Urban Fire Situation Forecasting: Deep sequence learning with spatio-temporal dynamics
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)
- et al.
Special issue on spatial analytical approaches in urban fire management
Fire Saf. J.
(2013) - et al.
Dynamic modeling of fire spread in building
Fire Saf. J.
(2011) Social and economic characteristics as determinants of residential fire risk in urban neighborhoods: A review of the literature
Fire Saf. J.
(2013)- et al.
Spatial forecasting of residential urban fires: A bayesian approach
Comput. Environ. Urban Syst.
(2010) - et al.
A convection model for fire spread simulation
Appl. Math. Lett.
(2005) - et al.
Evolutionary-statistical system: A parallel method for improving forest fire spread prediction
J. Comput. Sci.
(2015) - et al.
Exploratory and inferential methods for spatio-temporal analysis of residential fire clustering in urban areas
Fire Saf. J.
(2013) - et al.
Urban fire risk clustering method based on fire statistics
Tsinghua Sci. Technol.
(2008) - et al.
Reduced frequency and severity of residential fires following delivery of fire prevention education by on-duty fire fighters: Cluster randomized controlled study
J. Saf. Res.
(2012) - et al.
Urban ride-hailing demand prediction with multiple spatio-temporal information fusion network
Transp. Res. C
(2020)
Csan: A neural network benchmark model for crime forecasting in spatio-temporal scale
Knowl.-Based Syst.
Time series forecasting using a hybrid arima and neural network model
Neurocomputing
Urban fire and life safety task force
Enhancing sustainable urban development through smart city applications
J. Sci. Technol. Policy Manag.
Fighting fire with education: what is the best way to reach out to homeowners?
J. For.
A dynamic pipeline for spatio-temporal fire risk prediction
Evaluation procedures for forecasting with spatio-temporal data
Auto-encoding variational bayes
Sequence to sequence learning with neural networks
Fire detection in the urban rural interface through fusion techniques
A cellular automata model for fire spreading prediction
Latest Trends Urban Plan. Transp.
Modeling spatial–temporal dynamics of urban residential fire risk using a markov chain technique
Int. J. Disaster Risk Sci.
Statistical model for forecasting monthly large wildfire events in western united states
J. Appl. Meteorol. Climatol.
A Data Mining Approach to Predict Forest Fires Using Meteorological Data
Integrated spatio-temporal data mining for forest fire prediction
Trans. GIS
Fire risk modeling using artificial neural networks
Efficient deep cnn-based fire detection and localization in video surveillance applications
IEEE Trans. Syst. Man Cybern.: Syst.
Multi-task learning for spatio-temporal event forecasting
Deep learning for real-time crime forecasting and its ternarization
Cited by (18)
Urban hotspot forecasting via automated spatio-temporal information fusion
2023, Applied Soft ComputingDeep multi-view graph-based network for citywide ride-hailing demand prediction
2022, NeurocomputingCitation 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 SciencesCitation 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.