Glossary

BenchDataset

Abstract base class implemented by every dataset wrapper in torchgeo_bench.datasets. Declares static metadata (bands, number of classes, task type, default split sizes) and exposes a get_dataset() factory returning a PyTorch Dataset for a given split.

BenchModel

Abstract base class implemented by every backbone in torchgeo_bench.models. Sealed forward_patch_features method guarantees that input normalisation is always applied before the subclass-specific _forward_patch_features.

bootstrap

Resampling technique used to estimate a confidence interval around a metric. eval.bootstrap controls the number of resamples; per-dataset CIs are reported as (ci_lower, ci_upper) in the results CSV.

GeoBench V1

The original GeoBench benchmark released as a set of HDF5 files. Datasets in V1 use the m- prefix in the CLI (e.g. m-eurosat) and live under data/classification_v1.0/.

GeoBench V2

The second-generation GeoBench distribution, packaged as one directory per dataset under data/geobenchv2/. V2 datasets use no prefix (benv2, treesatai, pastis, …).

Hydra

The configuration framework used to compose model and run configs. See Configuration for the override syntax and https://hydra.cc for the full documentation.

intrinsic dimension

The geometric / statistical dimension of a manifold of feature embeddings, estimated by methods such as TwoNN, MLE, or lPCA. Optional in torchgeo-bench via eval.intrinsic_dim and the [id] extra; see torchgeo_bench.intrinsic_dim.

KNN-5

5-nearest-neighbour classifier evaluated on backbone-extracted features. Reported as method="knn5" in the results CSV.

linear probe

Logistic regression trained on frozen backbone features. We sweep C over eval.c_range and report the best test-set performance with best_c. Method label: linear.

mIoU

Mean Intersection-over-Union, the primary metric for segmentation datasets. Computed by evaluate_segmentation().

resume mode

When resume=true, the runner skips any (dataset, method, model, config) combination already present in the output CSV. See Results format for the exact key.