How we made it happen
1. Scanning and detecting the fields’ boundaries
To monitor individual fields to make accurate predictions, it is first necessary to properly delineate the fields. Knowing a field’s boundaries allows us to distinguish different fields, detect field-specific events such as tillage or irrigation, and helps to predict a field’s yield.
The challenges? The sparse amount of available data and its quality. Even though solutions already exist, these are usually hidden behind a subscription. We challenged the client to find and create new data sources that would help achieve the desired impact and implemented a solution ourselves. This is how we did it:
Collect satellite data (using Google’s Earth Engine) to extract fields
Draw boundaries on the fields (through VGG Image Annotator)
Data augmentation (where we increase the amount of data by adding slightly modified copies of existing data)
Create an instance segmentation model (localization of specific objects and the association of their belonging pixels), using PyTorch Vision
Train (using transfer learning +fine-tuning)
Evaluate the results
Our model is now 95% accurate in scanning and detecting the boundaries of up to 1.000 fields every hour. Based on this information, Our client makes accurate decisions on rewarding and incentivizing farmers for their efforts in reducing CO2.
2. Detecting the applied tillage practice
To ensure that the reported carbon emissions reflect reality, our client was looking to use machine learning and remote sensing data to analyze the actual practices applied to fields, such as tillage, the use of cover crops, and the presence of irrigation systems.
One of the main challenges in the project was the limited amount and quality of available data. We tackled this challenge by combining various sources and methodologies:
Collect labelled data on tillage practices
Collect satellite data to extract fields labelled with a certain tillage practice
Extract useful properties from the data
Create a Machine Learning model to make tillage predictions
Evaluate the results and re-iterate
Our model is now 88% accurate in detecting the applied tillage practice, which helps enormously to reward & incentivize farmers for their efforts in reducing CO2.
How does this apply to other industries?
The Field Boundary Detection project can be applied to any industry where you have to detect objects and their boundaries. Here are some examples:
Urban control: distinguishing between green and urbanized areas (e.g. between parks, trees, lawns, and buildings, parking lots, …), detecting houses with green rooftops etc.
Manufacturing: helping pick-up robots to detect objects.
Healthcare: helping the detection of diseases (e.g. cancers) or broken bones during radioscopy.
Safety/government: detecting drivers using their phones while driving.
Real estate: finding open areas for new buildings.
Any task where you want to detect, outline and classify object instances in an image is where we can create impact for you.
Want to create a similar impact at your organization? Book a call with our AI specialist below.