Periodic Reporting for period 1 - WomenTech-EO_AI (SpaceSense – Democratizing satellite imagery for Machine learning)
Berichtszeitraum: 2023-06-01 bis 2023-11-30
Climate change is our generation’s biggest challenge that every organization is trying to tackle. And earth observation (EO) is a crucial technology that will help these organizations to adapt and thrive in these uncertain times. But today developing EO systems from scratch is time consuming, expensive and requires very specific skills.
This is where SpaceSense comes in. SpaceSense is a NewSpace company helping organizations build geospatial AI solutions simpler. We provide numerous EO-based services and insights for the Agricultural and Environmental industries.
We are facing two major issues:
-The climate variability between our different customers, and our need to adapt our solutions for each. For example wheat will not grow the same in Tunisia as in Russia. So our models need to account for it.
-Most of our AI models are sensor dependant. A sensor is linked to a specific satellite or constellation of satellites. This is because different sensors have slightly shifted spectral detection ranges, resulting in images that are spectrally shifted. This means that the observations are not interchangeable between similar sensors. This is a major issue because it makes it significantly more expensive to develop solutions for these satellites.
This EU projects aims to solve both these issues, through the development of:
- A template for creating a deep learning segmentation solution for crop type detection using sentinel-2. This template will be able to enable the user to create a new custom crop type detection model with local data in a fully automated fashion. This will also include various aspects of feature engineering, data filtering and processing, pre-trained model and post processing, all coming from internal R&D of SpaceSense, reducing the time-to-market from months to weeks.
-An algorithm and a pipeline that would be able to take two different types of images (A and B), and by using an unpaired Image to Image GAN model, translate image type A into a spectrum range more similar to image type B. This means that every model we create in the future will be able to be done only for one image type, and that'll be able to use this algorithm to apply the model to other sources of imagery, without any need for retraining.
This is where SpaceSense comes in. SpaceSense is a NewSpace company helping organizations build geospatial AI solutions simpler. We provide numerous EO-based services and insights for the Agricultural and Environmental industries.
We are facing two major issues:
-The climate variability between our different customers, and our need to adapt our solutions for each. For example wheat will not grow the same in Tunisia as in Russia. So our models need to account for it.
-Most of our AI models are sensor dependant. A sensor is linked to a specific satellite or constellation of satellites. This is because different sensors have slightly shifted spectral detection ranges, resulting in images that are spectrally shifted. This means that the observations are not interchangeable between similar sensors. This is a major issue because it makes it significantly more expensive to develop solutions for these satellites.
This EU projects aims to solve both these issues, through the development of:
- A template for creating a deep learning segmentation solution for crop type detection using sentinel-2. This template will be able to enable the user to create a new custom crop type detection model with local data in a fully automated fashion. This will also include various aspects of feature engineering, data filtering and processing, pre-trained model and post processing, all coming from internal R&D of SpaceSense, reducing the time-to-market from months to weeks.
-An algorithm and a pipeline that would be able to take two different types of images (A and B), and by using an unpaired Image to Image GAN model, translate image type A into a spectrum range more similar to image type B. This means that every model we create in the future will be able to be done only for one image type, and that'll be able to use this algorithm to apply the model to other sources of imagery, without any need for retraining.
First, we'll go through the activities for the crop type detection model template:
-Creation of a robust multi-day ML model architecture that would work at creating a crop type detection model in most geographical contexts (based on seasons, cloud coverage, field average size...)
-Creation of a satellite data acquisition and pre-processing pipeline that would automatically prepare the satellite data for all fields boundaries provided by the user for the training dataset
-Creation of an interface for the user to provide the local training data and validate their particular settings (type of training, image size...). This was provided as a Jupyter Notebook
-Creation of a testing and inference pipeline to provide all the required analysis to review the model performance, and to deploy the model in a production environment if the model accuracy is satisfactory
-Orchestration of all the elements above so that everything can be done flawlessly by an unexperienced user.
The outcome of theses activities is satisfactory since several customers tested the solution and where able to produce an accurate model for theirs needs in a couple of weeks. One of the user had never worked on an AI project before.
Here are the the activites for the pairing of two images from different sensors (called SpectraSync):
-In depth litterature review to see what is available on the topic
-Definition of a methodology and models to test
-Creation of a training dataset on which to test the models
-Test of the different models and review of the performance
-Creation of a usable final pipeline with the best performing model
The outcome of these activities is mostly satisfactory. The first results are good, but there seems to be a lot of work remaining to be able to fully automate it.
-Creation of a robust multi-day ML model architecture that would work at creating a crop type detection model in most geographical contexts (based on seasons, cloud coverage, field average size...)
-Creation of a satellite data acquisition and pre-processing pipeline that would automatically prepare the satellite data for all fields boundaries provided by the user for the training dataset
-Creation of an interface for the user to provide the local training data and validate their particular settings (type of training, image size...). This was provided as a Jupyter Notebook
-Creation of a testing and inference pipeline to provide all the required analysis to review the model performance, and to deploy the model in a production environment if the model accuracy is satisfactory
-Orchestration of all the elements above so that everything can be done flawlessly by an unexperienced user.
The outcome of theses activities is satisfactory since several customers tested the solution and where able to produce an accurate model for theirs needs in a couple of weeks. One of the user had never worked on an AI project before.
Here are the the activites for the pairing of two images from different sensors (called SpectraSync):
-In depth litterature review to see what is available on the topic
-Definition of a methodology and models to test
-Creation of a training dataset on which to test the models
-Test of the different models and review of the performance
-Creation of a usable final pipeline with the best performing model
The outcome of these activities is mostly satisfactory. The first results are good, but there seems to be a lot of work remaining to be able to fully automate it.
For the crop type detection model template:
The results are very encouraging. We were able to sign several contracts for this solution, and we also discovered some use cases we did not anticipate: Most solutions for crop type detection on the market are focused on the main cereal crops. But for companies looking for uncommon crops, there are no solutions available. Ours makes it possible for them to build custom local models for their specific crops.
The main needs that remains on this product is to make it more user friendly (currently some customers request us to do the custom training), and include more customisation options.
For SpectraSync:
As mentioned above, the results are encouraging, and could help adapt a lot of our AI models to other sensors. But there is still a neeed of further research to make the solution more robust, and being able to implement it in production.
The results are very encouraging. We were able to sign several contracts for this solution, and we also discovered some use cases we did not anticipate: Most solutions for crop type detection on the market are focused on the main cereal crops. But for companies looking for uncommon crops, there are no solutions available. Ours makes it possible for them to build custom local models for their specific crops.
The main needs that remains on this product is to make it more user friendly (currently some customers request us to do the custom training), and include more customisation options.
For SpectraSync:
As mentioned above, the results are encouraging, and could help adapt a lot of our AI models to other sensors. But there is still a neeed of further research to make the solution more robust, and being able to implement it in production.