ESA title

Speckle filtering through Convolutional Neural Networks

The idea is to develop an AI model capable of filtering speckle noise from Sentinel-1 data. After creating a dataset of Synthetic Aperture Radar (SAR) satellite images containing thousands of time series, we used the dataset to get its speckle-free version by averaging over time (see figure below). The training set was created by applying common statistical models representing noise to the speckle-free dataset. 


The first column shows the noisy image, the second column the AI model output, and the third column the ground truth.

A Neural Network was created to learn how to filter speckle from SAR images. An example of the test on images is shown below. The figure shows the images with the corresponding labels, while the graphs highlight the evaluation of the proposed model by using two metrics. This gave rise to the understanding that the presence of speckle noise leads to bad results for the generative problem proposed for the Seeing Through Clouds project, and hence the reason for the development of this tool. The tool helps improve all the projects based on Sentinel-1 images, and in fact can be used to pre-filter the data faster.


Dataset creation

More building capacity

Subscribe to our newsletter

Share