GPI Workflow

This project builds a reproducible workflow for mapping grassland production intensity (GPI) from Sentinel-2 imagery and field-calibrated environmental measures. The model is calibrated with 2025 field surveys, where sampled zones were assigned observer-based management classes and measured for soil resistance, soil moisture, vegetation height, and plant richness. These field measurements are used to derive a three-class training dataset using KNN. A random forest classifier is then trained on pixels extracted from the labeled habitat polygons and applied to the full raster stack to produce a pixel-level GPI map.

Once calibrated, the saved model can be applied to other years without collecting new field data, provided that matching Sentinel-2 predictor rasters are available for the target year.

Quick Start

1. Check config.R for the calibration year, prediction year, image dates, class order, and expected predictor bands.
2. To rebuild the calibration model and make the configured prediction map, run:

source("scripts/run_calibration_and_prediction.R")

3. To apply the existing calibration model to a new prediction year, update prediction_year and prediction_image_date, add those rasters, then run:

source("scripts/run_prediction_only.R")

Workflow Logic

The project uses two spatial units:

• zone: sampled training polygons identified by polygon_id
• field: the reporting layer used to summarize the final pixel map

Sampled zones carry field measurements and observer labels. Those labels are converted into a weighted KNN-derived GPI target. The labeled zones are then used as training masks, so all predictor pixels inside each zone inherit the zone label. The trained random forest is applied pixel by pixel to the prediction-year raster stack, and the resulting raster is summarized back to fields for polygon-based reporting.

Step	Script	Role	Main outputs
00	scripts/00_export_remote_sensing_from_gee.js	Documents the raster export used by the R workflow. Run only when rasters need to be rebuilt.	<band>_<date>_mosaic.tif
01	scripts/01_build_anchor_training_data.R	Joins sampled-zone geometry, field observations, plant richness, and calibration raster summaries into the anchor training table.	anchor_zone_training_data_<calibration_year>.csv
02	scripts/02_validate_environmental_relationships.R	Checks whether s2rep has interpretable relationships with measured ecological variables.	validation summary CSV and figure
03	scripts/03_define_candidate_gpi_classes.R	Defines the supervised target with leave-one-out KNN over standardized field variables.	candidate_gpi_training_data_<calibration_year>.csv, KNN summary, boxplot
04	scripts/04_train_gpi_classifier.R	Extracts labeled pixels from sampled zones, tunes and fits the random forest with grouped cross-validation, and saves diagnostics plus the selected model.	model RDS and validation tables
05	scripts/05_build_prediction_stack.R	Builds the prediction-year predictor raster stack used directly by the pixel classifier.	pixel_predictor_stack_<prediction_year>.tif
06	scripts/06_predict_pixel_map.R	Applies the saved calibration model pixel by pixel, summarizes the classified raster back to fields, and writes the final map products.	pixel_gpi_map_<prediction_year>.tif, field_gpi_pixel_summary_<prediction_year>.csv, field_gpi_map_<prediction_year>.gpkg, PNG previews

Project Layout

GPI_Project/
├── config.R
├── README.md
├── data/
│   ├── raw/                 # sampled field inputs
│   ├── spatial/             # sampled-zone and full-field geometry
│   └── processed/
│       ├── models/          # fitted random forest model
│       ├── predictions/     # prediction stacks and field summaries
│       ├── rasters/         # Sentinel-2 predictor stack
│       ├── spatial/         # pixel maps and field geopackages
│       ├── training/        # anchor and candidate training tables
│       └── validation/      # diagnostics and model evaluation tables
├── figures/                 # validation and map preview figures
└── scripts/

Inputs

data/raw/environmental_field_data_<calibration_year>.csv

Contains sampled-zone field measurements such as soil moisture, soil resistance, vegetation height, and the observer label.

data/raw/plant_diversity_plots_<calibration_year>.csv

Contains plot-level plant richness observations that are summarized to polygon_id.

data/spatial/sampled_zone_geometry.gpkg

Contains the sampled training polygons used for raster extraction and joins to field observations.

data/spatial/field_geometry.gpkg

Contains the geometry of all fields in the study area.

data/processed/rasters/

Stores the predictor stack expected by calibration and prediction:

• s2rep_<date>_mosaic.tif
• ndvi_<date>_mosaic.tif
• ndwi_<date>_mosaic.tif
• savi_<date>_mosaic.tif
• evi_<date>_mosaic.tif
• msi_<date>_mosaic.tif
• ndmi_<date>_mosaic.tif
• mndwi_<date>_mosaic.tif

The calibration rasters use calibration_image_date. Annual maps use prediction_image_date. Keep the band names and index formulas consistent between calibration and prediction years.

Outputs

Training outputs:

• data/processed/training/anchor_zone_training_data_<calibration_year>.csv
• data/processed/training/candidate_gpi_training_data_<calibration_year>.csv

Validation outputs:

• environmental validation summary and figure
• KNN method summary and class counts
• random forest tuning, confusion matrix, class accuracy, variable importance, model comparison, and selected-model metadata

Model output:

• data/processed/models/gpi_best_model_<calibration_year>.rds

Prediction and map outputs:

• data/processed/predictions/pixel_predictor_stack_<prediction_year>.tif
• data/processed/predictions/field_gpi_pixel_summary_<prediction_year>.csv
• data/processed/spatial/pixel_gpi_map_<prediction_year>.tif
• data/processed/spatial/field_gpi_map_<prediction_year>.gpkg
• figures/pixel_gpi_map_<prediction_year>.png
• figures/field_gpi_map_<prediction_year>.png

Configuration

config.R centralizes the values that must stay consistent across scripts:

• calibration_year and calibration_image_date
• prediction_year and prediction_image_date
• expected predictor bands
• model predictor bands
• ordered three-class GPI levels
• knn_k
• canonical input and output paths

Update prediction_year and prediction_image_date for annual mapping. Update predictor_bands only when the raster stack changes.

Required R Packages

• tidyverse
• sf
• terra
• exactextractr
• janitor
• mgcv
• broom
• patchwork
• caret
• randomForest