Facebook IconFacebook IconFacebook IconFacebook Icon

Aerial imageries: Data fusion to enhance agricultural information

Sagacify Logo


Hyperspectral imagery is used to identify specific materials and substances, by analyzing the reflection across the light spectrum of each pixel in images. 

It is often used in remote sensing applications to characterize vegetation and soil types and in medicine to track changes in biological tissue like tumors.

The use of hyperspectral imagery has numerous advantages. It samples hundreds of spectral bands with a high spectral resolution, which enables practitioners to derive information on the object's chemical composition. 

ScanWorld aims to become a leading provider of space-based hyperspectral imagery. Frequently providing farmers with the most accurate information possible about their crops will allow them to respond immediately when problems arise and will prevent them from happening in the first place. Even more interesting is that they aim to improve day-to-day optimization (how much water is needed, what fertilizers to add, etc.)

In the agriculture industry, high performing imagery is a crucial resource. It can provide information on crop health, soil moisture, and rainfall patterns. But acquiring this imagery is costly, difficult to scale, and requires specialized technical knowledge. One solution to this problem is satellites.

Main problems with satellites are that they are expensive and that they take time to be launched. Launching a satellite can take years at a minimum and cost millions of dollars.

Currently, multispectral satellites are available (e.g. the Sentinel-2 multispectral satellites of the European Union's Copernicus program) but they provide less valuable information than hyperspectral systems could. However, the images of those satellites have advantages regarding their spatial resolution, which is quite high.

When analyzing satellite images, one characteristic of the ideal situation is to have ideal data: hyperspectral images (HSI) with  high resolution. 

This kind of image is not available yet, or at least very costly, which means that providing high-quality data to farmers is too expensive.

But what if we could create HSI with high-resolution data, by combining low-resolution hyperspectral images with higher resolution multispectral images containing all the information we desire in the spectral resolution ?


Together with ScanWorld, we wanted to achieve this goal through a proof of concept. 

The goal was to automate the fusion of multispectral and hyperspectral data without sacrificing spectral resolution or spatial accuracy. More precisely, the project aims to develop and train deep learning algorithms capable of extrapolating the complete set of 200+ spectral bands (i.e. hyperspectral) based on the sub-sample of 12 bands (i.e. multispectral) and so create a high resolution HSI at a moment only low resolution HSI et high resolution MSI are available. 

The expected result? High-quality imagery allowing for more accurate insights about what's happening on earth using satellite images.


Our model was able to integrate the information contained in multiple images of the same scene into one HS and HR composite image.

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, only some training data. This allows us to have a generalized model for many different images.

The main goal of this project was to reduce the cost of access to HSI data. We wanted to meet this goal by reducing the number of external data sources needed and making cheaper data available.

We were able to accomplish this through several different techniques and processes. With the elements we had selected to work together, we created a final model that surpassed our original expectations for performance.

Our high-performance model allowed us to fulfill our initial goal—it also allowed us to take on more work within our planning period. As a result, we developed another model that interpolates HSI captured at different periods and creates an intermediate HSI.

Don’t hesitate to read what ScanWorld had to say about this project and how Hyperspetral imagery is a key technology to enable a future of sustainable agriculture in the post by Pietro Ceccato Chief Scientist at ScanWorld and Business Unit Manager at Spacebel in his article: “Agriculture: can Hyperspectral Imagery change the game?”

Related cases