This fixed generator creates 30 independent spectra in each of four toy classes. A broad brightness nuisance, smooth latent variation and small noise accompany a class-specific absorption-like dip. It is not a sensor simulator or a field HSI benchmark. Samples 1–20 within each class train the learned methods; the last 10 are held out. A real HSI study needs splits by appropriate specimen, field, scene or acquisition groups.

The mean and optional per-band standard deviation always use only the 80 training spectra. PCA, Fisher LDA and the autoencoder fit those 80 rows. LDA receives only training labels. All 120 points participate in each stored t-SNE/UMAP fit; their hollow marks are original split identifiers, not out-of-sample projections. Class colours are annotations for every unsupervised method.

Neighbour overlap is the average fraction of five nearest neighbours in preprocessed 48-band space also found in the displayed 2D coordinates, over all 120 points. It is a descriptive geometry diagnostic, not accuracy, a test score or evidence of physical classes. For one retained component, axis 2 is zero. With more than two retained dimensions, the plot and overlap still use only the first two. Raw-space reconstruction MSE uses all 48 bands and all retained dimensions.

Changing settings after looking at held-out errors turns this toy holdout into a development set. Reserve a genuinely untouched test set for real conclusions. Rotation, reflection, visual cluster area and distances between t-SNE/UMAP islands are not reliable physical explanations.

t-SNE presets: scikit-learn exact solver, random initialisation, 750 iterations, automatic learning rate. UMAP presets: umap-learn, Euclidean distance, spectral initialisation, 300 epochs, one CPU thread. Seeds 7 and 29; all 36 actual outputs and version metadata ship with the source. The browser never pretends to fit these stored runs.