Abstract

Natural disasters like bushfires pose a catastrophic threat to the Australia and the world’s territorial areas. This fire spreads in a wide area within seconds, and therefore, it is complicated and challenging to mitigate. To minimize risk and increase resilience, identifying bushfire occurrences beforehand and takes necessary actions is critically important. This study focuses on using deep learning technology for predicting bushfire occurrences using real weather data in any given location. Real-time and off-line weather data was collected using Weather Underground API, from 2012 to 2017 (\(N= 128{,}329\)). The obtained weather data are temperature, dew point, pressure, wind speed, wind direction, humidity, and daily rain. An algorithm was developed to collect this data automatically from any destination. Six different optimizer models were analyzed that use in deep learning technology. Then, the comparison was carried out to identify the best model. Selecting an optimizer for training the neural network, in this case, deep learning is a challenging task. Six best optimizers were chosen to compare and identify the best optimizer to estimate potential fire occurrences in given locations. The six optimizers; Adagrad, Adadelta, RMSprop, Adam, Nadam, and SGD were compared based on their processing time, prediction accuracy and error. Our findings suggest Adagrad optimizer provides less prediction time which is a critical factor for fast-spreading bushfires. Our work provides a data collection model for disaster prediction, which could be utilized to collect climatic characteristics and topographical characteristics in with larger samples. The developed methodology could be utilized as a natural disaster prediction model for precise predictions with less error and processing time using real-time data. This study provides an enhanced understanding of finding the locations that fire starts or spot fires which are more likely to occur, and lead to identifying of fire starts that are more likely to spread.

Graphic abstract

中文翻译：

---

使用深度学习预测林区大火发生的最佳优化器选择

摘要

丛林大火等自然灾害对澳大利亚和世界领地构成灾难性威胁。火灾在几秒钟内在广阔的区域蔓延，因此，扑救工作复杂且具有挑战性。为了最大程度地降低风险并提高应变能力，事先确定森林大火的发生并采取必要的措施至关重要。这项研究的重点是使用深度学习技术通过任何给定位置的真实天气数据来预测森林大火的发生。使用Weather Underground API收集了2012年至2017年的实时和离线天气数据（\（N = 128 {，} 329 \））。获得的天气数据是温度，露点，压力，风速，风向，湿度和每日降雨。开发了一种算法来自动从任何目的地收集此数据。分析了深度学习技术中使用的六个不同的优化器模型。然后，进行比较以确定最佳模型。选择用于训练神经网络的优化器，在这种情况下，深度学习是一项艰巨的任务。选择了六个最佳优化器来比较和确定最佳估计器，以估计给定位置的潜在火灾发生。六个优化器；根据Adagrad，Adadelta，RMSprop，Adam，Nadam和SGD的处理时间，预测准确性和误差进行了比较。我们的发现表明阿达格勒优化程序提供较少的预测时间，这是快速蔓延的丛林大火的关键因素。我们的工作为灾难预测提供了一个数据收集模型，该模型可用于收集较大样本中的气候特征和地形特征。所开发的方法可以用作自然灾害预测模型，从而使用实时数据以更少的错误和更少的处理时间进行精确的预测。这项研究可以更好地了解如何查找起火地点或发现更可能发生火灾的地点，并有助于确定更可能蔓延的起火地点。

图形摘要