Learning from yields: Prevailing features for winter wheat yield variability and the role of farmers’ management decisions

The relationship between food security and climate change is a central concern for policymakers and society at large. Temperature fluctuations and extreme weather events significantly impact agriculture, notably affecting yield production. Effective management measures that enhance resilience of crop production to abiotic stress are thus highly important. This requires an appropriate understanding of the predominant stressors and their temporal impact on yield formation under given pedo-climatic conditions. Designing future climate-smart management systems will strongly profit from an appropriate evaluation of current yield variability, identifying the main underlying environmental and management related factors. Therefore, the two key questions addressed in this study are:     

• At which temporal stage does crop development indicate differentiation in biomass growth that impacts the attainable final crop yield?
• Are the distinctive crop growth and yield patterns in a region predominantly driven to environmental site effects (soil type, rainfall, temperature) and to what extent farmers’ management decisions (pre-crop, cover crop, seeding time) can influence the site-specific natural drivers? 

In recent decades, multiple approaches have been used to analyze factors driving crop yields, from classical replicated field trials over plot scale agroecosystem models to remote sensing-based machine learning approaches. This work is centered on a georeferenced polygon dataset, containing fields from 0.1 to 16.6 hectares in Lower Austria focused on the country's predominant staple crop, winter wheat, between the years 2013 and 2020. In total, the dataset contains 541 entries with winter wheat yield data and detailed management history of the respective fields. Using different types of feature selection techniques, from classical machine learning (i.e., random forest) to recent techniques (i.e., guided regularized random forest), we aim to (i) analyze the temporal growth pattern and extract the yield determinant features as well as their specific timing from several remote sensing derived indices (e.g., Enhanced Vegetation Index (EVI), Normalized Difference Vegetation Index (NDVI)), and (ii) the role of the site specific pedoclimatic information (e.g., surface air temperature, rainfall, soil data) as well as management data (e.g., previous crop, cover crop, seeding time, tillage type). 

Based on the most promising feature models, we will map the expected winter wheat yield variability for Lower Austria and evaluate yield predictability with regional winter wheat yield data from Lower Austria at NUTS3 level between 2015-2022. Since crop-specific crop yield maps are not currently available at the regional level, the validation data will be obtained by intersecting regional yield data and yearly land cover data.

From the results, we expect to provide an improved insight into yield-relevant time periods for winter wheat growth and their interplay with prevailing site-conditions such as soil type based on remote sensing indices. This can contribute to an improved understanding of winter wheat yield formation, thereby providing decision support for more targeted management adaptation and more realistic estimates of expectable management impacts over the unmanageable fate of natural site conditions.