Environment-Adaptive Preference Optimization for Wildfire Prediction
arXiv:2605.12435v1 [cs.LG] 12 May 2026
1
Enyi Jiang1,2 and Wu Sun3 Siebel School of Computing and Data Science, University of Illiois at Urbana-Champaign 2 Computer Science, Stanford University 3 Department of Global Ecology, Carnegie Institution for Science [email protected]/[email protected], [email protected]
Abstract 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 EnvironmentAdaptive 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
Climate change is fueling more extreme wildfires [2], with devastating impacts on climate [6], health [28, 30], and the economy [39]. Extreme wildfires are challenging to predict [41], as they emerge from the complex interplay of fire weather [9, 10, 37], topography [33], vegetation fuels [32, 37], and human factors such as ignition and fire suppression [16, 19, 26, 44], all of which are difficult to fully represent in process-based wildfire models. Whereas machine learning (ML) approaches such as XGBoost [8] have shown promise in wildfire prediction [5, 18, 21, 25, 42], outperforming process-based wildfire models [41], they typically require extensive historical fire data for training and may struggle to generalize to novel fire regimes that emerge under climate change [14]. As climate change causes shifts in the spatial pattern, seasonality, and statistical distribution of fire weather conditions [15] and fuel quantity and quality [4, 11, 12], static data-driven ML approaches are not well suited to predict fire behavior under novel regimes that fall outside historical distributions. Wildfire prediction is a long-tail, non-stationary problem, where rare, high-impact events need to be detected under evolving environmental distributions. From a machine learning perspective, it can be framed as a long-tail classification problem [45] under distribution shifts [35]. Existing methods address these challenges separately, using different loss adjustments for long-tail dataset [24, 43, 22] and domain generalization [27, 20, 46] or test-time adaptation [36, 40, 7] for the distribution shift. However, to the best of our knowledge, few studies address these challenges jointly within a unified framework. Further, these approaches remain grounded in empirical risk minimization (ERM), which Preprint.
prioritizes probability calibration. Such calibration often fails to transfer across environments, while rare extreme events remain under-emphasized. As a result, distinguishing high-risk from low-risk cases becomes more important than estimating precise probabilities. This motivates a shift to relative risk ranking, which aims to learn relative preferences among extreme, important and normal outcomes at the same time. Notably, preference optimization [31, 34] provides a natural framework for this objective, enabling the model to learn more fine-grained decision boundaries. To this end, we propose Environment-Adaptive Preference Optimization, which combines probability calibration with local preference-based ranking to improve extreme event prediction under distribution shifts. Given a test input distribution and a pretrained model on historical data, we first construct a distribution-aligned local extreme event dataset via k-nearest neighbor (KNN) retrieval [29] from historical data. We then perform joint supervised and preference-based optimization on this local subset, combining supervised loss with a localized DPO loss. By restricting learning to a coherent local manifold, our method enforces consistent preference signals and adapts decision boundaries to environment-specific conditions. We evaluate our framework on a wildfire dataset under distribution shifts. EAPO improves robustness under distribution shift, achieving superior performance (ROCAUC 0.7310) and outperforming various baselines. These results demonstrate the effectiveness of combining preference learning with environment-adaptive modeling for wildfire risk prediction.
2
Background
Wildfire prediction as a long-tail problem under distribution shift. Let (x, y) ∼ D denote meteorological input x ∈ X and the corresponding binary wildfire outcome y ∈ Y (0 for non-fire and 1 for fire). In extreme wildfire prediction, the target variable shows a long-tail distribution: P(y ∈ Y1 ) ≪ P(y ∈ Y0 ), where Y1 denotes rare, high-impact fire events (Y = Y0 ∪ Y1 ). Given a model fθ and a loss function ℓ, standard empirical risk minimization (ERM) minimizes LERM = E(x,y)∼Dtrain [ℓ(fθ (x), y)] , (1) which is dominated by the majority class Y0 , leading to suboptimal performance on Y1 . In addition, wildfire prediction is subject to a distribution shift in x, driven by internal climate variability and long-term climate change, such that Dtrain (x) ̸= Dtest (x). As a result, a model trained on Dtrain minimizes E(x,y)∼Dtrain [ℓ(fθ (x), y)] but is evaluated under E(x,y)∼Dtest [ℓ(fθ (x), y)], leading to a mismatch that degrades generalizability. The combination of long-tail distributions and distribution shifts poses a fundamental challenge: how to robustly predict in regions of the input space that are underrepresented or experiencing a distribution shift. Direct preference optimization (DPO). We denote fθ (x) as the model logit and convert it to the probability used in preference comparisons using a sigmoid function (policy) πθ (y | x) = σ(fθ (x)) for binary classification. DPO optimizes a parameterized policy πθ using preference data (x, y + , y − ), where y + is preferred over y − . Following [31], the objective is defined relative to a reference policy πref , where β controls the sharpness of the preference: πθ (y + | x) πθ (y − | x) LDPO (πθ ) = −E(x,y+ ,y− )∼D log σ β log − log . (2) πref (y + | x) πref (y − | x) This relative ranking mechanism is key to addressing distribution problems in wildfire prediction. Under extreme class imbalance between fire and non-fire outcomes, standard cross-entropy models may minimize global loss by predicting non-fire for all instances (“nothing ever happens”). In this case, DPO is useful because it acts as an implicit hard-negative miner, forcing the network to explicitly learn the subtle physical boundaries that differentiate a rare yet important event from normal conditions (non-fire). However, naïvely constructing preference pairs across heterogeneous conditions can introduce conflicting gradients, leading to overfitting and unstable decision boundaries. This motivates a localized approach to preference optimization, which leads to our EAPO.
3
Environment-Adaptive Preference Optimization
We introduce EAPO, given a reference policy πref obtained from standard supervised training. Specifically, πref (y | x) = σ(fθ (x)) is derived from a model fθ trained by minimizing the empirical risk LERM (Eq. 1) over the training distribution. 2
3.1
Dynamic Manifold Retrieval via KNN
Under distribution shift, the test distribution Dtest (x) may occupy regions poorly represented by Dtrain (x). To mitigate this, we construct a localized approximation of the test-time distribution via B non-parametric k-nearest neighbor (KNN) retrieval [29]. Given target inputs {xtest i }i=1 ∼ Dtest , we define the neighborhood operator: Nk (xtest ) = {xtrain : xtrain ∈ Top-k nearest neighbors of xtest }. (3) j j S train train The resulting distribution-aligned dataset is Dlocal = xtest ∼Dtest {(xtrain ∈ Nk (xtest )}, j , yj ) : xj which reduces distribution mismatch by restricting learning to an analogous meteorological regime. Within this local manifold, we further isolate rare extreme events Dextreme = {(x, y) ∈ Dlocal : y = 1}. Together, Dlocal and Dextreme provide a structured decomposition of the target-aligned data, separating the dominant background conditions from rare extreme events. The KNN retrieval forms the basis for constructing localized preference pairs in subsequent optimization. 3.2
Localized Preference Optimization
Given the localized dataset Dlocal and Dextreme , we pose EAPO as a fine-tuning procedure that combines supervised learning with preference-based optimization. First, we retain a supervised binary classification objective LSFT = E(x,y)∼Dlocal [ℓ(fθ (x), y)] over the local distribution, which preserves calibrated predictions within the target-aligned region. To emphasize target-aligned and rare events, we construct preference pairs from Dlocal and its extreme subset Dextreme . For each (x, y) ∈ Dlocal or Dextreme , we define: y + = y and y − = 1 − y, yielding a preference pair (x, y + , y − ). We then apply the DPO objective (Eq. 2) on preference pairs sampled from Dlocal and Dextreme . The overall objective combines supervised learning with preference-based refinement: LEAPO = LSFT + λ1 · LDPO-local + λ2 · LDPO-extreme .
(4)
This hybrid formulation preserves global predictive structure while selectively refining decision boundaries around rare extreme events, while improving robustness under distribution shift.
4
Experiment
Experimental setup. We use the GridMET dataset [1] for the meteorological inputs x, which consist of 12 features (e.g., temperature, humidity, precipitation, wind, pressure, burning index, and fuel moisture), and the GFED5 dataset [38] for wildfire labels y, with positive dry matter (DM) values indicating wildfire events (y = 1). We focus on the Yosemite region, historically prone to natural wildfire [23], to capture meaningful extreme events and reduce geographical variation in fire occurrence. We use 2001–2020 data for training and 2021 data for test to simulate distribution shift. We compare EAPO against various baselines, including Logistic Regression [13] and XGBoost [8] with class reweighting, as well as neural network models (TabNet [3]) trained with Binary CrossEntropy (BCE) and Focal loss [24]. Implementation and evaluation. We first train a TabNet model on the training data for 100 epochs for BCE and 50 epochs for Focal loss, with a learning rate of 0.005. We then fine-tune the model using EAPO for an additional 100 epochs, with a learning rate of 0.0001. We use an Adam [17] optimizer for both stages. For hyperparameters, we evaluate k ∈ {3, 5, 10} for nearest neighbor retrieval, and set β = 0.1, λ1 = 1.0, and λ2 = 0.1. For evaluation, we select the optimal decision threshold for each method based on the precision-recall (PR) curve on the training set, and report performance on the test set using the selected threshold. EAPO achieves the best overall performance. Table 1 shows that EAPO consistently outperforms strong baselines, with the largest gains in recall, F1, and ROC AUC. In particular, Focal + EAPO achieves the best recall (66.12%) and ROC AUC (73.10%), while BCE + EAPO attains the highest F1 (32.26%) and precision (22.76%). In extreme wildfire prediction, where the positive class is rare, recall, F1, and ROC AUC are more informative than accuracy. This is because recall measures the ability to detect rare events, F1 captures the trade-off between detection and false alarms, and ROC AUC reflects ranking quality across thresholds. From this perspective, EAPO achieves the most well-balanced performance. Furthermore, the consistent performance across varying neighborhood 3
Model
Acc (%)
Prec (%)
Recall (%)
F1 (%)
Logistic Regression 66.42 19.52 65.29 30.06 XGBoost 74.82 22.51 52.34 31.48 BCE 74.16 20.58 46.83 28.60 BCE + EAPO (k = 5) 74.31 22.76 55.37 32.26 Focal 72.42 20.84 53.44 29.98 Focal + EAPO (k = 3) 69.19 21.26 66.12 32.17 Focal + EAPO (k = 5) 69.19 21.16 65.56 31.99 Focal + SFT (k = 10) 71.45 20.87 56.75 30.52 Focal + EAPO (k = 10) 69.07 21.18 66.12 32.09 Table 1: Performance comparison on wildfire prediction.
ROC AUC (%) 70.40 71.09 68.39 71.97 70.67 73.10 73.08 70.19 73.07
Figure 1: Prediction breakdown across wildfire intensity levels (log-scaled DM). sizes (k = 3, 5, 10) demonstrates robust retrieval of local manifolds. These gains reveals that by restricting learning to a distribution-aligned manifold and emphasizing extreme events, EAPO has more robust decision boundaries for changing environments (X ). EAPO improves detection of extreme wildfire events. Figure 1 shows model performance across wildfire intensity levels (log-scaled DM), where higher DM consumed by fire corresponds to more extreme events. Despite the degradation of performance in extremely high DM regimes (> 109 ) across all models, EAPO achieves higher detection rates in moderately high DM bins (e.g., > 80 % for the bin 108 –108.5 , compared with < 70 % in the same bin for baselines). This means that EAPO better captures high-impact cases with robust performance in moderately high fire intensity regimes. Discussion. While our results demonstrate the effectiveness of EAPO, several limitations remain to be addressed in future work. First, our experiments focus on the Yosemite region, which simplifies spatial variability but limits the geographic extent. Extending the framework to larger regions and more diverse fire regimes would help better evaluate its robustness under spatial heterogeneity. Second, our current formulation does not explicitly distinguish between natural and human-induced wildfires. These two types of events are driven by different underlying mechanisms and may exhibit distinct patterns in the feature space. Incorporating additional signals or modeling strategies to differentiate ignition sources could further improve predictability and interpretability.
5
Conclusion
In this work, we introduce EAPO to address extreme class imbalance and distribution shifts in wildfire prediction. By combining dynamic KNN-based manifold retrieval with localized preference optimization, our approach shifts learning from probability calibration to relative risk ranking. Our experiments demonstrate that EAPO improves detection of rare events and achieves strong ranking performance (ROC-AUC) under distribution shifts. This framework provides a promising direction for adaptive, preference-based models in dynamic wildfire prediction systems. 4
References [1] John T Abatzoglou. Development of gridded surface meteorological data for ecological applications and modelling. International journal of climatology, 33(1):121–131, 2013. [2] John T. Abatzoglou, Crystal A. Kolden, Alison C. Cullen, Mojtaba Sadegh, Emily L. Williams, Marco Turco, and Matthew W. Jones. Climate change has increased the odds of extreme regional forest fire years globally. Nature Communications, 16(1), 7 2025. [3] Sercan Ö Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 6679–6687, 2021. [4] Jennifer L. Baltzer, Nicola J. Day, Xanthe J. Walker, David Greene, Michelle C. Mack, Heather D. Alexander, Dominique Arseneault, Jennifer Barnes, Yves Bergeron, Yan Boucher, Laura Bourgeau-Chavez, Carissa D. Brown, Suzanne Carrière, Brian K. Howard, Sylvie Gauthier, Marc-André Parisien, Kirsten A. Reid, Brendan M. Rogers, Carl Roland, Luc Sirois, Sarah Stehn, Dan K. Thompson, Merritt R. Turetsky, Sander Veraverbeke, Ellen Whitman, Jian Yang, and Jill F. Johnstone. Increasing fire and the decline of fire adapted black spruce in the boreal forest. Proceedings of the National Academy of Sciences, 118(45):e2024872118, 2021. [5] J. Buch, A. P. Williams, C. S. Juang, W. D. Hansen, and P. Gentine. SMLFire1.0: a stochastic machine learning (sml) model for wildfire activity in the western United States. Geoscientific Model Development, 16(12):3407–3433, 2023. [6] Brendan Byrne, Junjie Liu, Kevin W. Bowman, Madeleine Pascolini-Campbell, Abhishek Chatterjee, Sudhanshu Pandey, Kazuyuki Miyazaki, Guido R. van der Werf, Debra Wunch, Paul O. Wennberg, Coleen M. Roehl, and Saptarshi Sinha. Carbon emissions from the 2023 Canadian wildfires. Nature, 633(8031):835–839, 8 2024. [7] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 295–305, 2022. [8] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016. [9] Janice L. Coen, E. Natasha Stavros, and Josephine A. Fites-Kaufman. Deconstructing the King megafire. Ecological Applications, 28(6):1565–1580, 2018. [10] Andrea Duane, Marc Castellnou, and Lluís Brotons. Towards a comprehensive look at global drivers of novel extreme wildfire events. Climatic Change, 165(3–4), 4 2021. [11] Todd M. Ellis, David M. J. S. Bowman, Piyush Jain, Mike D. Flannigan, and Grant J. Williamson. Global increase in wildfire risk due to climate-driven declines in fuel moisture. Global Change Biology, 28(4):1544–1559, 2022. [12] Jessica E. Halofsky, David L. Peterson, and Brian J. Harvey. Changing wildfire, changing forests: the effects of climate change on fire regimes and vegetation in the Pacific Northwest, USA. Fire Ecology, 16(1), 1 2020. [13] David W Hosmer Jr, Stanley Lemeshow, and Rodney X Sturdivant. Applied logistic regression. John Wiley & Sons, 2013. [14] Piyush Jain, Sean C.P. Coogan, Sriram Ganapathi Subramanian, Mark Crowley, Steve Taylor, and Mike D. Flannigan. A review of machine learning applications in wildfire science and management. Environmental Reviews, 28(4):478–505, 2020. [15] Matthew W. Jones, John T. Abatzoglou, Sander Veraverbeke, Niels Andela, Gitta Lasslop, Matthias Forkel, Adam J. P. Smith, Chantelle Burton, Richard A. Betts, Guido R. van der Werf, Stephen Sitch, Josep G. Canadell, Cristina Santín, Crystal Kolden, Stefan H. Doerr, and Corinne Le Quéré. Global and regional trends and drivers of fire under climate change. Reviews of Geophysics, 60(3):e2020RG000726, 2022. e2020RG000726 2020RG000726. 5
[16] Jon E. Keeley and Alexandra D. Syphard. Causal analysis of fire regime drivers in California. International Journal of Wildland Fire, 34(12):WF25166, 12 2025. [17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. [18] Spyros Kondylatos, Ioannis Prapas, Michele Ronco, Ioannis Papoutsis, Gustau Camps-Valls, María Piles, Miguel-Ángel Fernández-Torres, and Nuno Carvalhais. Wildfire danger prediction and understanding with deep learning. Geophysical Research Letters, 49(17):e2022GL099368, 2022. e2022GL099368 2022GL099368. [19] Mark R. Kreider, Philip E. Higuera, Sean A. Parks, William L. Rice, Nadia White, and Andrew J. Larson. Fire suppression makes wildfires more severe and accentuates impacts of climate change and fuel accumulation. Nature Communications, 15(1), 3 2024. [20] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Metalearning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. [21] Fa Li, Qing Zhu, Kunxiaojia Yuan, Fujiang Ji, Arindam Paul, Peng Lee, Volker C. Radeloff, and Min Chen. Projecting large fires in the western US with an interpretable and accurate hybrid machine learning method. Earth’s Future, 12(10):e2024EF004588, 2024. e2024EF004588 2024EF004588. [22] Mengke Li, Yiu-ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6929–6938, June 2022. [23] Shu Li and Tirtha Banerjee. Spatial and temporal pattern of wildfires in california from 2000 to 2019. Scientific reports, 11(1):8779, 2021. [24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. [25] Y. Liu, H. Huang, S.-C. Wang, T. Zhang, D. Xu, and Y. Chen. ELM2.1-XGBfire1.0: improving wildfire prediction by integrating a machine learning fire model in a land surface model. Geoscientific Model Development, 18(13):4103–4117, 2025. [26] Rebecca K. Miller, Christopher B. Field, and Katharine J. Mach. Barriers and enablers for prescribed burns for wildfire management in California. Nature Sustainability, 3(2):101–109, 1 2020. [27] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International conference on machine learning, pages 10–18. PMLR, 2013. [28] Chae Yeon Park, Kiyoshi Takahashi, Shinichiro Fujimori, Thanapat Jansakoo, Chantelle Burton, Huilin Huang, Sian Kou-Giesbrecht, Christopher P. O. Reyer, Matthias Mengel, Eleanor Burke, Fang Li, Stijn Hantson, Junya Takakura, Dong Kun Lee, and Tomoko Hasegawa. Attributing human mortality from fire PM2.5 to climate change. Nature Climate Change, 14(11):1193–1200, 10 2024. [29] Leif E Peterson. K-nearest neighbor. Scholarpedia, 4(2):1883, 2009. [30] Minghao Qiu, Jessica Li, Carlos F. Gould, Renzhi Jing, Makoto Kelp, Marissa L. Childs, Jeff Wen, Yuanyu Xie, Meiyun Lin, Mathew V. Kiang, Sam Heft-Neal, Noah S. Diffenbaugh, and Marshall Burke. Wildfire smoke exposure and mortality burden in the USA under climate change. Nature, 647(8091):935–943, 9 2025. [31] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741, 2023. 6
[32] Krishna Rao, A. Park Williams, Noah S. Diffenbaugh, Marta Yebra, Colleen Bryant, and Alexandra G. Konings. Dry live fuels increase the likelihood of lightning-caused fires. Geophysical Research Letters, 50(15):e2022GL100975, 2023. e2022GL100975 2022GL100975. [33] Jason J. Sharples, Richard H. D. McRae, and Stephen R. Wilkes. Wind–terrain effects on the propagation of wildfires in rugged terrain: fire channelling. International Journal of Wildland Fire, 21(3):282–296, 01 2012. [34] Feifan Song, Bowen Yu, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18990–18998, 2024. [35] Masashi Sugiyama and Motoaki Kawanabe. Machine learning in non-stationary environments: Introduction to covariate shift adaptation. MIT press, 2012. [36] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–9248. PMLR, 2020. [37] Daniel L. Swain. A shorter, sharper rainy season amplifies California wildfire risk. Geophysical Research Letters, 48(5):e2021GL092843, 2021. e2021GL092843 2021GL092843. [38] Guido R van der Werf, James T Randerson, Dave van Wees, Yang Chen, Louis Giglio, Joanne Hall, Roland Vernooij, Mingquan Mu, Samiha Binte Shahid, Kelley C Barsanti, et al. Landscape fire emissions from the 5th version of the global fire emissions database (gfed5). Scientific Data, 12(1):1870, 2025. [39] Daoping Wang, Dabo Guan, Shupeng Zhu, Michael Mac Kinnon, Guannan Geng, Qiang Zhang, Heran Zheng, Tianyang Lei, Shuai Shao, Peng Gong, and Steven J. Davis. Economic footprint of California wildfires in 2018. Nature Sustainability, 4(3):252–260, 12 2020. [40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. [41] S. S.-C. Wang, Y. Qian, L. R. Leung, and Y. Zhang. Interpreting machine learning prediction of fire emissions and comparison with firemip process-based models. Atmospheric Chemistry and Physics, 22(5):3445–3468, 2022. [42] Sally S.-C. Wang, Yun Qian, L. Ruby Leung, and Yang Zhang. Identifying key drivers of wildfires in the contiguous US using machine learning and game theory interpretation. Earth’s Future, 9(6):e2020EF001910, 2021. e2020EF001910 2020EF001910. [43] Tong Wu, Qingqiu Huang, Ziwei Liu, Yu Wang, and Dahua Lin. Distribution-balanced loss for multi-label classification in long-tailed datasets. In European conference on computer vision, pages 162–178. Springer, 2020. [44] Xiao Wu, Erik Sverdrup, Michael D. Mastrandrea, Michael W. Wara, and Stefan Wager. Lowintensity fires mitigate the risk of high-intensity wildfires in California’s forests. Science Advances, 9(45):eadi4123, 2023. [45] Chongsheng Zhang, George Almpanidis, Gaojuan Fan, Binquan Deng, Yanbo Zhang, Ji Liu, Aouaidjia Kamel, Paolo Soda, and João Gama. A systematic review on long-tailed learning. IEEE Transactions on Neural Networks and Learning Systems, 2025. [46] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. arXiv preprint arXiv:2104.02008, 2021.
7