Joint analysis of SAR and optical satellite images time series for grassland mowing event detection.
Résumé
Throughout Europe, grasslands are a major component of the landscape comprising 40% of agricultural land. Permanent Grassland (PM) means land used to grow herbaceous forage crops naturally (self-seeded) or through cultivation (sown) and that has not been included in the crop rotation of the holding for five years or more. PM are major ecosystems associated with high biodiversity which provide a wide range of ecosystem services (e.g. carbon sequestration, water quality, flood and erosion control).
Grasslands have an important carbon storage capacity which is valuable for climate protection. Different studies have demonstrated that grassland managements such as grazing or mowing can cause significant effects on carbon storage in soils. Identifying and mapping grassland management practices over time can thus have important impact on climate studies.
Remote sensing allows a synoptic and regular monitoring through systematic acquisitions of Earth Observation imagery. The emergence of free and easily Sentinels satellite data provided by the European Copernicus programme offers new possibilities for grassland monitoring. Sentinel-1 (S1) and Sentinel-2 (S2) mission acquire radar and optical satellite image times series at high temporal resolution and fine spatial resolution. They fully match the requirements both for yearly and real-time monitoring.
In this work, we target to jointly exploit both data sources to dynamically detect mowing events (MowEve) on permanent grasslands. Thematic related analysis of the datasets will highlight strengths and weaknesses of both optical and radar imagery. (i) S2 appears efficient for MowEve detection, with significant variations in the vegetation status that can be easily detected in the spectral signal extracted from the time series of images. But the temporal revisit of S2 although nominally 5 days is often reduced even by half due to the frequent cloud cover (ii) SAR images acquisitions being independent of illumination conditions or cloud cover allows for systematic acquisitions and revisit rate of 6 days. Data consistency makes S1 data essential during fast phenomena such as MowEve. Yet, radar data appears very sensitive to soil moisture, precipitations and geometrical properties making interpretation of their time series more challenging.
MowEve detection being weakly supervised, the proposed methodology relies on applying traditional change detection strategies on a low-level fused S1 and S2 data representation. Recurrent Neural Networks will be trained to derive yearly or real-time synthetic S2 vegetation indices from both S2 and S1 observations. Furthermore, through attention mechanisms, our proposed RNN architecture will be able to take into account external data (climate, clouds, topography, etc.) so as to dynamically weight at parcel-level the contribution of optical and radar images. Such method will contribute to obtain dense temporal optical profiles without missing data and compatible with MowEve detection.
An experimental evaluation will be carried out on a test site covering an area of 110x110 Km in France (Mâcon region). Object-oriented analysis will be presented based on permanent grasslands derived from the Land Parcel Identification System. The proposed approach will be compared with traditional MowEve methods essentially based on thresholding independently the different modalities.