Monitoring forest health using hyperspectral imagery: Does feature selection improve the performance of machine-learning techniques?

Research study

See https://pat-s.github.io/2019-feature-selection/ for a detailed description including HTML result documents.

Project structure

📔 code/: R scripts

📔 docs/00-manuscripts/ieee: LaTeX manuscripts

📔 R/: R functions

📔 _drake.R: {drake} config file. Specifies execution order of all steps to reproduce this study.

📔 analysis/: Reporting documents (R Markdown)

📔 docs/: HTML docs created via {workflowr} using the .Rmd sources from the analysis/ directory.

The data is hosted at Zenodo and automatically downloaded and processed when invoking the workflow via drake::r_make().

Reproducibility

This study makes heavy use of the R packages {drake}, {renv} and {workflowr} to streamline workflow execution, manage R package versions and the creation of a research website to complete the study.

By calling drake::r_make() from the repository root, the creation of R objects used in this study is initiated (including data download from Zenodo). Intermediate/single objects can be computed by specifying their names explicitly in drake_config(targets = <target name>) in _drake.R.

While most targets are cheap to compute, the modeling part is pretty expensive. These were run on a High-Performance-Computing (HPC) system and attempting to create those on a local desktop machine is not recommended.

Known Issues

Parts of this work (and some targets) depend on the download of Sentinel2 images. For this task the R package {getSpatialData} was used. After a required refactoring to the latest version of the package in November 2020 (due to outdated/non-working functionality with the initial package version of {getSpatialData} from 2019), the Sentinel2 download is currently broken.

This issue does not affect the recreation of the targets used for the scientific manuscript.

Creating targets with {drake}

(Before creating any target/object, make sure to call renv::restore() to install all required packages.)

Calling r_make() will create targets specified in drake_config(targets = <target>) in _drake.R with the additional drake settings specified.

Important: If you do have access to a Slurm cluster, set options(clustermq.scheduler = "slurm") in _drake.R (around l.73).

Required disk space

The data/ folder will contain data about 5.5GB in size.

Important intermediate targets

Out of the 400+ intermediate targets/objects in this project, the following targets are considered important, i.e. they might want to be created/inspected in more detail.

• task_reduced_cor: List of all mlr tasks used for benchmarking.
• bm_aggregated: Aggregated benchmark results of all models using a 1 meter buffer for hyperspectral data extraction.
• eda_wfr: Creates the report which shows Exploratory Data Analysis (EDA) plots and tables.
• eval_performance_wfr: Creates the report which evaluates the model performances.
• spectral_signatures_wfr: Creates the report which inspects the spectral signatures of the hyperspectral data.
• feature_importance_wfr: Creates the report which inspects the feature importance of variables.
• filter_correlations_wfr: Creates the report which inspects correlations among filter methods.

Note that most reports require some/all fitted models. Creating these (e.g. target benchmark_no_models) is a costly process and takes several days on a HPC and way longer on a single machine.