Devastating wildfires have increased in frequency and intensity over the last few years, worsened by climate change and prolonged droughts. Wildfire susceptibility mapping with machine learning has been proven useful for fire-prevention plans, turning into an indispensable tool in wildfire prevention. However, applications of deep learning models in wildfire susceptibility prediction to date are scarce. This study proposes a new Ensemble model based on two deep learning networks previously presented in literature that achieved remarkable results for forest fire susceptibility and other environmental risks. We compare our model with each of its sub-models, two more deep learning networks, and other machine learning benchmark, namely, XGBoost and SVM. Furthermore, we analyze the effects that different sample patch sizes have on the predictive performance of the algorithms. As case study we selected the fire occurrences in two regions in Chile, from 2013 to 2019. Satellite imagery data for fifteen fire influencing factors in the study area were retrieved to build a dataset to extract the samples to train the models. These factors include elevation, aspect, surface roughness, slope, minimum and maximum temperature, wind speed, precipitation, actual evapotranspiration, climatic water deficit, NDVI, land cover type, distance to rivers, distance to roads and distance to urban areas. During training, the best sample patch size was found to be 25 × 25 pixels. As a result, the highest area under the curve (AUC) was 0.953 achieved by the Ensemble model, followed by CNN-1 with AUC = 0.902. The Ensemble model also achieved the best accuracy, sensitivity, specificity, negative predictive value and F1 score. Finally, the predicted susceptibility maps suggest that static variables can be considered as predisposing factors, while dynamic variables affect the intensity of the predicted probabilities, with an important role of the anthropogenic variables. These resulting maps may be useful to prioritize wildfire surveillance and monitoring in extensive high risk areas.

在过去几年中，毁灭性的野火频率和强度都在增加，气候变化和长期干旱使情况更加恶化。使用机器学习的野火敏感性映射已被证明对防火计划很有用，已成为预防野火不可或缺的工具。然而，迄今为止，深度学习模型在野火敏感性预测中的应用很少。本研究基于之前文献中提出的两个深度学习网络提出了一种新的 Ensemble 模型，该模型在森林火灾敏感性和其他环境风险方面取得了显着的成果。我们将我们的模型与其每个子模型、另外两个深度学习网络和其他机器学习基准，即 XGBoost 和 SVM 进行比较。此外，我们分析了不同样本块大小对算法预测性能的影响。作为案例研究，我们选择了 2013 年至 2019 年智利两个地区的火灾事件。检索研究区域内 15 个火灾影响因素的卫星图像数据，构建数据集以提取样本以训练模型。这些因素包括海拔、坡向、地表粗糙度、坡度、最低和最高温度、风速、降水、实际蒸散量、气候水分亏缺、NDVI、土地覆盖类型、与河流的距离、与道路的距离和与市区的距离。在训练期间，发现最佳样本补丁大小为 25 检索研究区15个火灾影响因素的卫星图像数据，构建数据集，提取样本训练模型。这些因素包括海拔、坡向、地表粗糙度、坡度、最低和最高温度、风速、降水、实际蒸散量、气候水分亏缺、NDVI、土地覆盖类型、与河流的距离、与道路的距离和与市区的距离。在训练期间，发现最佳样本补丁大小为 25 检索研究区15个火灾影响因素的卫星图像数据，构建数据集，提取样本训练模型。这些因素包括海拔、坡向、地表粗糙度、坡度、最低和最高温度、风速、降水、实际蒸发量、气候水分亏缺、NDVI、土地覆盖类型、与河流的距离、与道路的距离和与市区的距离。在训练期间，发现最佳样本补丁大小为 25 ×  25 像素。因此，Ensemble 模型实现的最高曲线下面积 (AUC) 为 0.953，其次是 CNN-1，AUC  =  0.902。Ensemble 模型还实现了最佳的准确性、敏感性、特异性、阴性预测值和 F1 分数。最后，预测的敏感性图表明，静态变量可以被视为诱发因素，而动态变量影响预测概率的强度，人为变量的重要作用。这些生成的地图可能有助于在广泛的高风险地区优先进行野火监视和监测。