An increasing number of space missions now regularly and systematically observe the Earth's surface. In terms of optical missions with a decametrical resolution, we benefit from two space missions, Landsat and Sentinel-2, the characteristics of which are summarised in the table below. These missions provide free, high-quality data that is relatively easy to use.

However, the simultaneous use of multiple sources of space data is still rare. For instance, searching all scientific literature published in 2025 for 'Sentinel-2' yields 6462 references, 'Landsat' yields 5569, and 'Landsat and Sentinel-2' yields just 987. This is probably due to the difficulty of combining data with different resolutions, spectral bands and acquisition dates.

A new method for spatio-temporal fusion

In their new article, Julien Michel and Jordi Inglada have proposed a solution to this problem. Imagine you have a year's worth of Sentinel-2 and Landsat-8 data with clouds, irregular acquisitions, different resolutions and spectral bands. Their proposed learning method makes it possible to provide all the data at the same resolution (10 metres), free of clouds, on any desired date throughout the year and for any available band in either sensor.

We will not go into detail here about the architecture of the AI method (Temporal Attention Multi-Resolution Fusion, or TAMRF), but a brief description can be found in this blog post, with more detail provided in the scientific aticle. The self-supervised learning strategy is inspired by that used in language models: a certain amount of data provided by the two satellites is hidden from the tool, which is then trained to estimate the content of this data from the other images in the time series. The model will therefore learn to interpolate the data temporally, spatially and spectrally. They have developed an original cost function with two innovative terms: one that promotes high spatial resolution details in the predicted images and another that teaches the model to ignore uninformative pixels, such as cloudy pixels or areas of missing data. All of this is explained in detail in Julien Michel's thesis (the thesis was succesfully defended on the 2nd of February)

A well validated fusion

Data fusion not only produces beautiful images, but also well-validated data. The accuracy of the reflectances was assessed by withholding data from the TAMRF and comparing the predicted image with the actual image. For instance, when predicting a Sentinel-2 image, the following performance was achieved:

• If a Landsat image was available on the same day, the accuracy was around 0.015 reflectance units. This accuracy is very close to that of Sentinel-2 surface reflectances.
• If no image is available during the month, the performance degrades slightly, ranging from 0.025 to 0.045 depending on the bands. While the method lacks the ability to predict specific events such as the exact date when a farmer will harvest his plot, it still provides an accurate estimate of the order of magnitude.

Currently, no other method offers equivalent capabilities; however, various published methods can perform more specialised subtasks, such as predicting a Sentinel-2 image when a Landsat image is available or the temporal interpolation of a Sentinel-2 series.

Performance comparisons have been carried out for some of these subtasks. Experiments show that a single TAMRF model trained once and for all offers similar or better performance than specialized models, with additional advantages such as robustness to clouds in images and the ability to predict unobserved dates and improve spatial resolution.

A useful fusion

This method should be very useful.

• It will allow the temporal revisit times of two space missions to be combined for improved monitoring of agricultural events, seasonal changes and natural disasters.
• It will provide a valuable alternative to monthly cloud-free syntheses.
• It could even transform satellite design, enabling missions to be carried out with frequent observations at medium resolution (e.g. 10 m) combined with missions involving less frequent (but still regular) revisits at higher resolution (e.g. 2 m).