|LATTARI FRANCESCO||Cycle: XXXIII |
Section: Computer Science and Engineering
Tutor: DANIEL FLORIAN
Advisor: MATTEUCCI MATTEO Major Research topic
:A Deep Learning Approach 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.
The huge volume of remote imagery raises new challenges for the research community in order to develop new intelligent techniques to extract and understand the information coming from earth
observation. The applications are many, including, but not limited to, automatic target recognition, terrain surface classification, geospatial object detection and 3D land surface reconstruction.
One of the recent research efforts is to exploit Synthetic Aperture Radar (SAR) complex images to extract temporal information from consecutive observations of the same earth region at different times in order to detect small and large surface deformations. Classic Interferometric SAR (InSAR) methods are based on the processing of the amplitude and phase signals of each pixel to detect changes with millimeter accuracy.
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 types 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. In addition classic methods are based on strong statistical assumptions which lead to several approximantion in the results.
This research work focuses on the developing of a new data-driven approach investigating the use of Deep Learning techniques to build a system able to analyze and extract useful information from time series representing the surface changes. The use of Recurrent Neural Networks (RNNs) is the starting point of the research due to their capability of capturing temporal information in sequential data. In particular, Long Short Term Memory (LSTM) networks showed to be very effective in learning long-term dependencies and they are well suited to process this kind of data which is characterized by a high temporal resolution. Furthermore, instead of considering the time serie of each point separately, this work aims to improve the accuracy in change detection by analyzing the correlation between spatially coherent pixels through time exploiting the power of Convolutional Neural Networks (CNNs). This allows to identify correlated regions in the images providing a good basis to solve specific problems such as segmentation or classification.
SAR images are also characterized by the contamination of a multiplicative noise known as speckle noise which makes the processing and interpretation of this data difficult; we aim to build a noise reduction system to be included in the temporal analysis to further improve the accuracy of the measurements derived from the SAR data.
Among the foreseen contributions, this research promotes the use of multi-modal data, e.g., multispectral and hyperspectral optical images captured by passive sensors, which could be complementary information sources to be fused with the one coming from the SAR images.
In short we aim to investigate a new technology to interpret the satellite radar data by correlating the information through space and time, thus providing an accurate data-driven system for extracting salient features to be used to solve specific computer vision tasks or to predict catastrophic events, containing time and computational costs.