Loss Functions

SemanticSeg4EO provides a comprehensive set of loss functions for both binary and multi-class segmentation. The loss function list automatically switches when you change the training mode.

Binary Mode Loss Functions

These losses are available when Mode = binary.

Loss Function	Description
binary_dice_bce	Recommended default. Combination of Dice loss and Binary Cross-Entropy (50/50 weight). Handles both overlap quality and pixel-level accuracy.
binary_focal_dice	Combination of Focal loss and Dice loss. Best for strongly imbalanced datasets where the foreground class is rare.
bce	Standard Binary Cross-Entropy. Simple and fast. Works well for balanced classes.
binary_dice	Pure Dice loss. Focuses entirely on region overlap. Insensitive to class imbalance.
binary_focal	Focal loss. Down-weights easy examples and focuses on hard-to-classify pixels. Use with the Focal Gamma and Focal Alpha parameters.
binary_tversky	Tversky loss. Generalisation of Dice that allows control over the trade-off between false positives (alpha) and false negatives (beta).
binary_focal_tversky	Combination of Focal and Tversky losses. Best for very imbalanced data where recall is critical.

Multi-class Mode Loss Functions

These losses are available when Mode = multiclass.

Loss Function	Description
dice_ce	Recommended default. Combination of multi-class Dice and Cross-Entropy. Robust on most datasets.
focal_dice	Combination of multi-class Focal loss and Dice loss. Good for imbalanced multi-class problems.
ce	Standard Cross-Entropy. Classic choice, works well for balanced classes.
dice	Multi-class Dice loss. Handles class imbalance well.
focal	Multi-class Focal loss. Focuses training on difficult examples.
tversky	Multi-class Tversky loss. Control FP/FN balance with alpha and beta.
focal_tversky	Multi-class Focal + Tversky combination.
combo	Multi-loss combination. A weighted sum of Dice, Focal, and Cross-Entropy.

Loss Function Parameters

The Advanced Parameters tab exposes tuning parameters for Focal and Tversky losses.

Focal Loss Parameters

Parameter	Default	Effect
Focal Gamma γ	2.0	Modulating factor. Higher values → more focus on difficult pixels (those close to the decision boundary). γ = 0 reduces to standard cross-entropy.
Focal Alpha α	0.25	Weight for the positive (foreground) class. Increase for highly imbalanced datasets (rare foreground).

Tversky Loss Parameters

Parameter	Default	Effect
Tversky Alpha (FP weight)	0.3	Controls false positive penalty. Lower = more false positives tolerated = higher recall.
Tversky Beta (FN weight)	0.7	Controls false negative penalty. Higher = stronger penalty for missed detections = higher recall.

Note

For Tversky loss: Alpha + Beta should sum to 1.0. The default values (0.3 + 0.7 = 1.0) bias the model towards recall, which is appropriate for rare class detection (e.g., small buildings, water bodies).

Choosing a Loss Function

Scenario	Recommended loss
Balanced binary dataset	binary_dice_bce
Imbalanced binary (rare foreground)	binary_focal_dice or binary_tversky
Balanced multi-class	dice_ce
Imbalanced multi-class (rare classes)	focal_dice or tversky
Need high recall (few missed detections)	binary_tversky (beta=0.7) or binary_focal_tversky
Need high precision (few false positives)	binary_tversky (alpha=0.7, beta=0.3)

Auto-Conversion

When you switch the training mode between binary and multiclass, the loss function dropdown automatically selects a sensible equivalent:

• binary_dice_bce ↔ dice_ce

• binary_focal ↔ focal

• binary_tversky ↔ tversky

You can always override this auto-selection manually.