share:

* We selected three different areas in those countries, and our cartographers marked up ground truth for models prediction comparison.

source

That’s it for now. We look forward to implementing this new model into Mapflow.ai, which is an excellent start if you want to learn more about a modern AI-mapping platform of the 21st century.

Share this post if you want this feature to be implemented sooner! :)