WHAT DOES SENX4 OFFER?
When we talk about extracting information of images from multispectral sensors deployed on satellites, their spatial resolution is one of the main factors that will define their future uses. The ESA constellation of Sentinel 2A satellites offers to the scientific community multispectral images every five days in an open way. But one of the drawbacks of these images is their spatial resolution (10 meters per pixel), limiting their future applications.
There are a lot of methods to increase the spatial resolution of the images. Some solutions try to improve the 13 bands of each satellite, taking all of them to 10 meters using the bands with the máximum resolution. Other solutions develop models with 2X and 4X scaling factors from the 13 bands, thinking that is possible to use this factor to take the bands from 10 meters to 2,5 meters. But nobody has yet created a model to transform the 10 meters bands into 2,5 meters with a contrast of real bands.
In Tracasa, with our SENX4 model we have improved the RGB and NIR bands resolution, going from 10 meters to 2,5 meters with the 4X factor.
HOW HAVE WE GOT IT?
Until now, to increase the resolution were used the traditional technics of interpolation and resampling: linear interpolation, bilinear, cubic and bicubic. In our case, to increase the resolution in a 4X factor, we have used deep learning technics and with cutting-edge convolutional neural networks we increase the resolution of the images without including unwanted patterns, improving the traditional methods.
We have not used sharpening or color adjustments. We avoid generating very synthetic images, far from the reality of the sensor. Our model only uses artificial intelligence to learn the patterns of the data and get an image as real as possible.
VALUING THE IMPROVEMENT
To interpret the results we have used a measuring system that is frequently used in the analysis of images: the PSNR and SSIM combination. These measures are indicators of the quality of the results based on an objective image to achieve. The higher the value in the metric, the better the image will be adjusted to the real result. Taking bicubic like the baseline to improve in the traditional models of interpolation, we have achieved an important improvement in the metrics, that offers, as a result, a more defined image, clearer and more adjusted to reality.
Semantic segmentation (Streets and buildings)
With the new image SENX4 (2,5 meters) we can have a muticlass semantic segmentation more correct and accurate. We improve the segmentation metrics that we got with the data of Sentinel 2A 10 meters.
With SENX4 resolution we optimize the results that we get with clustering algorithms. We add more precision in the spatial limits of the clusters and in their representative average.