Parameters data list

This page enumerates the key processing parameters and data items of the embedding pipeline. Defaults are sourced from configs/default.yaml; CLI flags override them at run time (see the Software user manual).

Processing parameters

Parameter	Value (default)	Where	Description
Tile size	256 px → (C, 256, 256)	change.tile_image(size=256)	Fixed tile edge; scenes are reflect/replicate-padded to multiples of 256.
Embedding dimension D	1024	CLAY_EMBED_DIM	Length of the class-token embedding vector (Clay v1.5).
Encoder	Clay v1.5 ViT (frozen)	embed.ClayEmbedder	Production backbone; timm vit_small_patch16_224 is the CPU smoke baseline.
Patch size / grid	8 px → 32×32 = 1024 patch tokens	embed.encode(return_patches=True)	Per-patch tokens for patch-level change maps (~80 m granularity).
Normalization	(x - means) / stds, stats [1, C, 1, 1]	clay_metadata	Per-band standardization with verified Clay statistics (S2 = 10 bands, S1 = 2 bands VV/VH dB).
Embed batch size	32	configs/default.yaml	Encoder forward-pass batch (16 for the change probe encode).
Search index	IndexFlatIP (cosine)	search.build_index	Exact brute-force inner product over L2-normalized vectors.
Search top-k	10	configs/default.yaml (search.py default 12)	Number of nearest neighbours returned per query.
Change metric	cosine (or l2)	change.embedding_change_score	Per-tile bitemporal distance between embeddings.
Change threshold	F1-max over 0.01–0.99 quantiles (99 points)	change.pick_threshold	Operating point selected on the train split (or held-out validation slice for the supervised probe), never on test.
Change label fraction	0.05 (5 %)	change.tile_mask_labels(frac=0.05)	A tile/patch is labelled “changed” if > frac of its pixels changed.
Delta features	abs / signed / concat → (N, D) or (N, 3D)	change.delta_features	Input representation for the supervised change probe.
Probe classifier	LogisticRegression(max_iter=2000)	probe.py	Few-shot linear probe on frozen embeddings.
Probe held-out fraction	0.2, seed=42	probe.heldout_split	Fixed stratified test split, held constant across all shot levels and seeds.
Baseline	ResNet-18, in_chans=10, from scratch	baseline.build_resnet18	Supervised CNN reference (epochs=60, lr=1e-3, weight_decay=1e-4).

Data items

Data item	Format	Schema / shape	Producer → consumer
Embedding store	Parquet	columns id, modality, vector (float32 [D]), optional label; one row per tile	store.save_embeddings → search / probe / change
Stacked matrix	NumPy ndarray	(N, D) float32	store.stack_vectors → FAISS / probe / change
FAISS index	in-memory IndexFlatIP	(N, D) L2-normalized	search.build_index → search.search
Probe artifact	probe.npz	coef (n_classes, D), intercept, classes	probe.save_probe → demo (load_probe)
Demo bundle	release archive	embeddings.parquet + probe.npz	demo.fetch_bundle → Gradio UI

The end-to-end flow that consumes these parameters is described in the DPM Introduction and the Context overview.