MeiBo002114/FIP1

The FIP 1.0 Data Set: Highly Resolved Annotated Image Time Series of 4,000 Wheat Plots Grown in Six Years

Dataset Details

Dataset Description

We provide time series data for more than 4,000 wheat plots, including aligned high-resolution image sequences totaling more than 153,000 aligned images across six years. Measurement data for eight key wheat traits is included, namely canopy cover values, plant heights, wheat head counts, senescence ratings, heading date, final plant height, grain yield, and protein content. Genetic marker information and environmental data complement the time series. Data quality is demonstrated through heritability analyses and genomic prediction models, achieving accuracies aligned with previous research.

• Curated by: Mike Boss, Lukas Roth, Norbert Kirchgessner
• License: CC-BY

Dataset Sources

• Data: https://doi.org/20.500.11850/697773

  Note that only the aligned inner plot images are contained in this repo, the original image paths point to /data/image and have to be downloaded separately from the ETH research collection.

• Paper: https://doi.org/10.1093/gigascience/giaf051

Uses

We see the largest re-use potential of the presented data set for the development and evaluation of new modelling and prediction approaches in crop genomics and phenomics. The multi-faceted data set allows modelling approaches on various levels:

• Genomic prediction approaches that include genotype-environment interactions: The presented data enhance the data by Gogna et al. by 6 environments, totalling to 14 environments that are characterized by environmental covariates. The presented benchmark of a genomic prediction with random regressions to environmental covariates provides a baseline that novel approaches can challenge.
• Modelling plant growth and development with longitudinal modelling approaches: The four low-level traits canopy cover, plant height, wheat head count and senescence cover the full growing season of winter wheat in 6 environments that are characterized by environmental covariates. Baseline approaches for plant height growth modelling, canopy cover growth modelling and senescence dynamics modelling for subsets of the presented data exist.
• Image-based phenomic predictions and combined phenomic and genomic prediction approaches: The dense time series of images allow training and analysing end-to-end modelling approaches (e.g., deep learning based) that predict target traits such as yield based on images.

Dataset Structure

The data is structured such that each row contains all data for a plot in a given year. The plot can be identified by it's plot_uid which is also available in parts as yearsite_uid, crop_type, etc. If data does not exist for a certain plot it is None, while if time series data does not exist for a certain date it is simply not present.

The image data is in the images and inner_plot_images colums. images are the original images while inner_plot_images are aligned cutouts of the inner wheat plot. Trait data is included as the values in trait_value(s) and the dates trait_date(s). Marker data includes an anyomized genotype_id string, biallelic codes and corresponding metadata strings. Enviroment variables are also included as _value(s), _date(s) and in addition _time(s).

The sowing date, harvest date and harvest_year are included. The data set includes additional data used for the creation of the data set itself such as the alignments.

Dataset Creation

Curation Rationale

Winter wheat provides a crucial share of calories for human nutrition, with global demand steadily increasing. However, crop production faces challenges due to limited resources like water, agrochemicals, and land. Climate change further threatens crop yields, necessitating responsible and efficient resource use.

Crop yields are substantially driven by complex interactions between plant genetics and environmental factors. For instance, genes involved in fruit formation interact with temperatures at flowering, influencing growth and yield potential. Limited phenotyping data is seen as the major reason for the incomplete understanding of such genotype-environment interactions.

HTFP was developed to address this data gap. Imaging HTFP platforms allow researchers to monitor crop canopy development over time, generating dense time series data of plant growth. There are many approaches to process such data ranging from extracting traits at critical time points to modeling growth dynamics and finally using end-to-end methods that directly analyze image time series.

This data set aims to provide a comprehensive foundation for these diverse approaches. Our goal is to foster collaboration between plant physiology, biometrics, and computer vision research, ultimately improving the ability to predict genotype-environment interactions for current and future climates.

Source Data

The FIP at ETH was established in 2015 to collect image time series of crops growing under realistic field conditions. The FIP's cable carrying system is capable of carrying a 90 kg sensor head. The original sensor head, hereafter referred to as the FIP 1.0 head, was equipped with a red, green, and blue (RGB) camera and a TLS, among other sensors. Wheat field experiments were observed using FIP 1.0 over an eight-year period from 2015 to 2022, yielding six years of data collection, with 2015 and 2020 excluded due to incomplete measuring seasons . Up to three times a week, RGB images of all experimental units (so-called `plots') were collected, and plant heights were measured simultaneously using either the TLS (2016, 2017) or drones (2018--2022) . In 2023, the FIP 1.0 sensor head was replaced with a new, multi-view RGB sensor head. The described data set includes all RGB and height data collected in winter wheat experiments up to this replacement.

The area of approximately one ha that the FIP can monitor is divided into six smaller parts (so-called `lots') that are integrated into a crop rotation. The two FIP lots dedicated to winter wheat provide space for ~350 genotypes, replicated once per lot. For the first three years (2016--2018), the GABI-WHEAT panel was grown as the genotype set. From 2019--2022, a subset of the GABI-WHEAT panel was grown in addition to other genotypes . The GABI-WHEAT panel consists of registered genotypes from different climatic regions of Europe . Genetic marker data and MET data from eight year-locations for GABI-WHEAT are publicly available.

The GABI-WHEAT panel was largely superseded by the Swiss breeding set in 2021 . This new set primarily consists of eighth-generation (F8) breeding genotypes. For the Swiss breeding set, genetic marker data exists but remains confidential. The remaining genotypes, linked to specific projects such as INVITE, were present throughout all years but were generally only grown in a single year each . These genotypes currently lack available marker data.

Regular measurements with the FIP 1.0 head were accompanied by reference measurement campaigns as part of several projects. The heading date and senescence ratings were performed to investigate the relationships of senescence dynamics and diseases . Yield measurements taken on the FIP field were combined with data from other locations to train phenomic prediction models . The plant height measurements served as a basis to quantify the temperature response of wheat genotypes in the stem elongation phase . The extracted plant height values demonstrated their usefulness in improving trait extraction methods from longitudinal data .

The images collected were used to quantify canopy covers and examine their relationship to frost damage events using CNNs. Using a combination of drone data and the high-resolution images the rows in the individual plots were identified . In a small subset (375 images), the wheat heads were annotated and the data was integrated into the public global wheat head detection data set . The image-based canopy cover values served as a test data set to evaluate the cultivar-specific extensions of the thermal time concept .

Dataset Card Authors

Mike Boss, Lukas Roth, Norbert Kirchgessner

Dataset Card Contact

Mike Boss ([email protected])