Satellite remote sensing priorities for better assimilation in crop growth models : winter wheat LAI and grassland mowing dates case studies

In a context of markets globalization, early, reliable and timely estimations of crop yields at regional to global scale are clearly needed for managing large agricultural lands, determining food pricing and trading policies but also for early warning of harvest shortfalls. Crop growth models are often used to estimate yields within the growing season. The uncertainties associated with these models contribute to the uncertainty of crop yield estimations and forecasts. Satellite remote sensing, through its ability to provide synoptic information on growth conditions over large geographic extents and in near real-time, is a key data source used to reduce uncertainties in biophysical models through data assimilation methods. This thesis aims at assessing possible improvements for the assimilation of remotely-sensed biophysical variables in crop growth models and to estimate their related errors reduction on modelled yield estimates. Assimilated observations are winter wheat leaf area index (LAI) and grassland mowing dates derived respectively from optical (MODIS) and microwave (ERS-2) data. These observations have been assimilated in WOFOST and LINGRA growth models. Observing System Simulation Experiments (OSSE) allowed assessing errors reduction on yield estimates after assimilation for different situations of accuracy and frequency of remotely-sensed estimates and for different assimilation strategies, indicating expected improvements with the currently available and forthcoming sensors; it supports also guidelines for future satellite missions. A main finding is the fact that yield estimate improvements are only possible for assimilation strategies able to correct the possible phenological discrepancies between the remotely-sensed and the simulated data. These phenological shifts arise mainly from uncertainties on the parameters and initial states driving the phenological stages in the models but are also due, in some situations, to lack of pixel purity because of the medium resolution of sensors such as MODIS. This thesis also identifies some of the main sources of uncertainties and assesses their impact on assimilation performances. The impact of water content and biomass on SAR backscattering of grasslands is specifically assessed. The backscattering of grasslands increases with the increases of water content and decreases with the biomass in dry conditions. Based on these results, methodologies to classify grasslands according to land use (mowing or grazing) and to detect mowings are designed and demonstrated. A good classification accuracy is observed (overall accuracy around 80%). Results for mowings detection are a bit lower as half of the mowings are correctly identified but the methodology can be considered as promising in particular in the perspective of very dense SAR time series as currently acquired operationally by Sentinel-1.