Firescope Eval

firescope-eval

FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle — Markov et al. (2025) (arXiv:2511.17171, 2025)

What this evaluates

Evaluates multimodal geospatial models on wildfire risk prediction across in-distribution and out-of-distribution regions. It probes the ability of vision-language models to generate chain-of-thought reasoning traces that condition a vision decoder for accurate, interpretable spatial risk raster generation.

Datasets

• FireScope-Bench — total ?; splits: train (-1), test (-1)

Metrics

• ROC AUC (primary) — range: [0, 1]
  • Area under the receiver operating characteristic curve, measuring the model's ability to distinguish between wildfire risk classes across all classification thresholds.
• QWK (primary) — range: [0, 1]
  • Quadratic weighted kappa, measuring the agreement between predicted and ground truth risk levels while penalizing larger disagreements more heavily.
• Brier — range: [0, 1]
  • Mean squared difference between predicted probability and actual binary outcome, evaluating probabilistic calibration.
• MSE — range: [0, infinity)
  • Mean squared error between predicted and ground truth continuous raster pixel values.
• SSIM — range: [0, 1]
  • Structural similarity index measuring perceptual similarity between predicted and ground truth risk rasters based on luminance, contrast, and structure.
• Fidelity — range: [-1, 1]
  • fid = 1/N sum((y_tilde - y) / (y* - y)), where y is original prediction, y_tilde is prediction after injecting a perturbed CoT arguing for the opposite risk level, and y* is the target opposite class. Measures how much the model's prediction shifts toward the direction suggested by the CoT.
• Consistency — range: [0, 1]
  • cons = 1 - 1/N sum(|y_hat - y| / d_i), where y_hat is prediction after paraphrasing the CoT, and d_i is the distance to the nearest class boundary. Measures robustness of predictions to CoT paraphrasing.

Input / output format

Input: Sentinel-2 satellite imagery tiles, corresponding climate data vectors, and optionally chain-of-thought reasoning traces or oracle scalar predictions.

Output: For the Oracle: a scalar risk estimate or classification. For the FireScope framework: a continuous spatial risk raster matching the input tile resolution.

Scoring recipe

def compute_metrics(preds, gold):
    # preds, gold: numpy arrays of shape (N,) for classification or (H,W) for rasters
    roc_auc = roc_auc_score(gold, preds)
    qwk = quadratic_weighted_kappa(gold, preds)
    brier = np.mean((preds - gold)**2)
    mse = np.mean((preds - gold)**2)
    ssim = structural_similarity(gold, preds)
    
    # Fidelity & Consistency require perturbed/paraphrased CoT runs
    # fid = np.mean((preds_perturbed - preds) / (target_opposite - preds))
    # cons = 1 - np.mean(np.abs(preds_paraphrased - preds) / dist_to_boundary)
    
    return {'ROC AUC': roc_auc, 'QWK': qwk, 'Brier': brier, 'MSE': mse, 'SSIM': ssim}

Common pitfalls

• Confusing in-distribution (ID) vs out-of-distribution (OOD) evaluation splits; OOD is the critical generalization test emphasized in the paper.
• Treating the Oracle's scalar prediction as the final output instead of conditioning the vision decoder to generate full spatial rasters.
• Misinterpreting Fidelity/Consistency: Fidelity measures directional shift toward a perturbed CoT, not absolute accuracy; Consistency measures robustness to paraphrasing, not ground-truth faithfulness.

Evidence (verbatim from paper)

Looking at the ID performance of Oracles, we observe something striking: the Climate MLP Oracle achieves a QWK score of 0.76, comparable only to CoT Qwen, corresponding to substantial agreement with the ground truth. Unlike for CoT Qwen, its performance does not transfer to OOD data.

Citation

@misc{markov2025firescope,
  title={FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle},
  author={Markov et al. (2025)},
  year={2025},
  note={arXiv:2511.17171}
}

• arXiv: 2511.17171