Table 12 Comparative analysis of Grad-CAM++ and LIME interpretability methods.

From: SpinachXAI-Rec: a multi-stage explainable AI framework for spinach freshness classification and consumer recommendation

Aspect

Grad-CAM++

LIME

Type

Visual Heatmap Overlay

Feature Attribution via Superpixel Perturbation

Interpretability scope

Global (entire image view)

Local (specific superpixels/features)

Visualization clarity

Smooth transitions, highlights full leaf structure

Sharp contours, sparse activation zones

Feature relevance

Captures spatially continuous attention

Highlights top contributing features only

IoU/Dice evaluation

Quantified using segmentation-like metrics

Not directly applicable (no region overlap scoring)

Use case suitability

Better for biomedical and plant structural features

Best for debugging model behavior and local causes

Back to article page