Keywords

1 Introduction

Landslides are one of the deadliest natural catastrophes with the potential for serious socioeconomic devastation, including loss of life and damage of property. Landslides occur in a diversity of landscapes around the world and rainfall is often the most prevalent trigger (Crosta and Frattini 2008; Tran et al. 2021). For example, the heavy rain caused by Hurricane Mitch in October and November 1998 induced thousands of landslides in Honduras, El Salvador, Guatemala, and Nicaragua (Bucknam et al. 2001). About 147 slope failures and 33 debris flows were recorded around Umyeon mountain in South Korea in July 27, 2011 due to heavy rainfall (Viet et al. 2016).

In general, efficient planning and management might significantly lessen the damage caused by landslides (Sultana and Tan 2021). To mitigate the landslide disaster, especially human casualties and injuries, the most effective method is probably the early identification of the locations and time of landslide occurrence. Brabb (1991) concluded that more than 90% of the damages could be avoided if landslides could be warned early.

Cetina and Nadim (2008) defined a landslide early warning system as “monitoring devices designed to avoid, or at least to minimize the impact imposed by a threat on humans”. Establishing an early warning system for landslide occurrence (location and time) plays a vital role in managing, preventing, and mitigating probable hazards and damages. There are several systems that integrated landslide susceptibility maps, rainfall thresholds, real-time and forecasted rainfall data with WebGIS (Zhang et al. 2011; Ahmed et al. 2018). While physically-based models are usually appropriate for small areas (a single slope or a small catchment), the empirical approach can be applied to much larger areas (city, province, country) if the historical data are available (Guzzetti et al. 2020).

Vietnam is one of the nations that have been deeply affected by natural hazards and climate change (Le et al. 2021). Principally, mountainous areas often suffer from abnormal weather phenomena causing many extreme rainfall events and accelerating the process of natural disasters. This trend results in the urgent requirement for effective effort in order to prevent and mitigate damage. However, one of the barriers to researching and developing a landslide early warning system in Vietnam is the heterogeneity and difficulty in data collection, especially data on historic landslide events (Hung et al. 2017).

To develop a landslide early warning system for Ha Long City, the landslide susceptibility map and rainfall thresholds triggering landslides need to be achieved. To perform these tasks, information on landslide locations (for landslide susceptibility mapping) and landslide occurrence times (for identifying rainfall thresholds) needs to be collected. However, the rapid urbanization in Ha Long City renders it more difficult to determine the locations of landslide occurrences. According to the survey results of the State-Funded Landslide Project (SFLP): “Investigation, assessment, and warning zonation for landslides in the mountainous regions of Vietnam” in 2015, only 20 landslide locations inside Ha Long City can be identified and surveyed. In contrast, numerous landslide spots (494 in total) may still be reliably found in other regions of Quang Ninh Province (Fig. 1). The time of occurrence is critical for determining the rainfall threshold triggering landslides (at least the date of landslide occurrence). However, most landslide sites in Vietnam are only properly statisticized and recorded if they result in severe damages such as human deaths, building collapses, and traffic bottlenecks on key and essential roads. As a result, the occurrence time of landslides may only be gathered if they occur in densely populated residential regions such as Ha Long City.

Fig. 1
A map of Quang Ninh province of Vietnam presents the network of rods and rivers. Several dots on the map indicate the locations of the landslides. The Ha Long is bordered by a dark line.

Landslide locations in Ha Long City and Quang Ninh Province

Full size image

Because of this discrepancy, the authors had to create a landslide susceptibility map for the whole of Quang Ninh province (Ha Long City included) and establish a dedicated rainfall threshold for Ha Long City.

2 Study Area

Being a province situated in the northeast of Vietnam, Quang Ninh features quite a diverse topography with types of coastal plains and swamps, midland hills, and mountains. The province has a typical tropical climate with a hot, rainy, and humid summer (from May to October) and cold, dry winter (from November to April). The province receives an average of 1995 mm of precipitation every year, the majority of which falls during the summer (80–85% of the total annual precipitation). Due to the influence of several natural, environmental, and societal variables, the mountainous region of Quang Ninh province is prone to numerous natural catastrophes, particularly landslides.

Ha Long is the capital city of Quang Ninh Province. With the benefit of coal resources and a favorable location for tourism, and trade, Ha Long City has recently had the fastest growth and urbanization in Vietnam. The topography of Ha Long City is also quite diverse and complex, including mountains, delta, coastal, and island. The hilly and mountainous areas are mainly in the north and northwest and account for about 70% of the land area. According to the Quang Ninh Department of Construction assessment, many households in Ha Long City live in landslide-prone areas (DOC 2016). Therefore, developing a system for landslide early warning in Ha Long City is very urgent.

3 Methodology and Results

3.1 Landslide Susceptibility Map by Spatial Multi-criteria Evaluation (SMCE) Method

The SMCE method is a semi-quantitative evaluation technique that is often used to determine the landslide susceptibility index (LSI). This method was developed from the analytical hierarchical process (AHP), and it can support users in performing multi-criteria assessment in a spatial model.

Technically, the SMCE method overcomes the limitations of quantitative assessment methods which require a large amount of accessible and comprehensive data. The method is heavily reliant on expert opinion, since each landslide causal component is employed as a mapping parameter, and each of those mapping parameters is assessed based on the subjective experience of experts (Castellanos 2008). The difference between semi-quantitative and qualitative methods is the association of the weight under certain criteria. In fact, the semi-quantitative technique has an advantage when the timing of occurrence of landslides cannot be determined and the available numeric data is restricted (Castellanos Abella and Van Westen 2007).

Input maps of the SMCE method are used as representative spatial criteria, which are grouped, normalized, and weighted in a “criteria tree”. The output of the SMCE method consists of one or more “composite index maps” that show the results of simulations (Castellanos Abella and van Westen 2007).

The SMCE can be represented, as a matrix as depicted in the diagram in Fig. 2, whereas:

  • Matrix A includes criteria (factors that influent the landslide initiation) represented on an axis (C1 to Cn) and a list of possible alternatives (to establish a hierarchy, score, assign weights, etc.). The decision is made from the set of criteria (the result of matrix A) showing a different axis (A1 to Am).

  • Each cell in the matrix (aij) represents performing a specific alternative for each specific criterion. The value of each cell in the matrix is the product of the multiplication of the value (normalized to values between 0 and 1) of each criterion for a particular alternative and the weight corresponding to each criterion (W1 to Wn).

  • As long as the cells in the matrix have been filled (all criteria have been evaluated), the final value equals the sum of all the cell values of the different criteria for each particular alternative (for example sum of the values of cells a11 to a1n for alternative A1).

The process of landslide susceptibility mapping for Quang Ninh Province follows 08 main steps as below:

  1. (1)

    Collect available information, data, and spatial factor maps for the study area

  2. (2)

    Compile and normalize all selected spatial factor maps and their attributes

  3. (3)

    Rasterize factor maps and divide them into different groups of criteria

  4. (4)

    Structure the problem into a criteria tree by assigning the spatial data that is relevant for each of the criteria

  5. (5)

    Standardize factor maps and design value functions using SMCE model in which the map or column values are converted to values between 0 and 1

  6. (6)

    Determine the weights of the criteria within one group and among groups using SMCE tools such as direct weights, pairwise comparison and rank ordering

  7. (7)

    Calculate and evaluate the output map as Landslide Susceptibility Index (LSI) map

  8. (8)

    Classify the results to generate the final map as the landslide susceptibility map

From the results of document collection and synthesis, and on the basis of the completeness of the factor maps in terms of data structure and attributes, nine landslide factor maps were selected as the main input data for landslide susceptibility assessment by the SMCE method. These factors are: Elevation, slope, land cover, amplitude of relief, lineament density, drainage density, groundwater availability, and weathering crust type. The maps of these factors were collected from the SFLP project at 1:10,000 and 1:50,000 scales and applied to the SMCE module using ArcGIS software with the exact resolution of 20 m (Figs. 3, 4, 5, 6, 7, 8, 9, 10 and 11).

Fig. 2
An illustration presents the S M C E method used to determine the landslide susceptibility index. A matrix has rows from A 1 to A m and columns from C to C n. A formula defines the A i j. On the right side, many indicators A i 1 to A m are arranged vertically.

Schematic procedure of the Spatial Multi-Criteria Evaluation (SMCE) method (Castellanos Abella and Van Westen 2007)

Full size image
Fig. 3
A heat map of the Quang Ninh Province of Vietnam presents the elevation of the different regions. Most of the regions have elevations between negative 101.07 and 94.52 meters.

Landslide factor map: Elevation

Full size image
Fig. 4
A heat map of the Quang Ninh Province of Vietnam presents the slope of the different regions. Most of the regions have slopes between negative 5.86 and 14.52 degrees.

Landslide factor map: Slope

Full size image
Fig. 5
A heat map of the Quang Ninh Province of Vietnam presents the drainage density of the different regions. Most of the regions have densities between 0.94 and 5.4 degrees.

Landslide factor map: Drainage density

Full size image
Fig. 6
A heat map of the Quang Ninh Province of Vietnam presents the groundwater availability of the different regions. The water availability is divided into 4 categories, very low, low, medium, and high. A major area of the province has very low groundwater availability.

Landslide factor map: Groundwater availability

Full size image
Fig. 7
A heat map of the Quang Ninh Province of Vietnam presents the land cover of the different regions. Most of the areas are covered with forests.

Landslide factor map: Land cover

Full size image
Fig. 8
A heat map of the Quang Ninh Province of Vietnam presents the weathering crust types of the different regions. Most of the regions fall under the Q and S A F S A types.

Landslide factor map: Weathering crust type

Full size image
Fig. 9
A heat map of the Quang Ninh Province of Vietnam presents the engineering geology of the different regions. Most areas have medium-engineering geology.

Landslide factor map: Engineering geology

Full size image
Fig. 10
A heat map of the Quang Ninh Province of Vietnam presents the Lineament density of the different regions. Most areas fall into the density range of medium to very high in the center. Near the boundaries, the density is either low or very low.

Landslide factor map: Lineament density

Full size image
Fig. 11
A heat map of the Quang Ninh Province of Vietnam presents the amplitude of relief of the different regions. Most areas have an amplitude of relief between negative 27.90 and 91.69 meters per square kilometer.

Landslide factor map: Amplitude of relief

Full size image

The role of nine selected factors susceptible to landslides in Quang Ninh Province has been analyzed and evaluated based on summarizing the assessment opinions of 02 groups of experts: field survey staff and professional research staff. The assessment of the influential role of the input factors for the SMCE model in the Quang Ninh area is summarized in the “criteria tree” presented in Fig. 12. The map layers are considered as the “main criteria”, and the attributes of each map layer are considered as “sub-criteria”. The susceptibility of each criterion to the landslide is expressed by the weight W. The values of weight W are calculated by the Analytic Hierarchy Process AHP (Saaty 1977, 2000), on the basis of pairwise comparisons between the “main criteria” and between the “sub-criteria” of each “main criteria”.

Fig. 12
A tree diagram presents the 9 factors that influence landslide susceptibility. Groundwater, engineering geology map, lineament density, land cover, slope, elevation, amplitude of relief, drainage density, and weathering crust type.

Criteria tree for criteria weight (W), showing the influence on landslide occurrence of 09 factors

Full size image

Based on the natural distribution of susceptibility values, four thresholds of susceptibility index to landslide have been determined to classify the LSI map (output of SMCE module) into five susceptibility classes: Very low, low, moderate, high, and very high (Table 1). The landslide susceptibility maps of Quang Ninh Province and Ha Long City were generated, as shown in Figs. 13 and 14.

Table 1 LSI thresholds for generating landslide susceptibility map in Quang Ninh Province
Full size table
Fig. 13
A heat map of the Quang Ninh Province of Vietnam presents the landslide susceptibility of the different regions. Most regions are exposed to medium to very-high landslide susceptibility.

Landslide susceptibility map of Quang Ninh Province

Full size image
Fig. 14
A heat map of Ha Long City presents the landslide susceptibility of the different regions. Most regions are exposed to very low landslide susceptibility.

Landslide susceptibility map of Ha Long City

Full size image

The landslide locations (identified through field survey and aerial image interpretation) (Fig. 1) were employed to validate the accuracy of the landslide susceptibility map. The distributions of the predicted landslide susceptibility classes (very low, low, moderate, high, and very high) in the susceptibility map compared with the observed landslides in Quang Ninh Province and Ha Long City were summarized in Tables 2 and 3 respectively. The result shows that 323 landslide events (63% of the recorded landslides) occurred in the zones with high and very high susceptibility, although these zones account for only 29% of the total area of Quang Ninh Province. Only 18 landslide events (4%) occurred in low susceptibility areas covering around 2916 km2 (48%) of the province area. Considering Ha Long City only, 40% (8/20) of observed landslides are distributed in the two most susceptibility zones that cover 8% of the city area.

Table 2 The distribution of the predicted landslide susceptibility classes and compared with the landslide inventory map in Quang Ninh Province
Full size table
Table 3 The distribution of the predicted landslide susceptibility classes and compared with the landslide inventory map in Ha Long City
Full size table

In the comparison tables for both Quang Ninh Province and Ha Long City, the largest proportions of landslides are located in the moderate susceptibility zones (34% in Quang Ninh Province and 50% in Ha Long City). However, it is worth noting that most of these landslide locations are on excavated slopes (for road and house constructions).

These results are consistent with the actual situation of the landslide occurring in Quang Ninh Province and Ha Long City. It shows that the landslide susceptibility map for Quang Ninh Province has high reliability to be applied in the early warning system for Ha Long City.

3.2 Empirical Rainfall Threshold

The empirical approach is applied to identify rainfall thresholds by analyzing historical rainfall events that triggered and did not trigger landslides. The main principle of this approach involves the assumption about the past relationship representing the rainfall-initiated landslides that will also induce landslide events in the future (Gioia 2013).

The method to identify landslide rainfall threshold includes three main steps as follows:

Step 1) Collect and process data:

  • Collect information related to landslides in Ha Long City from mass media, field investigation, and previous studies.

  • Collect rainfall data in the past. The rainfall data (hourly and daily) in rainy seasons (May–October) from 2005 to 2021 was used in this study. It comes from the Bai Chay National Hydro-meteorological station in Ha Long City’s center.

Step 2) Examine the correlation between rainfall data and landslide events in the study area from 2005 to 2016.

  • Collecting hourly and daily rainfall data during and before the dates of landslide occurrences.

  • Determining the relationship between the daily rainfall (on the dates of landslide occurrence) and the antecedent rainfall values (from 1 to 60 days).

  • Determining the relationship between maximum hourly and daily rainfall and cumulative rainfall event.

  • Generating the rainfall thresholds for the initiation of landslides based on the relationship between rainfall data.

Step 3) Evaluate the accuracy of the rainfall thresholds generated.

  • Based on rainfall data and landslide events in the 2017–2021 period, the appropriateness of the rainfall thresholds determined in Step 2 will be verified.

In Ha Long City, 26 rainfall-triggered landslide incidents could be accurately dated between 2005 and 2021 (refer to Table 4 for details). These occurrences all took place during the rainy season, and each downpour is deemed as having ended if there is no rain after 24 consecutive hours. The values of rainfall events (whether triggering or not specified to induce landslides) such as daily rainfall (R), maximum hourly rainfall (Ihmax), maximum daily rainfall (Idmax), rainfall duration (D), cumulative event rainfall (E) and antecedent rainfall for different durations (1–60 days) were extracted and analyzed.

Table 4 Landslide events and related rainfall data (from 2005 to 2021)
Full size table

The data in the rainy season of the 2005–2016 period will be used to identify rainfall thresholds. During this period, 418 rainfall events were recorded at Bai Chay national hydro-meteorological station, of which 16 events were identified as triggering landslides. The correlations of maximum daily/hourly rainfall (Idmax/Ihmax) and cumulative event rainfall (E) with landslides are shown quite clearly in Figs. 15 and 16. Information on landslide events (occurrence date, maximum hourly rainfall (Ihmax), maximum daily rainfall (Idmax), cumulative event rainfall (E) is presented in Table 4.

Fig. 15
A scatter plot of maximum daily rainfall intensity versus cumulative event rainfall. The maximum daily rainfall intensity is negative 0.47 times E plus 177. The solid dots represent the landslide, and the blank dots represent no landslide. A major part of the dataset for landslides is above (200, 100).

Relationship between maximum daily rainfall intensity (Idmax) and cumulative event rainfall (E) (Landslide and rainfall data from 2005 to 2016)

Full size image
Fig. 16
A scatter plot of maximum hourly rainfall intensity versus cumulative event rainfall. The maximum hourly rainfall intensity is negative 0.23 times E plus 70. The solid dots represent the landslide, and the blank dots represent no landslide. A major part of the dataset for landslides is above (100, 50).

Relationship between maximum hourly rainfall intensity (Ihmax) and cumulative event rainfall (E) (Landslide and rainfall data from 2005 to 2016)

Full size image

These landslide-related rainfall events mostly had significantly higher maximum daily/hourly rainfall and cumulative event rainfall than other events. This trend is consistent with the findings from Onodera et al. (1974), Canon et al. (2001) and the fact that heavy or prolonged rainfall is the most important trigger. Following earlier research which adopted a regression model to manually draw the thresholds (Marques et al. 2008; Elias et al. 2020), the threshold Td (based on Idmax and E) and the threshold Th (based on Ihmax and E) were proposed for Ha Long City:

$$ \mathrm{Threshold}\ \mathrm{Td}:\mathrm{Idmax}=-{0.47}^{\ast }\ \mathrm{E}+177 $$
(1)
$$ \mathrm{Threshold}\ \mathrm{Th}:\mathrm{Ihmax}=-{0.23}^{\ast }\ \mathrm{E}+70 $$
(2)

Parameters of the threshold Eqs. (1) and (2) were estimated based on lines delimited manually in the scatter plots in Figs. 15 and 16, respectively. In general, depending on the priority given to the accuracy of predicting non-landslide or landslide occurrences, empirical thresholds could be slightly adjusted to be closer to the part of the non-triggering rainfall events or to the part of triggering ones. Nevertheless, too low thresholds will lead to too many fail alarms, while too high will result in additional damage (Staley et al. 2012). For Ha Long City, due to the limitation of collected landslide data and for safety purposes, thresholds were fitted to the region of non-triggering rainfall events.

Among 23 rainfall events from 2005 to 2016 in Ha Long city exceeding the threshold Td, 14/16 landslide events (87.5%) could be correctly warned, and only 09 rain events (39.1%) did not trigger landslides in the vicinity. Even so, two rainfall events (Nos. 8 and 14) were reported as causing landslides in reality, although their (E, Idmax) data points remain considerably below the Td threshold.

Among 35 rainfall occurrences surpassing the threshold Th between 2005 and 2016, 100% (16/16) of landslide events could be reliably predicted, but 54.3% (19/35) of rainfall events could not be defined as connected to landslide dangers. The maximum daily rainfall and cumulative event rainfall of the two rainy events Nos. 8 and 14 were not really high (less than 100 mm). However, because to the height of maximum hourly rainfall, they could be properly projected with the threshold Th.

Data of 190 rainfall events in the rainy seasons from 2017 to 2021 were used in combination with ten landslide events to validate identified rainfall thresholds Td and Th. These data are shown and compared with thresholds Td (Fig. 17) and Th (Fig. 18).

Fig. 17
A scatter plot of maximum daily rainfall intensity versus cumulative event rainfall. The maximum daily rainfall intensity is negative 0.47 times E plus 177. The solid dots represent the landslide, and the blank dots represent no landslide. A major part of the dataset for landslides is above (300, 140).

Relationship between maximum daily rainfall intensity (Idmax) and cumulative event rainfall (E) (Landslide and rainfall data from 2017 to 2021)

Full size image
Fig. 18
A scatter plot of maximum hourly rainfall intensity versus cumulative event rainfall. The maximum hourly rainfall intensity is negative 0.23 times E plus 70. The solid dots represent the landslide, and the blank dots represent no landslide. A major part of the dataset for landslides is above (200, 40).

Relationship between maximum hourly rainfall intensity (Ihmax) and cumulative event rainfall (E) (Landslide and rainfall data from 2017 to 2021)

Full size image

Among a total of 190 rainfall events during the test period, 17 events surpassed the threshold Td, including 9 events (52.9%) that were not identified to be connected with landslides. Meanwhile, using the Td threshold, 8/10 rain-induced landslide incidents (80%) may be appropriately warned. When the threshold Th was applied to these rainfall and landslide events, the percentage of over-threshold but unidentified occurrences related with landslides climbed to 56.5% (13/23 events). The use of threshold Th in the warning system, on the other hand, yields a safer outcome, with 100% of rainfall-triggered landslide occurrences being able to be notified in advance.

Four possible cases of outcomes or contingencies (Staley et al. 2012) when a rain threshold is applied: a true positive (TP) is an event where the data point is above the rainfall threshold, and at least 01 landslide was recorded; a true negative (TN) is a rain event with data point is below the rainfall threshold, and no landslides were induced; a false positive (FP) is an event where the data point exceeds the rainfall threshold but did not trigger a landslide; a false negative (FN) is an event where the data point is below the rainfall threshold, yet a landslide was recorded.

To evaluate the accuracy of two rain thresholds Td and Th, based on common methods used to assess accuracy (Segoni et al. 2018; Staley et al. 2012), four skill scores are determined: (1) the probability of detection (POD = TP/(TP+FN)); (2) the probability of false detection (POFD = FP/(FP+TN)); (3) the probability of false alarm (POFA) and (4) The Hanssen and Kuipers (1965) skill score which calculates the accuracy of predicting events with and without landslides (HK = POD − POFD).

Contingencies and skill scores calculated for rainfall thresholds Td and Th are present in Tables 5 and 6. It can be seen that the POFD and POFA skill scores for thresholds Td and Th are quite similar. However, the POD and HK skill scores applied with the threshold Th are significantly higher than these scores with the threshold Td. Therefore, to be employed in the early warning system for Ha Long City, the rainfall threshold Th should be prioritized.

Table 5 Contingencies calculated for thresholds Td and Th
Full size table
Table 6 Skill scores calculated for thresholds Td and Th (best scores are shown in bold)
Full size table

3.3 Proposed Landslide Early Warning System

WebGIS is a product that combines geographic information systems (GIS) and internet technologies (Miao and Yuan 2013). WebGIS is a tool for representing and analyzing geographical data over the internet environment. With the rapid development of Web and internet technologies, exchanging geographic information and transmitting urgent decisions through the WebGIS system is regarded as one of the fastest and most efficient methods (Zhang et al. 2011; Chen et al. 2016; Mamai et al. 2017). As a result, the WebGIS system may play an essential and effective role in the prevention and mitigation of natural catastrophe damage (Martinelli and Meletti 2008; Pessina and Meroni 2009).

In many areas, the landslide susceptibility map is mainly used for land-use planning and the development of measures to minimize the damage caused by landslides. Furthermore, when the susceptibility map is co-related with rainfall thresholds, the hazard map will be formed to figure out when landslides are likely to happen, and where people should be ready to take precautions or leave.

Within the framework of the SFLP project, a WebGIS system has been developed with free and open-source software. The WebGIS system has a three-tier structure, including: (1) the data tier uses the Database Management System PostgreSQL; (2) the server tier employs an HTTP web server—Apache and a map server—GeoServer; (3) the interface tier uses OpenLayers to visualize and manipulate maps and geographic objects on Web browsers. Some members of the author team have succeeded in developing tools to help the WebGIS system of SFLP automatically collect real-time rainfall data from two meteorological monitoring stations (installed in Ha Long city) and rainfall forecast data from the National Centers for Environmental Prediction (NCEP) system—USA. The structure and method of developing a WebGIS system integrated real-time rainfall data collection and forecasted rainfall data collection tools could be referred to in detail in Ha et al. (2021).

In order to effectively use the landslide susceptibility map and triggering rainfall threshold, these are required to be added to the WebGIS system along with automatic meteorological monitoring stations and rainfall forecasting systems. Rainfall thresholds need to be evaluated on a regular basis with real-time and anticipated rainfall data to identify the occurrence time of a high-risk landslide. The landslide susceptibility of Ha Long City’s predefined zones will be able to assist local authorities in focusing on warnings and organizing evacuations for high-risk areas, hence limiting landslide damage.

4 Discussion and Conclusion

The Spatial Multi-Criteria Evaluation (SMCE) technique is chosen with the benefit of being able to leverage expert knowledge to perform a landslide susceptibility map at a scale of 1:50,000 for the region of Quang Ninh Province. In this work, the input data for the SMCE model includes nine attribute map layers indicating nine component characteristics influencing the landslide phenomena in the region. The generated susceptibility map showed high consistency with the inventory map when more than 60% of landslide locations occurred in high and very high susceptibility areas. A significant amount of landslides occurred along the main roads and fell within the moderate susceptibility zone. This issue may come from two main reasons: the influence of human activities (house and road construction) and the limitations of input maps (terrain changes were not updated, and the spatial resolution is not high enough to represent small slopes). The influence of human activities should be considered if survey data can be collected in more detail.

The ratio of triggered and non-triggered over-threshold occurrences is an essential consideration when using the triggering rainfall threshold. If the non-triggered landslide rate is high, the warning’s efficacy will be reduced owing to people’s lack of faith upon getting wrong warning information (no geohazard occurrences). When compared to the use of threshold Th with threshold Td, it is clear that threshold Th is more reliable in assessing the risk of landslide occurrence. Despite the greater risk, using a threshold Td may assist in reducing the number of warnings for over-threshold rainfall events without triggering landslides.

Although empirical methods cannot pinpoint exactly each landslide site, they may assist local authorities and the public in anticipating high-risk landslide areas and periods throughout the rainy season. The use of empiric approaches is primarily reliant on experience, expert knowledge, and data from previous catastrophes. However, this strategy may avoid many problems associated with data heterogeneity. If information and attributes of landslide occurrences in Ha Long City are regularly updated, the accuracy of susceptibility zonation and rainfall thresholds for the initiation of landslides may be steadily improved over time.

In this research, the WebGIS system for landslide early warning was proposed. Real-time rainfall data (from a municipal meteorological monitoring station) and forecast rainfall data (from systems such as NCEP) may be linked and gathered into the system. The measurements are computed continually inside the system and compared to rainfall thresholds to anticipate the onset of landslides. Before and during severe rainfall events with the potential to surpass the threshold, landslide warnings can be issued through the WebGIS system.

It seems impossible to eliminate landslide occurrences in mountainous areas. Therefore, preparation and prevention are imperative to reduce the loss of life and property. Based on the empirical approach, this paper suggests the WebGIS system for early warning. However, implementing such a system is not easy, given the requirements of close collaboration among landslide specialists, local governments, and the general public in collaboratively designing countermeasure strategies, sharing information, and conducting prompt reaction plans.