Scanworld aims to become a leading provider of space-based hyperspectral imagery.
Their objective is to use satellite images to provide farmers with the most accurate possible information about their crops. They also aim to improve day-to-day farming practice by providing information about how much water is needed, what fertilizers to add, etc...
To provide such an accurate prediction, Scanworld uses satellite or aerial hyperspectral imagery, which contains rich information about the reflection of light from each pixel across the light spectrum. By analyzing the content of these images, it is possible to identify specific properties of the materials studied, and have a precise characterization of the vegetation state, soil types, crops maturity, ...
While the use of hyperspectral imagery has numerous advantages, acquiring them is costly, and the hyperspectral images suffer from a low spatial resolution.
Indeed, most imaging devices found on satellites are multispectral cameras, who provide cheaper images, with a better spatial resolution. Unfortunately, these images lack the data content needed to make the required crop predictions.
This is why Scanworld asked Sagacify to build AI models capable of combining low-resolution hyperspectral images with higher-resolution multispectral images in order to infer the hyperspectral image that we would observe, starting from a multispectral image.
To achieve this goal, Sagacify created models capable of merging multispectral and hyperspectral data without sacrificing spectral resolution or spatial accuracy.
The developed algorithms are capable of extrapolating the complete set of 200+ spectral bands of a hyperspectral image based on the 12 bands of a multispectral image.
With this process, it is possible to create a high-resolution hyperspectral image from a low-resolution hyperspectral image and a high-resolution multispectral image.
Our final model showed to be capable of integrating the information contained in multiple images of the same scene into one composite hyperspectral image of high resolution.
The main advantage of the new model compared to other existing ones is that it is a learning-based approach, which does not require any prior knowledge about the target image, only some training data. This allows to have a generalized model, valid for many different images.
More information about the project can be found in this article: “Agriculture: can Hyperspectral Imagery change the game?”
