Remote sensing data assimilation for in-season wheat yield predictions

In-season information on expected crop yields is important for farmers' crop management and business planning, as well as for the entire agricultural and food sector. In addition, timely information on possible extreme yield losses in specific production regions allows early decisions in European agricultural policy, e.g. on possible aid payments to producers. Yield losses are mainly caused by adverse and extreme weather conditions such as heat, drought, late frost, heavy rainfall and floods as well as by pests and diseases. Such events are difficult to predict and their actual impact on yields depends on a variety of factors. With ongoing climate change, such adverse conditions and the risk of yield losses are likely to increase (Lüttger and Feike, 2018).

Process-based crop simulation models (CSM) simulate crop growth, development and yield formation, taking into account local soil and weather conditions and potential abiotic stress. For local applications, where actual growth conditions and crop management (e.g., sowing date, cultivar, fertilisation) are known, CSM can be utilized during the season to provide insights into expected crop yields. However, for large-scale applications these information are mostly unknown, which hampers site-specific yield predictions on regional or even national scale. Furthermore, the accuracy of site-specific weather and soil data is limited and actual growing conditions may differ from those assumed in a CSM-based assessment based on such data from national databases.

Point-specific current data on actual crop status derived from remote-sensing can be used to fill those data-gaps and inaccuracies (Guo et al., 2018). Utilizing the newly available high-resolution hyperspectral data from the Environmental Mapping and Analysis Program (EnMAP), a radiation-transfer-model is used to derive pixel-specific state variables of LAI. The actual LAI values are then integrated into a CSM for in-season crop yield predictions on pixel-level. As remote sensing derived crop status data will only be available in the second half of the project phase, we will first use existing observation data from field experiments to mimic the remote sensing data and establish two common data assimilation routines (ensemble Kalman and particle filter) and respective processing pipelines in an ex-post modeling study. After evaluation of the most promising data assimilation technique, the approach will be extended to develop a German-wide winter wheat yield forecast for the current season by using climate forecast data from the DWD. The approach can later be extended to other crops and crop state variables.