PhD thesis proposal: The PhD aim is to exploit time-series of images acquired by the latest generation of Earth Observation sensors for the characterization of agricultural environments. The high spatial and temporal resolutions and the complementarity of the new optical and radar sensors allow both a relatively accurate crop delineation and semantization. The analysis of these dense times series offers new opportunities to monitor and characterize land cover/use changes. Accordingly, this research is focus on the trajectory analysis of these multi-source image time series covering multiple years to study grassland areas both in supervised and unsupervised ways. Specifically, the main goal is to characterize seasonal and long-term trends, as well as abrupt changes and anomalies so as to detect human activities. The interleaved applications are two-fold:
(i) monitoring human activities that are subject to highly controlled rules and European fundings ;
(ii) improving crop classification systems that exhibit low performances for grasslands.
Remote sensing offers the possibility to provide information on landscapes over large scales, thanks to the large spatial coverage and regular revisit frequency of spaceborne sensors. Radar and optical images time-series have already proved to be invaluable and complementary inputs for monitoring land cover/use changes in vegetation and cropland areas. Such relevance assessed with MODIS sensor in the 2000s has been recently exacerbated with the launch of the new Sentinel constellation and missions dedicated to specific environments (such as Venμs for vegetated areas), as well as the spread of free and open policies for the distribution of satellite images (Landsat, Copernicus, Spot World Heritage).
In agricultural environments, grasslands are a land cover class of crops of utter interest since they provide many ecosystem services such as quality feed, animal health or quality biomass (Si et al.,2012; Socher et al., 2013; Lopes et al., 2017). Accordingly, grasslands are frequently affected by human activities (e.g., moving, grazing, (Li et al., 2013; Dusseux et al., 2014)), and subsequently are highly controlled for regulation and European funding issues (Common Agricultural Policy, see(Kleijn & Sutherland, 2003)). Their characterization and discrimination have been barely investigated in the literature (G.mez Gim.nez et al., 2016). Recent works and large scale experiments in several European countries (Ukraine, Czech Republic, Germany) have highlighted the challenges of such a difficult task : the spatial heterogeneity, the distinct phenological cycles, and various anthropogenic interventions should be careful considered and discriminated (Lopes et al., 2017; Stenzel et al., 2017). Existing literature has demonstrated that time series of images (Dusseux et al., 2014; Lopes et al., 2017) are particularly suited for detecting grasslands and documenting human activity, especially in a multi-modal context (Hadria et al., 2009). However, they remain limited to very few epochs (G.mez Gim.nez et al., 2017) while dense datasets are more beneficial, especially with multiple points of view and very high spatial resolution (Schuster et al., 2015). To top it all, when grassland delineation is at stake, agricultural practices are ignored or implicitly handled and when anthropogenic events are detected, existing and accurate grassland maps are available (Hadria et al., 2009; Esch et al., 2014), while both issues are strongly interleaved.
The analysis of satellite image time-series can be roughly divided in two main categories of approaches:
In this work, the main goal is the exploitation of dense optical and radar image times series such as Sentinel-2/-1 and Venμs for grassland monitoring (delineation and event detection) over large areas. Specifically, the work will be performed on intra-annual and inter-annual scales and it will focus on:
The student will be hosted in IGN/LaSTIG lab. in Saint-Mandé and in UMR CESBIO, Toulouse (probable 50% time splitting). LaSTIG gathers 4 research teams with IGN researchers, engineers as well as University Paris Est research. The associated doctoral school is ED MSTIC (Mathématiques et Sciences et Technologies de l'Information et de la Communication) of University Paris Est.
(c) GdR 720 ISIS - CNRS - 2011-2018.