PaperlayerView exact revision on arXiv

Source-linked AI summary

Environment-Adaptive Preference Optimization for Wildfire Prediction

Enyi Jiang, Wu Sun

arXiv:2605.12435v1cs.LGcs.CE

TL;DR

Wildfire prediction must detect rare, high-impact events as environmental distributions evolve, while few studies jointly address long-tail imbalance and distribution shift. EAPO adapts a supervised model using KNN-retrieved local data and preference optimization that emphasizes extreme events. It reports ROC-AUC 0.7310 and improved detection in moderately high fire-intensity regimes.

  • Problem

    Wildfire prediction must detect rare, high-impact fires under evolving environmental distributions, while few studies jointly address long-tail imbalance and distribution shifts.

    Source: §1 · p. 1, §2 · p. 2

  • Method

    EAPO uses KNN retrieval to form target-aligned local data, then combines supervised learning and preference optimization with emphasis on rare extreme events.

    Source: §3.1 · p. 3, §3.2 · p. 3

  • Results

    ROC-AUC 0.7310; EAPO improves robustness under distribution shift and detects more events in moderately high fire-intensity bins than baselines.

    Source: §1 · p. 2, §4 · p. 4

  • Takeaways & Limitations

    EAPO provides a promising direction for adaptive, preference-based models in dynamic wildfire prediction systems.

    Source: §5 · p. 4

    Experiments focus on Yosemite, limiting the geographic extent evaluated and leaving robustness across larger regions and more diverse fire regimes to be assessed.

    Source: §4 · p. 4

Abstract

from arXiv · show

Predicting rare extreme events such as wildfires from meteorological data requires models that remain reliable under evolving environmental conditions. This problem is inherently long-tailed: wildfire events are rare but high-impact, while most observations correspond to non-fire conditions, causing standard learning objectives to underemphasize the minority class (fire) that matters most. In addition, models trained on historical distributions often fail under distribution shifts, exhibiting degraded performance in new environments. To this end, we propose Environment-Adaptive Preference Optimization (EAPO), a framework that adapts prediction to the target environment with long-tail distribution. Given a new input distribution, we first construct distribution-aligned datasets via $k$-nearest neighbor retrieval. We then perform a hybrid fine-tuning procedure on this local manifold, combining supervised learning with preference optimization, as well as emphasizing on rare extreme events. EAPO refines decision boundaries while avoiding conflicting signals from heterogeneous training data. We evaluate EAPO on a real-world wildfire prediction task with environmental shifts. EAPO achieves robust performance (ROC-AUC 0.7310) and improves detection in extreme regimes, demonstrating its effectiveness in dynamic wildfire prediction systems.

1 Introduction

Wildfire prediction must detect rare, high-impact fires under shifting environmental regimes, yet existing approaches typically address long-tail imbalance and distribution shift separately. EAPO combines local preference-based ranking with supervised learning to adapt prediction to the target environment.

1 Introduction

  • Static data-driven models may generalize poorly to novel fire regimes as climate change shifts fire-weather conditions and fuel characteristics.Extreme wildfires also emerge from interacting weather, topography, vegetation fuels, and human factors that are difficult to represent fully in process-based models.Source: §1 · p. 1
  • Existing wildfire methods address long-tail data and distribution shifts separately, and few studies jointly address both challenges in a unified framework.The paper frames wildfire prediction as detecting rare, high-impact events under evolving environmental distributions.Source: §1 · p. 1
  • Relative risk ranking is motivated by probability calibration failing to transfer across environments while rare extreme events remain under-emphasized.The introduction argues that distinguishing high-risk from low-risk cases matters more than estimating precise probabilities.Source: §1 · p. 2
  • EAPO combines supervised learning with local preference-based ranking, using KNN-retrieved historical data to adapt decision boundaries to the target environment.The method emphasizes rare extreme events while restricting learning to a coherent local manifold.Source: §1 · p. 2

2 Background

Wildfire prediction combines rare positive events with changing input distributions, creating a challenge for models trained on historical data. The section motivates preference-based ranking alongside localized learning to address these issues.

2 Background

  • Standard empirical risk minimization is dominated by the majority non-fire class, yielding suboptimal performance on rare fire events.The target distribution has far fewer fire outcomes than non-fire outcomes.Source: §2 · p. 2
  • The combined long-tail and distribution-shift problem is to predict robustly in input-space regions that are underrepresented or shifting.Training and test distributions differ, so a model optimized on training data faces a mismatch at evaluation.Source: §2 · p. 2
  • Direct preference optimization learns from preferred and dispreferred outcomes relative to a reference policy, with β controlling preference sharpness.For binary classification, the model logit is converted to a probability with a sigmoid function.Source: §2 · p. 2
  • Naively pairing preferences across heterogeneous conditions can produce conflicting gradients and unstable decision boundaries, motivating localized preference optimization.The section describes DPO as an implicit hard-negative miner that highlights boundaries between rare fires and normal conditions.Source: §2 · p. 2

3 Environment-Adaptive Preference Optimization

EAPO adapts a supervised reference model using target-aligned local data and preference comparisons that emphasize rare extreme events. KNN retrieval defines the local manifold, and a hybrid objective refines its decision boundaries.

3 Environment-Adaptive Preference Optimization

  • EAPO starts from a reference policy obtained by standard supervised training over the training distribution.The reference policy is the sigmoid-transformed model output.Source: §3 · p. 2

3.1 Dynamic Manifold Retrieval via KNN

  • KNN retrieval approximates the test-time distribution with nearby historical inputs, then isolates the local fire examples as an extreme-event subset.The local and extreme datasets support subsequent construction of localized preference pairs.Source: §3.1 · p. 3

3.2 Localized Preference Optimization

  • EAPO combines local supervised loss with DPO preference pairs drawn from both the local dataset and its rare-fire subset.Each example forms a pair in which its observed label is preferred over the opposite label; this retains calibrated predictions while emphasizing target-aligned and rare events.Source: §3.2 · p. 3
  • The hybrid objective preserves global predictive structure while refining decision boundaries around rare extreme events and improving robustness under distribution shift.The preference-based refinement is performed on pairs sampled from the local dataset and its extreme subset.Source: §3.2 · p. 3

4 Experiment

Experiments test EAPO on Yosemite wildfire data with a temporal train/test shift and compare it with classical and neural baselines. EAPO leads on several rare-event metrics, while evaluation remains geographically limited.

  • The evaluation uses GridMET meteorological inputs and GFED5 wildfire labels for Yosemite, training on 2001–2020 data and testing on 2021.TabNet is fine-tuned with EAPO after BCE or Focal-loss training; comparisons include Logistic Regression, reweighted XGBoost, and TabNet baselines.Source: §4 · p. 3
  • 66.12% recall and 73.10% ROC AUC for Focal + EAPO were the best, while BCE + EAPO achieved highest F1 (32.26%) and precision (22.76%).The authors identify recall, F1, and ROC AUC as more informative than accuracy for this rare-positive-class task.Source: §4 · p. 3, Table 1 · p. 4
  • Above 80% detection in the 108–108.5 DM bin versus below 70% for baselines; all models degrade above 109 DM.The figure compares prediction breakdown across wildfire intensity levels using log-scaled DM.Source: Figure 1 · p. 4, §4 · p. 4
  • Neighborhood sizes k = 3, 5, and 10 demonstrate robust retrieval of local manifolds; the authors say restricting learning to a distribution-aligned manifold and emphasizing extreme events gives EAPO more robust decision boundaries in changing environments (X).Source: §4 · p. 4
  • The Yosemite-only focus limits geographic extent; larger regions and more diverse fire regimes are needed to evaluate robustness under spatial heterogeneity.The current formulation also does not explicitly distinguish natural from human-induced wildfires.Source: §4 · p. 4

5 Conclusion

EAPO combines dynamic KNN-based manifold retrieval with localized preference optimization to address extreme class imbalance and distribution shifts. The experiments report improved rare-event detection and strong ranking performance, and the authors present the framework as a promising direction for dynamic wildfire prediction.

  • EAPO combines dynamic KNN-based manifold retrieval with localized preference optimization to address extreme class imbalance and distribution shifts.Source: §5 · p. 4
  • The experiments demonstrate improved detection of rare events and strong ranking performance (ROC-AUC) under distribution shifts.Source: §5 · p. 4
  • The authors describe adaptive, preference-based models as a promising direction for dynamic wildfire prediction systems.Source: §5 · p. 4
Loading 2605.12435v1…