Global wheat normalization. A robust approach based on Growing Degree Days accumulation | Globalización global del trigo. Una aproximación robusta basada en la acumulación de grados día

Global agriculture is mainly driven by climate. Hence, under the different climate change scenarios, it is facing several challenges such as a greater frequency of extreme events (e.g. floods, droughts, frosts, heat waves), a rise in mean global temperatures, and destabilization of crops’ phenology. But also, the global food system has to ensure food availability for a continuous growing population and with higher nutritional demands, increases of the production costs due to fuel volatility, and wars in exporting countries. Thus, the knowledge about where and when a crop season occurs, along with its requirements, extent, and condition; raises as crucial for ensuring food security. Specifically, knowing the timing and the location of a crop season is key to enable further studies (e.g. water balance, predicting yield, or estimating crop requirements), which usually are represented in crop calendars (e.g. tables or maps) and crop type classifications. Crop calendars are a static approximation to crop's phenology that gather regional and national information about sowing, planting, and harvesting dates of specific crops. However, current crop calendars have disadvantages that limit their applicability at fine scales: they are outdated, have regional or national scale, lack of documentation, or a low time resolution. Pertaining crop type classification, the number of works in literature with a global scope and fine resolution (i.e. 10 m) is scarce, with the maps of WorldCereal being an example. Such scarcity has its roots in two main problems related to ground-truth observations: the latitudinal shift of phenology, which limits the spatial representativeness of the observations, and the lack of observations in many regions. In other words, ground-truth observations are constrained in space and time. Because of the above, the main objectives of this thesis are: (a) generating a global crop calendar map for wheat based on existing knowledge and (b) normalizing time-series of wheat signal measured by remote sensing in order to improve the robustness and transferability of global classification algorithms. These objectives are meant to shed some light onto both topics: crop calendars' interpolation and constraints in the ground-truth observations. Nonetheless, each one has its own problematics. As far as crop calendars are concerned, we harmonized four different crop calendars to create our reference crop calendar. Then, in each national or sub-nation region, we assigned its planting and harvesting dates to the areas with the highest proportion of crop. In addition, we had to consider the circular nature of the reference crop calendar, since models usually are unable to deal with this kind of data. Regarding the constraints in the ground-truth observations, we leveraged the relation between temperature and crop development to reduce their associated spatial and temporal autocorrelation, allowing us to apply a random forest model globally regardless of where observations were located. As a result, we not only achieved the objectives of this thesis, but also opened new lines of research that are currently in development: the interpolation of crop calendars and the normalization of crop time-series that can be used to feed classification models with information closely related to the crops, reducing temporal and spatial autocorrelations and enabling global applications.