Africa Soil Diversity Atlas

An interactive web atlas for exploring predicted soil bacterial diversity across Sub-Saharan Africa, derived from 16S rRNA amplicon sequencing of ~810 soil samples across 9 countries.

Built with R and Shiny. Deployed on Posit Connect.

Overview

This atlas serves spatially continuous predictions of Hill diversity indices and evenness at 0.05° resolution (~5 km), estimated from Bayesian hierarchical models fitted to soil microbiome data from BioProject PRJNA807934. Users can query predicted diversity values at any location in Sub-Saharan Africa by clicking the map or hovering the cursor over it.

Countries covered

Benin, Botswana, Côte d'Ivoire, Kenya, Mozambique, Namibia, South Africa, Zambia, Zimbabwe

Diversity indices available

Layer	Symbol	Unit
Species richness	q = 0	ASVs
Shannon diversity	q = 1	Effective species
Simpson diversity	q = 2	Effective species
Evenness	logE	-

All layers are served as mean prediction alongside a 90% posterior predictive interval (PPI), reflecting the full uncertainty from the underlying Bayesian model.

Repository structure

.
├── global.R          # Startup: loads rasters, defines metadata & helpers
├── ui.R              # Full-screen map layout with floating panels
├── server.R          # Reactive logic: layer swap, map click, cursor tooltip
├── R/
│   ├── mod_map.R     # Map Shiny module
│   ├── mod_predict.R # Prediction Shiny module
│   └── utils.R       # Utility functions
├── www/
│   └── atlas.css     # Custom CSS (dark forest-green theme)
├── data/             # Pre-computed raster predictions (not tracked by git)
│   ├── pred_mean_D0.tif
│   ├── pred_mean_D1.tif
│   ├── pred_mean_D2.tif
│   ├── pred_mean_logE.tif
│   ├── pred_mean_D0_even.tif
│   ├── pred_ppi90_D0.tif
│   ├── pred_ppi90_D1.tif
│   ├── pred_ppi90_D2.tif
│   ├── pred_ppi90_logE.tif
│   └── pred_ppi90_D0_even.tif
└── sda.Rproj

Note: The data/ directory is excluded from version control (see .gitignore) due to file sizes. Raster files must be obtained separately, see Data below.

Local development

Prerequisites

• R ≥ 4.2
• The following packages:

install.packages(c(
  "shiny", "leaflet", "leaflet.extras",
  "terra", "sf", "bslib"
))

terra requires GDAL and PROJ system libraries. On Ubuntu/Debian:

sudo apt-get install libgdal-dev libproj-dev

On macOS with Homebrew:

brew install gdal proj

Running the app

Clone the repository and place the raster files in data/, then from the project root:

shiny::runApp(".", launch.browser = TRUE)

Or open any of global.R, ui.R, server.R in RStudio and click Run App.

Fix the port to keep a stable URL across restarts during development:

shiny::runApp(".", port = 4242, launch.browser = TRUE)

Data

Raster predictions are archived on Zenodo:

The pre-computed raster predictions are generated by a Bayesian hierarchical model (Student-t likelihood, non-centred parametrisation) fitted in Stan via CmdStanR. The modelling pipeline is maintained in a separate repository.

Raster specifications:

• CRS: EPSG:4326 (WGS 84)
• Resolution: 0.05° (~5 km at the equator)
• Format: Cloud-Optimised GeoTIFF (COG)
• Coverage: Sub-Saharan Africa

To convert local TIF files to COG format before deployment:

library(terra)

tif_files <- list.files("data", pattern = "\\.tif$", full.names = TRUE)

for (f in tif_files) {
  tmp <- paste0(f, ".tmp.tif")
  r   <- terra::rast(f)
  terra::writeRaster(r, tmp, filetype = "COG", overwrite = TRUE)
  file.rename(tmp, f)
  message("Converted: ", basename(f))
}

Deployment

The atlas is deployed on Posit Connect. To push an update:

rsconnect::deployApp(
  appDir  = ".",
  appName = "africa-soil-diversity-atlas"
)

Runtime settings (Posit Connect dashboard)

Setting	Recommended value
RAM	≥ 2 GB (4 GB preferred)
Max worker processes	≥ 2
Connection timeout	60 s
Idle timeout	120 s

These are set under App => Settings => Runtime in the Connect dashboard.

Key design decisions

Single-language stack: the entire pipeline from raster loading to UI rendering is in R, with no Python or JavaScript framework dependencies beyond what Leaflet and Shiny provide.

Pre-computed rasters over on-the-fly prediction: all diversity indices are pre-computed on a fixed grid. The app performs only fast terra::extract() lookups at clicked or hovered coordinates, keeping response time sub-second regardless of model complexity.

Student-t Bayesian model: the underlying statistical model uses a Student-t likelihood rather than Normal, which proved decisive for this dataset: switching reversed several ecological conclusions and improved ELPD by ~1,470 units relative to Normal-likelihood models.

Bivariate evenness model: richness and evenness are modelled jointly, allowing the detection of aridity effects that dissociate the two indices in opposite directions — a result not detectable with univariate models.

Citation

If you use this atlas or the underlying data in your work, please cite:

[Citation to be added upon publication]

The source sequencing data are available at NCBI BioProject PRJNA807934.

License

Code: MIT Data and model outputs: CC BY 4.0