# Current students

 LATTARI FRANCESCO Cycle: XXXIII

Section: Computer Science and Engineering
The development of increasingly accurate remote sensing instruments together with nowadays computing technologies led to an exponential availability of multispectral satellite and airborne data in the spatial, spectral and temporal domains. In particular, satellite equipped with synthetic aperture radar (SAR) systems allow to collect huge amount of complex images of the Earth's surface through the transmission of electromagnetic pulses to the ground. Being able to extract the information contained in them means being able to understand the environment we live in and to find suitable solutions to complex tasks. Thanks to the satellites revisiting time, and to the numerous Earth observation (EO) missions, e.g., Sentinel-1, terabytes of remote sensing data are available everyday, whose analysis can be accomplished only by using timely and automatic tools. The conducted research provides solutions to well-known Earth observations tasks based on the analysis of SAR-derived data. To this purpose, we focused on the investigation of deep learning methodologies, by demonstrating their effectiveness when employed to solve the problems considered in this research work. One of the main challenges encountered during the conducted work is lack of ground truth data for training. We account for this problem by developing specific data-simulation strategies, depending on the considered tasks, which are parts of the contributions of the presented work. The first task we considered in this research is known as SAR images despeckling. Satellite images acquired by SAR systems are contaminated by a multiplicative noise known as speckle, which makes the interpretation of this data very difficult. The problem of reducing the impact of the speckle in these images is called despeckling and it is a well-known problem in the SAR community. We successfully solve SAR images despeckling, by demonstrating how a deep network can be trained on simulated data and adapted to reduce the impact of the speckle in real images. The other problem we faced in this work, which has acquired increasing interest in the EO community, is the analysis of interferometric SAR (InSAR) time series to detect small and large surface deformations. This kind of analysis finds its application in several problems such as natural hazards detection (e.g., earthquakes, volcanic eruption, etc.), urban planning, vegetation mapping, environment monitoring, and many others. Processing these kinds of data is very challenging due to the uncertainties in the measurements (e.g., the atmospheric delay, orbital artifacts) and noise due to the material characteristics. Moreover the high data resolution makes classic methods both power and time consuming. We demonstrate that training deep recurrent networks is a winning strategy to detect ground deformations in InSAR time series and we further provide a carefully designed procedure to simulate realistic interferometric data. Research on trend change detection in InSAR time series has been carried out during a project funded by the European Space Agency (ESA). Finally, we provide an innovative solution to the phase unwrapping problem, which is the problem of recovering the absolute phase from a phase signal wrapped modulus-$2\pi$. We integrate an existing state-of-the-art unwrapping algorithm based on graph cut, i.e., PUMA, in a standard deep learning pipeline to train a deep network to improve the solution of the reference algorithm, for which a differentiable version has been implemented to allow gradients computation. The resulting NeuralPUMA approach is outperforming and we demonstrate that it has several advantages, among which the increased robustness to the lack of labelled data. Furthermore, for the phase unwrapping work we improved the phase-simulation procedure used n literature to generate more realistic samples. All the proposed methods have been validated in our research through a large suite of experiments and the main results are detailed and discussed opportunely.