A data-driven high spatial resolution model of biomass accumulation and crop yield: Application to a fragmented desert-oasis agroecosystem

Information on crop yield is important for food security, in particular under the conditions of climate change and growing population worldwide. We developed a new fully distributed, high spatial resolution, model of biomass accumulation and crop yield applicable to a highly heterogeneous desert-oasis agroecosystem. The bulk of required input data is obtained by retrieving pixel-wise biogeophysical variables from a suite of very diverse satellite data. Both temperature and water stress conditions at field-scale are given full consideration, while the model was designed to strike a balance between model applicability and satisfactory characterization of the heterogeneous desert-oasis system to estimate field-scale yield. The development of this model relies on three main innovations. First, the start and end of the growing season were estimated for each pixel by calibrating the high spatial and temporal resolution observations of Normalized Difference Vegetation Index (NDVI) by Sentinal-2 (S2) MSI (Multi-Spectral Instrument) against limited local phenological information. Second, to monitor crop water stress, account taken of irrigation, a process-based water and energy balance model was applied to estimate the actual evapotranspiration (ET). This requires knowledge of soil water availability, which is characterized by downscaling the ASCAT (Advanced SCATterrometer) soil moisture data product. To capture the dominant features of the eco-hydrological conditions in the desert and oasis agroecosystem, ET was further downscaled from the 1 km resolution. Third, likewise the water stress indicator, the air temperature stress indicator was mapped after characterizing the thermal contrast and heterogeneity of the desert-oasis system, by generating time series of air temperature at 1 km spatial resolution using the MODIS (Moderate Resolution Imaging Spectroradiometer) Land Surface Temperature (LST) data product. In the temporal dimension, gaps were mitigated by applying time series analysis techniques to reconstruct cloud-free time series of LST, NDVI, fAPAR and albedo. These innovations add up to a high resolution characterization of crop response to the geospatial variability of weather and climate forcing in the desert-oasis agroecosystem. The model was applied to the dominant crops, i.e., spring wheat, maize, sunflower, and melon, in the oases of the Shiyang River Basin (northwestern China) characterized by a rather fragmented land use. The high resolution of pixel-wise ecohydrological parameters, i.e., crop phenology, temperature stress and water stress factors successfully reflect differences of crops with different phenology and location in the oases. The relative errors for wheat and maize yields compared to the census data are less than 5% at district level. At the county level, the relative errors of wheat yields of Liangzhou, Minqin, Gulang, Jinchuan, and Yongchang equal to 0.87%, 24.2%, 9.7%, 12.5%, and 7.2%. For maize, the dominant crop, the error on estimated yields was less than 5%, except in Gulang. The relative error on estimated yield for sunflower was less than 10% compared to agricultural census data. The relative error on estimated melon yield was 16%. This performance highlights the applicability of the model to estimate field-scale yields in agroecosystems characterized by fragmented land use.