Urban Fire Situation Forecasting: Deep sequence learning with spatio-temporal dynamics

https://doi.org/10.1016/j.asoc.2020.106730Get rights and content

Highlights

  • It is the first exploration to process large-scale urban fired dataset with deep sequence learning models.

  • We first proposed FSFN to capture the spatio-temporal latent dynamics of fires.

  • We proposed Adversarial FSFN-A based on FSFN, which incorporates the offline geographical, social attributes and spatio-temporal dependencies.

Abstract

Understanding the evolving discipline of urban fire situations is a basic but challenging task for urban security and fire-fighting decisions. Traditional methods forecast the urban fire situation through mathematical modeling and statistical learning, which could be interpretable but generally lack of efficiency and practicality. Recently, some deep neural network methodologies, especially convolutional neural network (CNN) and recurrent neural network (RNN), are presented as paradigms to capture dynamics in spatial–temporal complex phenomenon, which tally with the characteristics of fire situation forecasting. In this paper, we propose a novel deep sequence learning model as the fire situation forecasting network (FSFN) to better process the information and spatio-temporal correlations in regional urban fire alarm dataset. FSFN model integrates structures of Variational auto-encoders and context-based sequence generative model Seq2seq to obtain the latent representation of the fire situation and learn the spatio-temporal dynamics. Furthermore, we augment the network structure of FSFN from a simple deep sequence generative model to adversarial fire situation forecasting network with auxiliary information(Adversarial FSFN-A). The experimental studies demonstrate the effectiveness of Adversarial FSFN-A has superior spatio-temporal distribution prediction of multi-type urban fire situation.

Introduction

According to the official fire situation report, Fire Loss in the United States During 2017, from National Fire Protection Association, United States fire departments responded to an estimated 1,319,500 fires in 2017. These fire incidents resulted in 3400 civilian fire fatalities, 14,670 civilian fire injuries and an estimated $23 billion in direct property loss. Videlicet, there was a civilian fire death every 2 h and 34 min and a civilian fire injury every 36 min in 2017. Among them, nearly 60% of the fires occurred in urban areas [1]. Compared with wildfires, urban fires are smaller but more difficult to predict, since they always feature with dynamics and uncertainties. It can be seen that urban fire forecasting and prevention has become a long term but significant task for urban security and smart cities construction [2].

Space and time are the two indispensable dimensions of urban fire forecasting, but have not yet been investigated adequately. In the space perspective, a city can be divided into residential districts, commercial districts, industrial districts and so on. Districts with higher-level fire-fighting conditions tend to maintain a lower ratio in fire incidents. Meanwhile, the incentives of fires are different in each district and may be affected by neighbor districts such as chemical fires in chemical plants, circuit fire in the industrial district and boiler fire in residential areas, which reveals the complexity of space correlations [3]. In the time dimension, for instance, the hot and dry seasons are more prone to fire than usual. During the day, working hours in industrial districts may have a higher probability of fire than non-working hours at night due to equipment’s operation or other reasons. The difficulties of forecasting urban fires are modeling spatio-temporal dynamics and short-time predictions. In some recent literature, urban fire theories are attractively revisited, such as building fire spread theory [4], social and economic characteristics theory [5], fire education theory [6]. Besides, another trend of developing methods is evolving from the social mathematical model to the data-driven statistical learning model [7], [8], [9]. However, traditional statistical learning models are still difficult to capture complex spatio-temporal correlations and short-time predictions.

In this paper, we present a series of novel deep sequence learning approaches, which introduces latent space sequence generation model to cope with urban fire forecasting. We initially arrange different types of spatio-temporal fire data into different sequences of Fire Situation Awareness Graphs (FSAGs) in the first step. Then, Convolutional Variational Auto-Encoder (VAE) [10] is applied to embed the spatial information from the FSAGs into latent space to generate soft computational spatial representation. Meanwhile, temporal correlation can be tackled by Seq2seq model [11], which can capture long-time dependencies from some previous FSAGs to some future FSAGs. Subsequently, the spatial and temporal dynamics correlation has been established in Fire Situation Forecasting Network (FSFN). To improve the performance of fire situation forecasting, we involve the adversarial loss [12] as well as some auxiliary information to strengthen the reality and confidence of predicted FSAGs. Adversarial FSFN with Auxiliary Information (Adversarial FSFN-A) model was proposed based on FSFN. In Adversarial FSFN-A model, geographic information, timestamp information and auxiliary social attributes are integrated. To summarize, we make the following main contributions:

  • To the best of our knowledge, it is the first exploration to process large-scale dataset of urban fire situation in the period of five years with deep sequence learning approaches.

  • We first proposed FSFN, a short-time deep sequence learning urban fire forecasting framework to capture the spatio-temporal dynamics of different fire types in latent FSAG form.

  • We proposed a hybrid deep sequence learning model Adversarial FSFN-A based on FSFN, which incorporates the offline geographical, social attributes and spatio-temporal dependencies with deep fusion.

  • We evaluate our approaches on the real-word datasets. The results show that our approach reduces the almost all evaluation metrics error compared to traditional state-of-art methods, which demonstrates that the Adversarial FSFN-A model has superiority in urban fire forecasting.

In the remainder of the paper, we begin with a review of the related work on the traditional urban fire forecasting methods in Section 2. Then we review background of VAE, Seq2seq and generative adversarial network (GAN) in Section 3. The dataset processing and research methodology adopted in this study are presented in Sections 4 Dataset processing, 5 Proposed methodology respectively , followed by the experimental results and analysis in Section 6. The final section concludes the achievements of this research and proposes future work.

Section snippets

Related work

Early methods focus on fire forecasting were based on some simulation, stochastic process modeling and some heuristic algorithms. Asensio et al. investigated the simulation model for fire spread [13], which presented a convection model coupled with fire propagation models to take into account the wind and the slope, two of the most relevant factors affecting surface fire spread. In [14], a sensor fusion technique with mathematical modeling was applied for fire detection and fire location

Prerequisite

In this section, we briefly call the related methods of this research. In particular, the basic principles of Variational Auto-Encoder, Sequence Generative Model and Generative Adversarial Network. Some commonly-used abbreviations and notations are shown in Table 1.

Variational Auto-Encoder: Variational Auto-Encoder is a deep unsupervised generative approach for disentangled factor learning that can automatically discover the independent latent factors of variation in unsupervised data, proposed

Dataset processing

In this section, we present some details about data processing with urban fire situation awareness graph (FSAG), which is the most fundamental element in spatiotemporal prediction tasks.

The urban fire dataset in this research is sorted from urban database managed by San Francisco government [43], whose region is defined in San Francisco city. The incident reports released from local Fire Department recorded all of the urban fire calls since 2005. Though the dataset is still updating, we only

Proposed methodology

In this section, we provide details of our proposed baseline Fire Situation Forecasting Network (FSFN) and Adversarial FSFN with Auxiliary Information framework.

Experiments and analysis

In this section, the training details are described, and some numerical analysis are presented to show the performance of our model.

Conclusion

In this study, we gain deep comprehension into the urban fire situation modeling and a series of novel deep sequence learning methods are proposed for fire prediction. Though it is still difficult to precisely forecast urban fire situation in both numerical scales and spatiotemporal distribution scales, the experimental results of Adversarial FSFN-A are remarkable in comparison to other traditional models in this domain and implies that combination of auxiliary information, latent sequential

CRediT authorship contribution statement

Guangyin Jin: Conceptualization, Methodology, Software, Visualization, Writing - original draft. Qi Wang: Conceptualization, Methodology. Cunchao Zhu: Writing - original draft, Writing - review & editing. Yanghe Feng: Supervision, Project administration. Jincai Huang: Supervision. Xingchen Hu: Writing - review & editing, Validation.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (60)

  • WangQ. et al.

    Csan: A neural network benchmark model for crime forecasting in spatio-temporal scale

    Knowl.-Based Syst.

    (2020)
  • ZhangG.P.

    Time series forecasting using a hybrid arima and neural network model

    Neurocomputing

    (2003)
  • Urban fire and life safety task force

    (2017)
  • AngelidouM. et al.

    Enhancing sustainable urban development through smart city applications

    J. Sci. Technol. Policy Manag.

    (2018)
  • McCaffreyS.M.

    Fighting fire with education: what is the best way to reach out to homeowners?

    J. For.

    (2004)
  • Singh WaliaB. et al.

    A dynamic pipeline for spatio-temporal fire risk prediction

  • OliveiraM. et al.

    Evaluation procedures for forecasting with spatio-temporal data

  • KingmaD.P. et al.

    Auto-encoding variational bayes

    (2013)
  • SutskeverI. et al.

    Sequence to sequence learning with neural networks

  • M. Zhu, A. Alperovich, O. Johannsen, A. Sulc, B. Goldluecke, An epipolar volume autoencoder with adversarial loss for...
  • ZervasE. et al.

    Fire detection in the urban rural interface through fusion techniques

  • QuartieriJ. et al.

    A cellular automata model for fire spreading prediction

    Latest Trends Urban Plan. Transp.

    (2010)
  • ArdiantoR. et al.

    Modeling spatial–temporal dynamics of urban residential fire risk using a markov chain technique

    Int. J. Disaster Risk Sci.

    (2019)
  • PreislerH.K. et al.

    Statistical model for forecasting monthly large wildfire events in western united states

    J. Appl. Meteorol. Climatol.

    (2007)
  • CortezP. et al.

    A Data Mining Approach to Predict Forest Fires Using Meteorological Data

    (2007)
  • ChengT. et al.

    Integrated spatio-temporal data mining for forest fire prediction

    Trans. GIS

    (2008)
  • FernandesL.C. et al.

    Fire risk modeling using artificial neural networks

  • MuhammadK. et al.

    Efficient deep cnn-based fire detection and localization in video surveillance applications

    IEEE Trans. Syst. Man Cybern.: Syst.

    (2018)
  • ZhaoL. et al.

    Multi-task learning for spatio-temporal event forecasting

  • WangB. et al.

    Deep learning for real-time crime forecasting and its ternarization

    (2017)
  • Cited by (18)

    • Deep multi-view graph-based network for citywide ride-hailing demand prediction

      2022, Neurocomputing
      Citation Excerpt :

      Considering the long-term and short-term dynamics of ride-hailing demand, deep models can be designed to capture short-term dependencies and long-term dependencies respectively. However, in most previous traffic prediction tasks, RNN-based models were still the most popular methods [16,18,28–35], which can only have enough learning capability for short-term dependencies. In this paper, we propose Deep Multi-View Spatio-temporal Virtual Graph Neural Network (DMVST-VGNN), which is stacked by three hierarchical structures, short-term temporal dynamics view, spatial dynamics view and long-term temporal dynamics view.

    • Adaptive Dual-View WaveNet for urban spatial–temporal event prediction

      2022, Information Sciences
      Citation Excerpt :

      The video-like input format is compatible with CNN and the spatial–temporal joint features can be captured by the deep neural network stacked with CNNs and RNNs. Many previous works have demonstrated the effectiveness of these hybrid deep learning models [10–15]. However, the most significant limitation of the video-like modeling approach is that only the geographic proximity feature is considered.

    View all citing articles on Scopus
    View full text