SAR temporal series interpretation and backscattering modelling for maize growth monitoring

The research objective is to develop a methodology for the extraction of bio- and geophysical variables from Synthetic Aperture Radars (SAR) and for their use in the perspective of maize monitoring in an operational context. SARs interest agronomists because they present some advantages for vegetation monitoring. However, the actual revisit cycle of SARs is not sufficient for crop monitoring. The image processing chain we developed overcomes this issue and meets the 4 requirements for operational crop monitoring: a high temporal resolution, a high geometric accuracy, a short processing time and the preservation of the signal content. From the literature, we know that the interactions between the signal backscattered by the vegetation and by the underlying soil are very complex. To understand these interactions, we carried out very intensive ground campaigns. The resulting data set is very rich. It covers 3 growing seasons during which 30 ERS SAR images and 13 RADARSAT SAR images were acquired and processed. In total, 612 fields, i.e. 581 maize fields and 31 sugar beet fields were located and visited. These field campaigns represent 2500 field visits and more or less 7500 measurements of 8 variables. One of the major outputs of this research comes from the analysis of the temporal behaviour of the SAR signal distribution at both field and regional levels. The SAR signal is analysed by the mean of the per-field backscattering coefficient. Previous results concerning the respective contribution of soil and crop are confirmed. The research also addresses the use of several regional indicators. We point out a drop of per-field variation coefficient averaged at regional level and we link it to the decrease of the infra-parcel variability of the soil roughness and to the progressive masking effect of the crop canopy on different sources of variability. The spatial variability of the ERS per-field backscattering coefficients is related to the variability of the sowing dates. Finally, existing and new versions of the cloud model are calibrated and validated. The cloud model is adapted to account for the data available from the field campaigns. The results show that SAR do not allow the prediction of the maize biomass at the field level but they can be used to give an indication on the crop status at a regional level.