Disease phenotyping for resistance breeding

This repository contains functions and analysis scripts required to reproduce results reported by Anderegg et al. (2019).

Author

Jonas Anderegg
Crop Science Group
ETH Zürich

Dependencies

Scripts build primarily on functions from the R packages prospectr, caret and tidyverse. Some analysis scripts are designed to run in parallel on a cluster, but can be easily serialized. Long computation times should be expected in this case.

Content

Folder Utils contains functions for the pre-processing of spectra, calculation and evaluation of spectral indices, training and evaluation of full-spectrum models, recursive feature elimination and analysis of variance.

Folder Analysiscontains scripts to implement the analysis and obtain results contained in the study.

Utils

001_spectra_utils.R

1. f_spc_smooth applies the Savitzky-Golay smoothing filter to raw spectra (wrapper for prospectr::SavitzkyGolay)
2. f_spc_avg averages replicate measurements
3. f_calc_sicalculates spectral indices
4. f_scale_si scales spectral indices and predictions obtained from full-spectrum model to range from 0 to 10
5. f_spc_bin computes average values of a signal in pre-determined bins (wrapper for prospectr::binning)
6. f_spc_trim removes noisy parts of the signal in pre-determined ranges
7. Several data wrangling helper functions.

002_pls_utils.R

1. f_splsda fits PLS-DA and S-PLS-DA models for classification into healthy and disease canopies. This is essentially a wrapper for functions of the mixOmics package.
2. assess_mod_perf evaluates model performance on the main experiment.
3. assess_mod_perf2 evaluates model performance on FPWW023 (used to quantify model robustness over time).

003_dynpars_utils.R

1. get_dyn_pars extracts dynamics parameters and plots model fits for each plot and spectral index. This function by default fits a Gompertz model with asymptotes constrained to 0 and 10. This is a wrapper for nls_multstart::nls_multstart.
2. get_dyn_pars_lin extracts dnyamics parameters for each plot and spectral index using linear interpolation. This function should be used if the temporal resolution of spectral measurements does not allow to fit parametric models.
3. Various helper functions called by 1. and 2.

004_plot_predobs.R

1. assess_mod_perf extracts various model performance metrics.
2. plot_predobs creates predicted vs. observed plots.
3. Helper functions.

005_rfe_utils.R

1. perform_rfe performs recursive feature elimination.
2. tidy_rfe_output gathers results of resamples.
3. plot_perf_profile creates a performance profile plot based on the tidy output of perform_rfe.
4. Helper functions

Analysis

1. 01_prep_spcdat.R Prepare spectral datasets for subquent analysis: (1) spectral index datasets, (2) spectral datasets for calibration and validation of (S)PLS-DA models.
2. 02_pls_class.R Fit (S)PLS-DA models, extract variable importance, assess "internal", "general" and "dynamic" model performance.
3. 03_extract_dynpars.R Fit Gompertz model (optionally different models) and linear interpolation to spectral indices, extract model and dynamics parameters.
4. 04_subset_selection.R Select subsets of sensitive and insensitive spectral indices.
5. 05_generate_preds.R Extract spectral-temporal features.
6. 06_class_spctemp.R Fit and evaluate classification models (healthy, non-inoculated plots vs. disease, inoculated plots) based on spectral-temporal features.
7. 07_class_spctemp_rfe.R Perform recursive feature elimination for classification models.
8. 08_regr_spctemp.R Fit and evaluate regression models to predict disease severity based on spectral-temporal features.
9. 09_regr_rflt.R Fit and evaluate regression models to predict disease severity based on spectral reflectance at t3.
10. 10_regr_spctemp_rfe.R Perform recursive feature elimination for regression models.
11. 11_feat_validation.R Validation of the most predictive feature(s) on a test dataset from 2016.