The influence of temporal resolution on crop yield estimation with Earth Observation data assimilation

Crop growth simulation models are often used to estimate crop yield.
For most models, this requires crop, water, and soil management information, though this information is often lacking in many regions of the world. Assimilation of Earth observation (EO) data in crop growth models can generate field-level yield estimates over large areas. The use of EO for assimilation often requires a trade-off between spatial and temporal resolution. Spatiotemporal data fusion can provide higher spatial (≤30m) and temporal resolution data to avoid this trade-off. In this study, we evaluated the timing and frequency of EO data assimilation in the Simple Algorithm for Yield Estimation (SAFY) in a persistently cloudy and fragmented agroecosystem of Ethiopia for 2019 and 2020 growing seasons. We used Landsat and MODIS data fusion to obtain frequent and spatially detailed LAI estimates and assimilated at each main maize growth stage to evaluate the effect of timing and frequency of LAI assimilation. The jointing to grain filling stage observations were more important (RMSE = 117 g/m2, rRMSE = 16%) than other growth stages to improve yield estimation. Using LAI estimates at key crop growth stages was more influential than the frequency of LAI estimates. Reasonably accurate yield estimation (rRMSE = 20%) was obtained using the pre-peak growth stage LAI observations, suggesting that the method is suitable for in-season yield forecasting. LAI retrieval errors from EO data, particularly at the early and late growth stages, were the source of yield estimation uncertainty. Therefore, assimilation of other EO-derived biophysical variables and improving LAI retrieval accuracy from EO data could further improve crop growth model performance in smallholder agricultural systems.