Model data augmentation
Data augmentation creates randomized variants of training spectra. It is disabled by default and affects training inputs, not validation, testing, or inference inputs.
Open a draft model version and select Overview > Advanced Settings > Data Augmentation. Set Enable data augmentation to Yes, then configure one or more methods.
Available methods
| Method | Effect | Controls |
|---|---|---|
| Random Cutout | Sets one contiguous band-value segment to zero | Probability, Mask size |
| Smooth Signal | Applies a moving-average filter along band order | Probability, Window size |
| Scale Signal | Multiplies a spectrum by one random factor | Probability, Scale min, Scale max |
| ZScore Normalize | Standardizes each selected spectrum independently | Probability |
| Min Max Normalize | Rescales each selected spectrum independently to 0-1 | Probability |
| Shift Signal | Circularly rolls values along band indices without changing wavelengths | Probability, Shift max |
| Add Noise | Adds Gaussian noise scaled by the spectrum mean | Probability, Noise factor |
| Flip Signal | Reverses band-value order without changing wavelengths | Probability |
| Magnitude Warp | Multiplies values by a smooth random curve | Probability, Sigma, Knot |
| Window Slice | Crops a contiguous band window and resizes it to the original length without changing wavelengths | Probability, Reduce ratio |
| Baseline Wander | Adds a sinusoidal baseline over normalized band order | Probability, Magnitude, Frequency |
Each probability controls how often that method is applied to an eligible training example. Multiple enabled methods can be applied to the same example.
Use augmentation safely
- Start from a model with a documented validation baseline.
- Add methods that represent variation your sensor or acquisition process can plausibly produce.
- Change a small number of controls at a time.
- Compare validation metrics and spatial predictions with the same evaluation data.
Aggressive transformations can erase diagnostic features or create physically implausible spectra. More augmented variation is not automatically better.
Shift Signal, Flip Signal, and Window Slice move spectral features relative to the unchanged wavelength metadata. Circular wraparound, reversed spectral order, and stretched band windows are generally not physically plausible acquisition effects for wavelength-calibrated hyperspectral data. Avoid these methods unless a task-specific study validates the exact distortion. Random Cutout also writes zeros rather than a missing-data mask; use it only when that representation matches the intended robustness test.