XAI Components & Workflow

The XAI Climate Identifier framework is structured around the following core components:

• Model Interface: Supports climate classification models and exposes prediction logits and gradients required for explanation methods.
• Attribution Methods: Implements a range of explanation techniques including Saliency Maps, Grad-CAM, Integrated Gradients, Occlusion, and SmoothGrad.
• Preprocessing Pipeline: Normalizes climate variables, maintains spatial resolution, and ensures reproducible input transformations for XAI consistency.
• Visualization Engine: Produces heatmaps, overlays, and comparative plots for analyzing explanation patterns across different methods.

Evaluation Metrics

The system incorporates a set of quantitative evaluation metrics to assess the reliability and usefulness of explanations:

• Faithfulness Metrics: Measures how closely explanations reflect model behavior through perturbation-based tests and sensitivity analysis.
• Robustness Metrics: Evaluates the stability of explanations under noise, transformation variance, and input perturbations.
• Complexity Metrics: Assesses the structural simplicity or density of explanation maps to ensure interpretability.
• Localization Metrics: Compares explanations against expected spatial regions in climate fields when reference maps are available.
• Sanity Checks: Verifies explanation validity through model and layer randomization tests.

Project Note

This project was originally planned to utilize a small-scale, region-specific climate dataset. Due to limited availability of suitable regional datasets with the required spatial and temporal consistency, this approach was set aside, and the project proceeded using broader and more accessible climate data sources instead.