|LATTARI FRANCESCO||Cycle: XXXIII |
Section: Computer Science and Engineering
Tutor: AMIGONI FRANCESCO
Advisor: MATTEUCCI MATTEO Major Research topic
:Deep Learning Approaches for Spatio-Temporal Analysis of Satellite Radar DataAbstract:
The development of increasingly accurate remote sensing instruments together with nowday computing technologies led to an exponential availability of multispectral satellite and airbone data in the spatial, spectral and temporal domains. In particular, satellite equiped 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 EO missions, e.g., Sentinel-1, terabytes of remote sensing data are available everyday, allowing for the developing of data-driven methods. The conducted research aims at developing Deep Learning methods for the automatic analysis of this kind of data, by providing novel solutions to well-known SAR-based complex tasks. The considered problems can be grouped in the macro area of spatio-temporal analysis for change detection, which is the main objective for which SAR data is exploited. However, developing deep learning methods is very challenging with this kind of data, for which obtaining the ground truth required for training can be very difficult depending on the specific task. Thus, one of the attempt of this work is to provide solutions to the lack of ground truth, by studying and developing suitable semi-supervised approaches. Our research started by developing an automatic approach to pre-process SAR images. This images are characterised by the contamination of a multiplicative noise known as speckle, which makes the interpretation of this data very difficult. The problem of reducing the effect of the speckle in these images is called despeckling and it is a well-known problem in the SAR community. We faced the problem by means of Convolutional Neural Networks (CNNs), by designing specific architectures and learning strategies. 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 faced the problem by considering recurrent architectures to model the temporal correlation between measurements on the ground in order to detect points subject to a change w.r.t. the historical trend. Also in this case we had to overcome the problem of the lack of ground truth, which is very difficult to obtain. Semi-supervised techniques have been studied and developed during our research also in terms of segmentation task. Indeed, we faced the problem of identify and classify pixels belonging to water bodies in SAR images. Water bodies segmentation is mainly related to change detection in the context of environmental monitoring, e.g., for the floodings detection, and it a challenging task considering that few course labels are available for training. Finally, our research involved developing novel solutions for interferogram filtering, which allows to obtain accurate measurements of points on the surface from pairs of SAR acquisitions with millimiter accuracy. Summing up, the objective of the conducted research is to investigate the use of deep learning for solving SAR-related complex tasks, by providing innovative methodologies and solutions to well-known problems in this field.