Changelog • sfclust

sfclust 1.1.1

Rewrote the “Advanced features” vignette (vg02-advanced-features.Rmd) to use nclust directly instead of an explicit genclust() call, and to demonstrate the path_save/nsave checkpointing arguments; regenerated the corresponding cached vignette data (gaussian-mcmc1.rds, gaussian-mcmc2.rds).
Fixed the “Standardized daily temperature at Canadian stations” vignette (vg12-temp-canada.Rmd): added sort = TRUE to plot_clusters_series() so its cluster numbering matches plot_clusters_map()/plot_clusters_fitted(), and corrected the resulting narrative description of cluster shapes.
Replaced a dead link to the INLA documentation in vg02-advanced-features.Rmd.

CRAN release: 2026-07-28

Added support for raster stars input (x/y grid dimensions), alongside the existing vector-geometry input, including a new raster_adjacency() internal helper and dispatch in genclust()/sfclust().
Added chapa, a real NDWI2 vegetation-index raster dataset from El Chaparrillo, Spain, and a new article vignette demonstrating raster clustering (vg13-ndwi-chaparrillo.Rmd).
Cells with missing/NA response values (e.g. raster cells outside a study region) are now automatically excluded from clustering via response / valid_ids, instead of requiring the user to pre-filter their data.
Added update() methods for sfclust objects: niter continues MCMC sampling from where a previous fit left off, and sample refits the full INLA models for a specific stored sample without further sampling.
Generalized spatial/functional dimension handling (spnames/fnames) so spatial dimension names no longer need to match a fixed convention.
plot_clusters_fitted()/fitted() gained inv_link/aggregation options, returning cluster-level (mean_cluster) and inverse-link (mean_cluster_inv) summaries.

Internal architecture reworked: sfclust() now dispatches to sfclust.data.frame() (core interface) and sfclust.stars() (spatial convenience wrapper), replacing the previous single-path implementation. data_all()/filter_df() replace the old stnames-based long-format conversion.

Fixed corner cases with NULL fitted models and raster inputs.
Fixed adjacency handling for custom/user-supplied adjacency matrices and point-geometry plotting.
Fixed cluster-level aggregation in fitted values and plots.
fitted()/plot() no longer require INLA to be installed when applied to an already-fitted sfclust object (e.g. one loaded from a saved .rds); the inverse-link transform for standard families (gaussian, binomial, poisson) is now computed natively instead of via INLA::inla.link.inv*().

CRAN release: 2025-05-19