Data Augmentation

SemanticSeg4EO implements a custom augmentation pipeline specifically designed for multi-channel geospatial imagery. Five progressive levels are available, each building on the previous one.

Augmentation Levels 

Level	Description
`none`	No augmentation. Patches are fed to the model as-is. Use when your dataset is already large and diverse, or for debugging.
`basic`	Geometric only: horizontal flip, vertical flip, 90° rotation. Safe starting point for most datasets. Recommended default.
`advanced`	`basic` + scaling, brightness, contrast, gamma, per-channel noise. Adds radiometric variability to handle sensor differences and atmospheric conditions.
`aggressive`	`advanced` + random crop, elastic deformation, Gaussian blur, Gaussian noise, coarse dropout (Cutout), channel dropout, minority class oversampling. Stronger regularization for small datasets.
`extreme`	`aggressive` + free rotation (±45°), grid distortion, MixUp, CutMix, channel shuffle (if > 3 bands). Maximum regularization. Use only if the model overfits severely.

Detailed Transform Catalogue 

Geometric Transforms 

Transform	Level	Description
Horizontal flip	basic+	Mirror image and mask left-right (p=0.5)
Vertical flip	basic+	Mirror image and mask top-bottom (p=0.5)
90° rotation	basic+	Rotate by 90°, 180°, or 270° (p=0.5)
Scale	advanced+	Random zoom in/out (0.85–1.15×, p=0.4)
Random crop	aggressive+	Random crop and resize to patch size (scale 0.7–1.0, p=0.4)
Elastic deformation	aggressive+	Local deformations simulating terrain distortion (alpha=120, sigma=12, p=0.25)
Free rotation	extreme	Rotate by any angle ±45° (p=0.3)
Grid distortion	extreme	Grid-based spatial distortion (limit=0.3, p=0.2)

Radiometric Transforms 

These transforms apply only to the image, not the label mask.

Transform	Level	Description
Brightness	advanced+	Multiply all pixel values by a factor in [0.85, 1.15] (p=0.4)
Contrast	advanced+	Adjust contrast around the mean (±0.15, p=0.4)
Gamma	advanced+	Apply gamma correction (0.85–1.15, p=0.3)
Per-channel noise	advanced+	Add Gaussian noise to a single random channel (std=0.02, p=0.2)
Gaussian noise	aggressive+	Add Gaussian noise to all channels (std=0.025, p=0.25)
Gaussian blur	aggressive+	Smooth image with Gaussian kernel (sigma 0.5–2.0, p=0.25)

Dropout and Cutout 

Transform	Level	Description
Coarse dropout	aggressive+	Zero out up to 5 rectangular regions (max 32×32 px, p=0.2). Simulates cloud shadows or missing data.
Channel dropout	aggressive+	Zero out 1–2 random channels (p=0.15). Only when `in_channels > 2`. Simulates missing spectral bands.

MixUp / CutMix 

Transform	Level	Description
MixUp	extreme	Linearly blend two training samples and their labels (alpha=0.3, p=0.2)
CutMix	extreme	Replace a random rectangular region with content from another sample (alpha=1.0, p=0.2)
Channel shuffle	extreme	Randomly permute band order (p=0.1). Only when `in_channels > 3`.

At aggressive and extreme levels, minority class oversampling is enabled: patches containing the minority class (class index 1 by default) are identified and augmented with extra transforms to improve model sensitivity to rare classes.

Choosing an Augmentation Level 

Scenario	Recommended level
Large diverse dataset (>10,000 patches)	`none` or `basic`
Medium dataset (1,000–10,000 patches)	`basic` or `advanced`
Small dataset (<1,000 patches)	`advanced` or `aggressive`
Very small dataset (<200 patches)	`aggressive` or `extreme`
Model overfits quickly	Increase by one level
Model underfits or unstable training	Decrease by one level

Tip

Start with basic augmentation and increase if you observe overfitting (widening gap between training and validation loss after early epochs).