Metode neuronale multinivel pentru teledetecție: reconstruirea și predicția imaginilor satelitare / Sensor deep learning: reconstruction and prediction of multispectral satellite images

The technological evolution in the field of remote sensing has generated an exponential increase in the number of multispectral images and their usefulness has been demon- strated in a multitude of applications. However, most of these products are covered by atmospheric phenomena or are damaged due to technical defects of the sensors. The atmospheric phenomena encountered can be in the form of clouds, haze or smoke, and the physical damage in the form of noise, stripe lines and in more serious cases even the total absence of a band.Starting from these premises, the thesis pursued the valorization of the information available in a multispectral image for two purposes: obtaining an optimal visualization and reconstructing a missing band. Both research directions have in common the use of neural networks. The first direction proposes a set of different methods to offset the effects caused by ambiguities, fog, light clouds and smoke by transferring relevant information between bands in order to visually reconstitute those parts of the image affected by atmospheric phenomena. The general concept shared by these methods implies a stacked autoencoder that successfully encompasses the information from all spectral bands into a latent representation used for visualization. All methods use Sentinel-2 data, and one of them is also tested on Landsat 8 data. The second direction proposes a band reconstruction spectral based method using a neural network with UNet architecture. The method takes into account the worst case scenario, that of the missing band and based on the other bands, performs the reconstruction. Using multiresolution multispectral images, in addition to the successful reconstruction of the missing band, in some cases an improvement of the resolution was also obtained.