Abstract
In this work a methodology aimed at land cover mapping over
geographically wide regions, leveraging multitemporal Sentinel-1 SAR
data, is presented. The paper describes an effective way to process SAR
multitemporal data in order to obtain a set of spatio-temporal features,
which well-summarize the temporal patterns of different land cover
classes. Moreover, in this paper an innovative approach to smartly
select training points from an existing Medium Resolution Land Cover
(MRLC) map is presented. Both qualitative and quantitative results over
four regions of interest, with the geographical extension of 100x100
square kilometre, confirm the validity of the proposed procedure and the
potential of SAR data for land cover mapping purposes. These regions,
located in Siberia, Italy, Brazil and Africa, were selected to test the
methodology in completely different climate environments. The
experimental results show that the proposed approach allows to increase
the overall accuracy by 16%, on average, with respect to existing
global products.