Scribd
Upload
18 views1 page

Sentinel 4

The document describes improvements made for V2 over V1 of a deep learning model, including Europe-wide normalization of satellite imagery, co-registration of input time-series data to reduce misalignments, use of open label datasets, automatic curation of reference data, super-resolution of probability maps, optional rectangularization of parcels, and automation of execution.

Uploaded by

bailogru
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
18 views1 page

Sentinel 4

The document describes improvements made for V2 over V1 of a deep learning model, including Europe-wide normalization of satellite imagery, co-registration of input time-series data to reduce misalignments, use of open label datasets, automatic curation of reference data, super-resolution of probability maps, optional rectangularization of parcels, and automation of execution.

Uploaded by

bailogru
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
You are on page 1/ 1

The complete list of improvements for V2 over V1 is as follows:

Europe-wide normalization of satellite imagery, to improve convergence and


generalization of the deep network;
Co-registration of input satellite time-series to reduce frame-to-frame
misalignments;
Use of open data from the AI4Boundaries and EuroCrops label datasets, which
provide reference data of agricultural parcels for some European countries;
Automatic curation of reference data;
Super-resolution of estimated probability maps for the parcels’ extent,
boundary and distance to boundary;
Optional “rectangularization” of estimated parcels;
Automation of execution.

The following paragraphs briefly describe the aim of each of the applied
improvements.
Band normalization
Fig 3. Visual examples of different normalization methods evaluated against the
convergence of the deep neural network and the achievable accuracy. Normalization
values are computed over Europe.

We have investigated the influence of different normalization approaches on the


convergence of the deep network and on the achieved accuracy for the estimated
parcels. We found that linear normalization using percentiles of the band
distributions (third picture in the top row of Figure 3) performed the best. To
evaluate these methods, we derived band distribution for different land covers over
Europe, resulting in a set of normalization factors that can be applied to a very
large area, therefore improving the generalization of the model. You can find all
details about this research in this blog-post.
Frame-to-frame co-registration

Geometric errors cause misalignments between consecutive satellite image


acquisitions, which produces artefacts in FD results. They affect the accuracy of
the training dataset, since temporal acquisitions are affected differently by
geometric errors. For the same reasons, geometric errors affect the accuracy of the
temporally merged probability image, which can also be seen when visualizing the
results.
Fig. 4. Example of geometric errors for a stack of satellite imagery compared to
the reference polygons for agricultural parcels (top row), and the improved
alignment after co-registering the image frames (bottom row).

An example of the noise introduced by geometric errors is shown in Fig. 4, where


the 3D spatio-temporal stack of the input satellite imagery is cut along one of the
spatial dimensions. The vertical axis shows the time progression over different
image acquisitions, while the horizontal axis shows the spatial index of the other
spatial dimension, in this case northing. The yellow lines represent the boundaries
of the reference polygons denoting agricultural parcels. For the unregistered stack
shown in the top row, misalignments between the boundaries and the underlying
images can be noticed. Such misalignments are compensated for when applying co-
registration of consecutive frames, as shown in the bottom row. More details about
the co-registration research can be found in this blog-post.

You might also like