Returns a data frame describing every pretrained encoder published
by the edaphos project: its name, the Zenodo DOI, the raster
AoI it was trained on, the number of input channels it expects,
the feature dimension of its embeddings, and the SHA-256 digest
of the hosted artefact.
Details
New encoders are added by the edaphos maintainers on each minor
release and propagated through the package. Users with bespoke
pretrained encoders can bypass the registry entirely by passing a
local .pt path directly to foundation_weights_load().
Examples
foundation_weights_list()
#> name
#> 1 edaphos-cerrado-moco-v1
#> 2 edaphos-cerrado-moco-v2
#> description
#> 1 MoCo v2 encoder pretrained on 50k 16x16 Cerrado tiles (SoilGrids 250m soc/clay/sand/phh2o/bdod + WorldClim 2.1 monthly prec/tavg + SRTM elev/slope), aligned to a 0.01-deg grid. 31 channels in, 64-dim feature embedding out. 20000 InfoNCE steps on an Apple M1 Max MPS; final InfoNCE loss ~1.64.
#> 2 MoCo v2 encoder, same dataset / hyperparameters as v1 but 200000 InfoNCE steps (10x v1's budget). Targets the v1.3.2 roadmap promise of a fair benchmark of the contrastive representation against an undertrained baseline. Final InfoNCE loss expected ~0.4 at step 200000 (was ~1.64 at v1's step 20000). 31 channels in, 64-dim out.
#> n_channels feature_dim proj_dim patch_size
#> 1 31 64 32 16
#> 2 31 64 32 16
#> aoi
#> 1 Cerrado core (lon -53 to -43, lat -23 to -10)
#> 2 Cerrado core (lon -53 to -43, lat -23 to -10)
#> url
#> 1 https://zenodo.org/records/19701276/files/encoder_q.pt
#> 2 <NA>
#> sha256
#> 1 44ace7f78c658b6028f1cf5ccfa624023295e5576f681d0135db64726c6738e8
#> 2 <NA>
#> doi license published_at edaphos_version
#> 1 10.5281/zenodo.19701276 CC-BY-4.0 2026-04-22T21:59:12Z 1.2.0
#> 2 <NA> CC-BY-4.0 <NA> 3.11.0