FIP1 - Winter Wheat Field Phenotyping

This dataset is a structured extract from the FIP 1.0 dataset (Boss, Roth & Kirchgessner, ETH Zurich, 2024; doi:10.3929/ethz-b-000697773), prepared for use in causal structure learning analyses of winter wheat growth and yield formation. It was used in the study "Hierarchical Causal Structure Learning" (Hermes et al., arXiv:2511.20021). The data covers more than 4,000 individual wheat plots grown across six seasons (2016–2018, 2019–2022) at the ETH Field Phenotyping Platform in Zurich, Switzerland. Each row corresponds to one plot in one growing season. The dataset contains nine variables spanning environmental drivers and plant phenotypic traits, described below. Environmental variables are provided as hourly time series over the full growing year; phenotypic traits are provided as scalar summaries or sparse time series depending on the measurement type. The data is structured in Parquet format to preserve the nested time series structure. The six splits (train, validation, test_plot, test_genotype, test_environment, test_genotype_environment) follow the benchmark splits defined in the FIP 1.0 HuggingFace release (doi:10.57967/hf/3191). Task: The dataset can be used to study causal discovery methods. Summary: Size of dataset: 4,058 x 9 Task: Causal Discovery Data Type: Continuous Data Dataset Scope: Standalone Dataset Ground Truth: Known Graph Temporal Structure: Time Series Data License: CC BY 4.0 Missing Values: No Missing Values Features: Solar irradiance — Hourly incoming solar radiation measured at 200 cm above ground (W/m²). Recorded by the FIP meteorological station throughout the growing year. Drives photosynthesis and crop energy balance. Precipitation — Hourly precipitation measured at 200 cm above ground (mm). Recorded by the FIP meteorological station. A key water input variable for crop growth and yield formation. Air temperature — Hourly air temperature measured at 200 cm above ground (°C), the standard meteorological screen height. Controls the rate of crop development, including heading timing and grain filling. Days till heading — Scalar. The number of days from sowing to the heading date (the moment when the wheat ear emerges from the flag leaf sheath). A key phenological trait reflecting genotype-specific and temperature-driven developmental timing. Height — Final plant height (m) measured at the end of the growing season before lodging, derived from terrestrial laser scanning (2016–2017) or drone surveys (2018–2022). Reflects both genetic potential and growing conditions. Canopy cover — Fractional ground cover by the wheat canopy (0–1), derived from RGB images taken up to three times per week by the FIP sensor head. Represents canopy development dynamics over the season. Spike count — Number of wheat spikes (heads) per plot area, counted from RGB images at approximately 9 time points per season. A direct component of yield, reflecting tiller survival and ear formation. Yield — Grain yield per plot (t/ha or kg/plot), measured at harvest. The primary agronomic outcome variable. May be provided as raw or spatially adjusted values. Protein — Grain protein content (%), measured from harvested grain samples via near-infrared spectroscopy. An important quality trait for wheat, often negatively correlated with yield. Files: train.gzip: N = 2930. Training data validation.gzip: N = 380. Hyperparameter tuning test_plot.gzip: N = 250. Generalization to unseen plots (same genotypes/environments seen in training) test_genotype.gzip: N = 246. Generalization to unseen genotypes test_environment.gzip: N = 190. Generalization to unseen environments (year-lot combinations) test_genotype_environment.gzip: N = 62. Generalization to both unseen genotype and unseen environment simultaneously Hermes_Graph_FIP1.png: Hermes et al., arXiv:2511.20021 estimated this graph in their paper Hierarchical Causal Structure Learning