modis-lai-forecast

project team: John Smith, David Durden, Emma Mendelsohn, Carl Boettiger, EFI Cyberinfrastructure and Methods and Tools working groups

This repo holds code for a spatially explicit forecasting challenge pipeline to benchmark spatial models using MODIS leaf index data. In this example we focus on locations of wildfire burns and recovery.

🗺️ Example notebook
source

Workflow overview

Site selection

Our goal is to select locations across a variety of environments and burn conditions. Currently we have two sites selected from Monitoring Trends in Burn Severity (MTBS). These shapefiles are available in the /shp directory.

California August complex fire
Colorado East Troublesome

Functions

Functions are stored in the R/ directory.

fire_bbox() reads in a fire boundary shapefile and determines a bounding box for grabbing MODIS data with a padding option.
ingest_planetary_data() downloads data from Microsoft planetary comuputer and returns a gdalcube data cube proxy object.
create_target_file() subsets the data cube, pulls data for a given data and serializes target geotiff to disk.
spat_climatology() creates climatology predictions and serializes prediction geotiff to disk. Predictions are created using an ensemble of historical data within a given month. If historical data is missing, values are treated as NA and bootstrap re-sampling is performed using previous monthly data.
scoring_spat_ensemble() assigns CRPS (Continuous Ranked Probability Scores) and Logarithmic Scores for a given target file and ensemble forecast. Serializes scored geotiff to disk.
na_bootstrap_fun() is used internally for re-sampling during creation of climatological forecasts. The function takes a vector x of (possibly missing) data and fills NA values using a bootstrap re-sampling of non-NA values.

Environment

This project uses renv for package management. Use renv::restore() to load project packages.

Next steps