融合ASPP机制的可解释卷积神经网络在野火蔓延预测中的应用

Application of Interpretable Convolutional Neural Networks Incorporating ASPP Mechanism in Wildfire Spread Prediction

摘要: 森林生态系统因野外火灾而遭受的破坏在全球范围内引发了重大关注。随着近年来野火的严重性和发生频率的增加，对高效预测模型的需求变得尤为迫切。本研究介绍了一种集成了有向金字塔池化(Atrous Spatial Pyramid Pooling, ASPP)机制的卷积神经网络(Convolutional Neural Network, CNN)模型，旨在提高野火蔓延预测的准确性。鉴于现有人工神经网络算法的“黑箱”特性限制了其在解释性方面的应用，本研究提出了一种新型的可解释CNN模型，即融合ASPP机制的CNN-ASPP。该模型利用包含环境变量的“次日野火蔓延”数据集进行性能评估，与当前先进的机器学习方法，如随机森林(Random Forest, RF)、支持向量机(Support Vector Machine, SVM)、人工神经网络(Artificial Neural Network, ANN)以及另一种CNN模型进行了比较。实验结果表明，在7 × 7邻域大小条件下，CNN-ASPP模型的F1-score高达97%，显著优于其他机器学习方法的90% F1-score。此外，本研究还通过梯度加权类激活映射(Gradient-Weighted Class Activation Mapping, Grad-CAM)算法，对CNN-ASPP模型中的不同卷积层进行了解释，揭示了较大扩张率(Dilation Rates, DR)在提取输入数据中更有意义特征方面的优势。本研究的发现对于开发更透明、更精确的野火蔓延预测模型具有重要意义，对森林资源管理和野火预防策略的制定具有深远的影响。

Abstract: Forest ecosystems have been persistently affected by wildfires, leading to a concern worldwide. The severity and frequency of wildfires have escalated in recent years, necessitating more effective prediction models. This study presents an application of convolutional neural networks (CNNs) for wildfire spread prediction, focusing on the use of atrous spatial pyramid pooling (ASPP) mechanisms in these networks, aiming to improve the accuracy of wildfire spread prediction. However, the black-box nature of these algorithms has not been fully explored. To bridge this gap, we proposed an explainable CNN model with an ASPP mechanism (CNN-ASPP) in this study. More specifically, we utilize the Next Day Wildfire Spread dataset, which includes environmental variables, to evaluate the performance of our model. The proposed model is compared with state-of-the-art machine learning (ML) methods, including random forest (RF), support vector machine (SVM), artificial neural network (ANN), and another CNN model. Our results showed that CNN-ASPP achieved an F1-score of 97%, outperforming the ML methods with an F1-score of 90% for a neighborhood size of 7 × 7. We also opened the black box and tried to explain different convolutional layers based on the gradient-weighted class activation mapping (Grad-CAM) algorithm. Our findings indicate that larger dilation rates (DRs) can extract more meaningful features from the input data. This study contributes to the development of more transparent and accurate models for wildfire spread prediction, which could have significant implications for forest management and wildfire prevention strategies.