Yearly cropland mapping over large areas with high resolution satellite image time series

With human population growth, shifts in diets, biofuel development and climate change, the food supply system is subject to increasing pressures. In this context, timely and dependable information on crop production becomes crucial for market stability and food security. In spite of advances in satellite systems and data processing, there is a disconnect between operational cropland mapping and the state-of-the-art. This thesis seeks to bridge this gap by capitalizing on available land cover maps and by optimizing the satellite inputs. First, priority areas are identified for all countries to strategically allocate future mapping efforts. Second, methods are proposed to enable regular cropland mapping over large areas in the absence of in situ calibration data. The combination of calibration data carefully selected from available land cover maps and spectral-temporal features derived from the satellite image time series yields spatially consistent results with an accuracy that varies depending on the landscape being mapped. The features are stable over space and time and reduce the intra-class variability which in turn improves generalization. Finally, the spatial resolution requirements of the imagery can be anticipated by quantifying the fragmentation of the agricultural landscapes. This relationship allows optimizing the resolution to reduce the data volume and selecting the appropriate imagery to achieve reliable area estimation. These developments are essential building blocks towards an operational cropland mapping system and an improved global agriculture monitoring.