Environment-Adaptive Preference Optimization for Wildfire Prediction

Problem
This paper addresses the challenge of predicting rare extreme events, specifically wildfires, from meteorological data. The authors highlight a significant gap in existing literature regarding the reliability of predictive models under evolving environmental conditions. The problem is exacerbated by the long-tailed nature of wildfire events, where the minority class (wildfires) is underrepresented in training datasets, leading to suboptimal model performance. Furthermore, models trained on historical data often fail to generalize when faced with distribution shifts, resulting in degraded predictive accuracy. This work is presented as a preprint and has not yet undergone peer review.

Method
The authors propose a novel framework called Environment-Adaptive Preference Optimization (EAPO). The core technical contribution involves a two-step process: first, constructing distribution-aligned datasets using $k$-nearest neighbor retrieval to adapt to the target environment; second, employing a hybrid fine-tuning procedure that integrates supervised learning with preference optimization. This approach emphasizes the importance of rare extreme events by refining decision boundaries while mitigating conflicting signals from heterogeneous training data. The training compute details are not disclosed, but the methodology is designed to enhance model robustness in dynamic environments.

Results
EAPO was evaluated on a real-world wildfire prediction task, demonstrating a robust performance with a ROC-AUC score of 0.7310. This performance was benchmarked against standard models that do not incorporate the proposed adaptive techniques. The results indicate a significant improvement in the detection of wildfires, particularly in extreme regimes, showcasing the effectiveness of EAPO in addressing the challenges posed by environmental shifts and long-tailed distributions.

Limitations
The authors acknowledge that while EAPO improves performance in dynamic environments, it may still be sensitive to the quality of the retrieved datasets and the choice of $k$ in the $k$-nearest neighbor retrieval process. Additionally, the framework’s reliance on historical data for initial training could introduce biases if the historical distribution does not adequately represent future conditions. The paper does not address potential computational overhead introduced by the hybrid fine-tuning process, which may limit scalability in real-time applications.

Why it matters
The implications of this work are significant for downstream applications in environmental monitoring and disaster response. By improving the predictive capabilities for rare events like wildfires, EAPO can enhance decision-making processes in resource allocation and risk management. This research contributes to the broader field of machine learning for environmental science, particularly in developing adaptive models that can maintain performance despite changing conditions. The framework could serve as a foundation for future studies aimed at optimizing predictions for other rare events across various domains.

Authors: Enyi Jiang, Wu Sun
Source: arXiv:2605.12435
URL: https://arxiv.org/abs/2605.12435v1