Current students


Section: Telecommunications

Major Research topic:
Novel Methods for Information Retrieval from Multitemporal SAR Images

Nowadays an ever-growing interest of private companies and public institutions in Earth Observation (EO) products is registered, indeed the EO market is expected to reach $15 billion by 2026, subject to a Compound Annual Growth rate of about 14% [1]. The European Space Agency alone in 2017 has injected 1.5 billion (26.9% of its budget) in the market, absorbed mainly by the costs of the Sentinel constellations.   
This excitement in the EO market, due to the unparalleled capabilities of EO products in delivering geospatial and geophysical information, is generating a data flood. Consider that Sentinel-1 (S-1) Synthetic Aperture Radar (SAR) constellation is generating free remote sensed data at 10 TB/day rate and by the end of April, about 20 PB of data has been downloaded by the users [2]. This near vertical growth in data availability has been supported by a comparable advance in the computational techniques and resources (i.e. Machine Learning, Deep Learning, Cloud and Parallel computing), often resulting in an unconditioned brute force approach for the transformation of the data in value-added information, against a better understanding of the intrinsic properties of the data itself [3]. 
A deeper knowledge of the data properties allows to point out particular subsets of the data space capable to well describe specific geophysical properties of the Earth surface, nevertheless, pre-digested data might enhance the performance of machine-learned results both in terms of computational costs and accuracy. For instance, S-1 SAR constellation is regularly acquiring data, independently from weather and illumination source, every 6 days all-over Europe at 5-by-20 (range-azimuth) meters spatial resolution in VV and VH polarization. More than one year of full capability observations have been already accumulated and, being a coherent system, through SAR it is observed both the amplitude and the phase of the scattered signal, extending this capability in two polarization, generating a high dimensionality dataset.  
In the light of this unprecedented availability of remote sensed SAR data, this project aims to understand the different contributions carried in the description of the Earth surface scattering properties by each component of the data hyperspace, generated by S-1 SAR observations. Furthermore, specific capabilities of different data hyperspace subset (e.g. temporal coherence, polarimetric coherence, decorrelation time) in discriminating different families of scatterers will be investigated. 
Thus, will be addressed questions such how much information is contained in the different dimensions of the data space and its redundancy or whether it is possible to define new parameters capable of discriminate specific families of scatterers and in conclusion if and to which extent ML techniques might obtain better performances than classical digital signal processing techniques.
This project will deliver a deep understanding, especially of the S-1 data content, possibly allowing a more efficient application of the current Machine Learning techniques, both in terms of time and results accuracy. The principal outcome, in terms of value-added products, will be Land Cover Land Use (LULC) classification maps and change detection maps at unprecedented accuracy.   


[1]        “Global Satellite-Based Earth Observation Market Prospects 2017-2026 - Data and Services Market to Reach $8.5 Billion.” PR    
             Newswire: News Distribution, Targeting and Monitoring, PRNewswire, 20 Dec. 2017,


[3]       L. Zhang, L. Zhang and B. Du, "Deep Learning for Remote Sensing Data: A Technical Tutorial on the State of the Art," in IEEE  
           Geoscience and Remote Sensing Magazine
, vol. 4, no. 2, pp. 22-40, June 2016. doi: 10.1109/MGRS.2016.2540798