The Construction of Urban Rainstorm Disaster Event Knowledge Graph Considering Evolutionary Processes

By inergency On Mar 25, 2024

3.1. Analysis of Urban Rainstorm Disaster Events

Heavy rainstorm is a common natural geographical phenomenon that can directly trigger rainstorm disasters. Typically, a rainstorm disaster can set off a series of destructive impacts. The combination of these impacts and the disaster itself constitute an urban rainstorm disaster event. Moreover, an urban rainstorm disaster event is a progressive occurrence, during which the location, area, and intensity of the precipitation are changing constantly. This situation leads to the variation of disaster duration, condition, and intensity of disaster-inducing factors. The interaction between a rainstorm and its related elements collectively shapes the overall urban rainstorm disaster event. The outcomes of such interactions drive the continuous evolution of the disaster event. Notably, urban rainstorm disaster events are distinct from rainstorms or heavy precipitation. A rainstorm does not necessarily lead to an urban rainstorm disaster event, whereas the occurrence of an urban rainstorm disaster event implies that a rainstorm has already occurred or is currently happening.

The evolution of an urban rainstorm disaster event differs from that of a rainstorm. The former is examined from a disaster studies perspective, which emphasizes the series of disaster events caused by precipitation. However, the latter is predominantly described from a meteorological perspective, with a focus on the generation and development of such natural phenomenon. Thus, this study summarizes the evolution of an urban rainstorm disaster event and roughly divides its lifecycle into four stages, namely, pregnant, development, continuous, and decline stages (Figure 1). Specifically, the pregnant stage refers to the period in which a rainstorm gradually transforms into an urban rainstorm disaster event. In this research, the pregnant stage starts from the first precipitation of a disaster event. Within this stage, the natural and human environments (climate, soil, hydrology, population density, and emergency management measures, etc.) collectively contribute to the formation of this rainstorm-related hazard event. With the increase in rainstorms, the disaster event turns into the development stage. During this stage, certain infrastructures suffer damage from the rainstorm, and the increased accumulation of urban water heightens the risk of waterlogging. As time progresses, the rainstorm disaster event moves into the continuous stage. During this stage, the disaster is persistently enhanced, leading to tremendous destruction, and the chain effect triggers various direct and secondary disasters. For example, a heavy rainstorm may lead to a building collapse, which, in turn, can damage roads and further hamper relief efforts. Following the continuous stage, the urban rainstorm disaster event enters the decline stage. During this stage, rainstorm intensity gradually decreases, ultimately ceasing, and post-disaster reconstruction work becomes the primary focus. Given that reconstruction is a gradual process, the end of an urban rainstorm disaster event often lags behind the end of a rainstorm.

3.2. Knowledge Representation Model for Urban Rainstorm Events

Based on urban rainstorm event analyses, a knowledge representation model for these events must be constructed to model the composition elements of disaster events. This model should not only express the different objects involved in urban rainstorm events throughout their lifecycle but also capture the variations between these objects, depicting the changes in the entire urban rainstorm event during different stages.

Therefore, this study adopts a hierarchical framework to divide the composition elements of urban rainstorm events into four categories: event, object–state, feature, and relationship layers (Figure 2). The event layer includes different urban rainstorm events. The object–state layer constrains the types of object included in urban rainstorm events, where each object is a collection of one or more different states. These states describe the specific attribute characteristics of the objects under varying spatiotemporal conditions. The feature layer defines the attributes and behaviors possessed by different objects, as well as the spatiotemporal characteristics contained in the states. Lastly, the relationship layer denotes the types of relationship in the knowledge representation model of rainstorm events. With the proposed model, the fundamental compositions and relationships of urban rainstorm events are uniformly expressed.

Each event in the event layer represents a real-world urban rainstorm event that has already occurred or is expected to occur. The event layer is located at the top level of the model, and within the knowledge graph, all elements of urban rainstorm events can be abstracted as collections of different urban rainstorm event elements. Using events as the base elements for urban rainstorm knowledge offers two advantages. (1) From a logical perspective, events provide a comprehensive overview of urban rainstorm knowledge. Urban rainstorm events involve various types of objects. Given the inherent spatiotemporal complexity of these objects, it becomes challenging to distinguish which object is specifically associated with a particular urban rainstorm event amid multiple urban rainstorm events. (2) From a data perspective, the entities involved in a single urban rainstorm event differ greatly from those in other urban rainstorm events. If real-time data input or knowledge updates are performed for the graph, efficiency will be optimized.

Within the event layer, each event is independent of other events. Let the set of all elements in the urban rainstorm event knowledge graph be denoted by F. Let each urban rainstorm event be denoted by E_n, where n represents the number of urban rainstorm events in the graph. This can be expressed as Equation (1):

F = <E₁, E₂, E₃, …, E_n>,

(1)

2.: Object–State Layer

Objects are nonvolitional entities in reality and constitute significant components of urban rainstorm events. These events not only include the primary object (rainstorm) that triggers the event but also incorporate the natural environment that affects rainstorm changes, as well as various individuals and objects affected by urban rainstorm disasters. According to the basic principles of disaster science, objects involved in the events can be classified into three categories, the disaster-pregnant environment, disaster-inducing factor, and disaster-bearing body, each of which contains multiple subcategories (Table 1).

Let E represent an urban rainstorm event, O be the objects related to urban rainstorm events, and m be the number of objects in the urban rainstorm event process within the graph. E can be represented as Equation (2):

E = <O₁, O₂, O₃, …, O_m>,

(2)

States refer to specific changes, actions, or records that occur to objects over time and space. Each object has at least one state. In this study, the object–state layer divides the urban rainstorm event into multiple independent objects. It allows the portrayal of changes in an urban rainstorm event via modifications in the plurality of objects contained in the event. The object-state layer decomposes each object into individual states. It represents the changes of objects by describing the sets of different states an object can have. This expresses the changes of the event itself through the variations of different objects involved in the event.

Let O represent the objects related to urban rainstorm disasters, S as the states of object O, and u as the number of states contained in an object. This can be represented as Equation (3).

O = <S₁, S₂, S₃, …, S_u>,

(3)

3.: Feature Layer

Features are special symbols or signs that express the identifiable characteristics of objects. This study classifies object features into four categories: time, position, attribute, and behavior. Time (T) and position (P, mainly referring to spatial location in this context) serve as the prerequisites for the existence of objects and establish the basic framework for expressing the evolutionary features of objects. Attributes (A) represent the inherent properties of objects, whereas behaviors (B) describe the various activities and actions generated by objects. Therefore, the representation of object features can be expressed as Equation (4):

Cities are prone to catastrophic rainstorm events, where objects experience varying temporal and spatial states. These objects undergo changes in position, attributes, and behavior as time progresses. Thus, in specific temporal and spatial conditions, the combination of an object’s inherent attributes and behavior leads to its various states. Assuming a certain point in time and position, an object exhibits specific attributes and behavior, which can be represented as Equation (5):

S_O = <k_t, k_P, k_a, k_b>,

(5)

4.: Relationship Layer

Urban rainstorm disaster events constitute a unified whole, with its various elements often being interconnected. Objects can refer to concepts, elements, or characteristics, whereas relationships primarily focus on the connections among multiple objects. The identification of a relationship requires the involvement of at least two objects and can be expressed as Equation (6):
where O_A and O_B represent different objects within an urban rainstorm disaster event, and r signifies the type of relationship existing between O_A and O_B.

The relationships within urban rainstorm disaster events include not only conceptual relationships within the knowledge system but also hierarchical, associative, temporal, spatial, inferential, and mapping relationships among its constituting elements. Particularly, temporal and spatial relationships have subcategories. Specifically, Allen’s interval algebra is employed to classify temporal relationships into 13 types, namely precede, precede by, meet, met by, overlap, overlap by, start, start by, during, contain, finish, finished by, and equal.

The spatial relationships mainly involve topological relationships (using the nine-intersection model), distance relationships (qualitative and quantitative), and orientation relationships (using an eight-cardinal direction system) (Figure 3).

3.3. Knowledge Extraction Model

Knowledge extraction is the basis for constructing a knowledge graph. It can be divided into three types, namely a rule-based or dictionary-based method, semi-supervised identification method, and deep learning model. The rule-based or dictionary-based methods require sufficient prior knowledge, which makes it impossible to establish a complete corpus. The semi-supervised identification methods require a large amount of urban rainstorm disaster knowledge as research support, which poses higher requirements for researchers. In contrast, deep learning models can effectively extract contextual features with superior accuracy and recall rates. Thus, we selected the deep learning models to extract urban rainstorm disaster information.

To acquire more accurate information, we integrated several mainstream models in this study, namely Bidirectional Encoder Representations from Transformers (BERT), Bidirectional Long Short-Term Memory (BiLSTM), and Conditional Random Field (CRF). This hybrid model (BERT–BiLSTM–Attention–CRF) allowed us to capitalize on the complementary strengths of each component, and can extract the disaster entities accurately. Specifically, it aims to leverage the advantages of BERT for contextual representation, BiLSTM for sequential information processing, and CRF for capturing dependencies among extracted entities. The attention mechanism enhances the model’s focus on relevant information, which can improve the overall performance.

The structure of the BERT–BiLSTM–Attention–CRF model is shown in Figure 4 [54]. The initial step involves inputting characters into BERT, which generates a word vector by combining word embedding, segment embedding, and position embedding. Subsequently, the word vector, enriched with semantics, is fed into the BiLSTM network. BiLSTM enables the model to learn temporal information and predict the subsequent output. Following this, a self-attention mechanism layer is utilized to extract local features identified by the BiLSTM network, which captures interaction relationships within the output feature vector. This enhances the global features of the feature vector and complements the output vector’s features from the BiLSTM layer. Lastly, the CRF captures rules governing the interaction between tags, which can ensure the logical predictions, such as avoiding the connection of B-SUBJECT after I-SUBJECT. Moreover, it can enhance the model’s logic of predictive tag sequence for an optimal output. Additionally, the BERT–Attention–CRF model is employed for relationship extraction in the context of urban rainstorm disasters. Such a relatively simple model performs well in the relationship extraction.