Enhanced Wildfire Detection Using a Mixture of Experts Approach

Beihua Kong, Yu ZhangJournal of Student Research 2025 · DOI

Abstract

With increasing global wildfire severity, effective fire detection methods are essential to mitigate widespread environmental and health impacts. A recent solution to this phenomena is the application of ensemble machine learning methods, which combine several models to create a more effective one. However, this raises several questions, notably whether an ensemble method is more effective than an individual model or if increasing the number of constituent models leads to overfitting. This paper conducts an ablation study on the Mixture of Experts (MoE) approach for forest fire detection via satellite imagery across a Canadian dataset. The model (MoE6) constitutes all six state-of-the-art architectures, including InceptionNet, ResNet, Vision Transformer (ViT), AlexNet, VGG-Net, and a baseline CNN. Experts of the MoE6 will be systematically removed to form MoE4 and MoE2, which constitute only the top four and top two performing constituent models respectively. We hypothesize that the MoE ensemble approach will outperform any constituent model (two heads are better than one). Furthermore, among the MoE architectures, we hypothesize MoE2 as the top model as it comprehensively integrates characteristics from top model architectures while mitigating overfitting. However, the results show that the original MoE6 was the top performer, achieving a peak accuracy of 93.13% and ROC-AUC of 0.9303. This work provides a promising solution for improving wildfire detection accuracy and response times, potentially reducing the devastation caused by wildfires globally.

Introduction

As global wildfire severity escalates—exemplified by Canada’s record-breaking 2023 season—early and accurate detection is critical for mitigating environmental and public health disasters. While individual Convolutional Neural Networks (CNNs) have shown promise in classifying satellite imagery, they often face a trade-off between computational efficiency and accuracy.

The fundamental challenge in ensemble learning is determining whether adding more "expert" models to a system leads to better performance or simply introduces noise and overfitting. This research investigates whether a Mixture of Experts (MoE) approach can synthesize the strengths of various architectures—such as the spatial feature extraction of CNNs and the global context of Vision Transformers—to create a more reliable wildfire detection tool.

Methodology

The core of this research is an ablation study on a Mixture of Experts (MoE) framework. Unlike a simple average of model outputs, an MoE uses a gating network that applies a softmax function to assign weights to each "expert" model based on the specific input image.

Schematic of the Mixture of Experts (MoE) architecture showing the routing mechanism and weighted summation of expert outputs.

The study tested three MoE variants:

  • MoE6: Includes all six constituent models: InceptionNet, ResNet, Vision Transformer (ViT), AlexNet, VGG-Net, and a baseline "Vanilla" CNN.
  • MoE4: Includes only the top four performers (ResNet, CNN, VGG-Net, and Inception).
  • MoE2: Includes only the top two performers (ResNet and CNN).

The study compared these ensembles against the individual performance of each constituent architecture to determine if the "two heads (or six) are better than one" hypothesis held true.

Data and Preprocessing

The study utilized the Quebec Image Dataset, a collection of 42,850 satellite images of Southern Quebec spanning from 1972 to 2021. The dataset is relatively balanced, containing 22,710 wildfire instances and 20,140 non-fire instances.

To reduce computational overhead, the original 350x350 pixel images were downsampled to 32x32 pixels. Data was normalized to a [0,1] scale and split into training, validation, and testing sets using a 7:1:2 ratio.

Results and Discussion

The authors initially hypothesized that MoE2 would be the top performer, reasoning that it would integrate the best features while avoiding the overfitting risks of a larger, more complex model. However, MoE6 actually performed the best, achieving a peak accuracy of 93.13% and an ROC-AUC of 0.9303. This suggests that the gating network effectively managed the "weaker" experts, extracting value even from models that performed poorly in isolation.

Among individual models, ResNet was the standout performer (93.13% accuracy). Interestingly, while MoE6 matched this accuracy, it achieved a slightly higher ROC-AUC (0.9303 vs. 0.9286), indicating better overall discriminative capability across different classification thresholds.

Conversely, the Vision Transformer (ViT) was the weakest individual performer (82.31% accuracy). While it showed near-perfect recall, its poor precision suggests it struggles with false positives in this specific satellite context. This is likely due to the low resolution (32x32) of the input data, which may hinder the transformer's ability to learn complex global dependencies.

Conclusion and Future Work

The MoE approach provides a more balanced performance across precision and recall metrics compared to individual architectures. However, the primary limitation identified was overfitting; training curves for MoE6 showed training accuracy approaching 1.0 while validation accuracy stagnated around 0.94.

Future iterations could benefit from:

  • Regularization: Implementing data augmentation, dropout layers, or stricter weight decay to bridge the gap between training and validation performance.
  • Architectural Diversity: Incorporating more diverse transformer-based models, such as the Swin Transformer, to provide the gating network with a broader range of "opinions."
  • Resolution: Testing the models on higher-resolution imagery to see if the Vision Transformer's performance improves when more spatial detail is available.

By refining these ensemble methods, researchers can develop more robust tools for early wildfire detection, potentially saving lives and ecosystems.

Summary updated

Abstract

With increasing global wildfire severity, effective fire detection methods are essential to mitigate widespread environmental and health impacts. A recent solution to this phenomena is the application of ensemble machine learning methods, which combine several models to create a more effective one. However, this raises several questions, notably whether an ensemble method is more effective than an individual model or if increasing the number of constituent models leads to overfitting. This paper conducts an ablation study on the Mixture of Experts (MoE) approach for forest fire detection via satellite imagery across a Canadian dataset. The model (MoE6) constitutes all six state-of-the-art architectures, including InceptionNet, ResNet, Vision Transformer (ViT), AlexNet, VGG-Net, and a baseline CNN. Experts of the MoE6 will be systematically removed to form MoE4 and MoE2, which constitute only the top four and top two performing constituent models respectively. We hypothesize that the MoE ensemble approach will outperform any constituent model (two heads are better than one). Furthermore, among the MoE architectures, we hypothesize MoE2 as the top model as it comprehensively integrates characteristics from top model architectures while mitigating overfitting. However, the results show that the original MoE6 was the top performer, achieving a peak accuracy of 93.13\% and ROC-AUC of 0.9303. This work provides a promising solution for improving wildfire detection accuracy and response times, potentially reducing the devastation caused by wildfires globally.
BibTeX
@article{kong2025enhanced,
  author = {Beihua Kong and Yu Zhang},
  title = {{Enhanced Wildfire Detection Using a Mixture of Experts Approach}},
  journal = {Journal of Student Research},
  year = {2025},
  doi = {10.47611/jsrhs.v14i1.8612},
}
Figure