Comparative Analysis of Local and Transferred ANN Models in Landslide Susceptibility Prediction in a Tropical Region

2.1 Study area

Sabah and Sarawak exist on the same island of Borneo [20, 21], whilst Selangor is situated in Peninsular Malaysia. Target 1 is located approximately 824.97 km away from Western Sarawak, and Target 2 is located 1009.89 km away from the region. Although seemingly far apart from each other, each one of the study areas exhibits the same dominant type of landslide, which is shallow rapid landslides. This is due to the shared primary landslide triggering factor which is heavy rainfall [22]. The three regions share the same wet season, which is the Northeast monsoon season that generally occur from November to February annually [23]. According to the rainfall intensity data provided by the Climate Hazards Center InfraRed Precipitation with Station Data (CHIRPS) during the Northeast monsoon season of 2022-2023 in mm per five days (pentad), the transfer area receives the most rainfall, with a minimum of 65.753 mm/pentad, a mean of 85.167 mm/pentad, and a maximum of 121.266 mm/pentad followed by Target 1 and Target 2, as seen in Table 1 [24].

Table 1. Rainfall data statistic in each study area in mm/pentad

Geologically, the conditions in each study area varied vastly from one to the other according to the Harmonised World Soil Database version 2.0 [25]. The only constant geological properties that was present throughout the study areas was Technosols with Geocode 7001 (Table 2). Between the Transfer region and Target 1, the other similarities were 4461, 4463, and 4498 (Table 2, Table 3). This is likely since both areas are located within the same island of Borneo, where some geological properties can be similar [26]. Furthermore, there were no similarities between the Transfer region and Target 1, to Target 2, excluding 7001. Unlike Target 1, Target 2 is in Peninsular Malaysia, which explains the difference between the geological properties of the region to the study areas in Borneo.

Economically, the gross domestic product (GDP) of Target 2 is 145,659 USD, which greatly outweighs the entire state of Transfer region, and Target 1 with a GDP of 31,636 USD, and 18,489 USD respectively [20, 27]. Furthermore, in terms of population density, the density of Transfer region and Target 1 combined could not surpass the population density of two cities in Target 2 that exceeds 10,000 people per square kilometre (km²), in contrast to Transfer region and Target 1 with a population density of 5,393 people per km² and 1,703 people per km² respectively [28]. As anthropogenic variables have been determined to influence landslides, it was important to include it in this study [29].

Table 2. Geological property type in each study area

Table 3. Geocode dominant composition

Elevation and slope angle are two topographical variables that frequently coincide with landslides, and have been deployed in multiple landslide susceptibility studies through ML models [30, 31]. With the data provided by the National Space Agency (NASA), Shuttle Radar Topography Mission (SRTM), the elevation and slope angle of the study areas were evaluated [32]. In terms of elevation, Target 1 was determined to have a maximum elevation of 3973m, a mean of 530.988m, and a minimum of -12m. Target 2 has a maximum elevation of 1262m, a mean of 112.277m, and minimum of -35m. The transfer region has a maximum elevation of 1262m, a mean of 86.010m, and a minimum of -72m. Thus, the Transfer region is more similar in terms of elevation to Target 2, instead of Target 1. The minimum slope angle in each study area were constant with 0°. In Target 1, the maximum slope angle was 85.132°, with a mean of 17.282°. In Target 2, the maximum slope angle was 74.688°, with a mean of 7.429°. In the Transfer region, the maximum slope angle was 75.650°, with a mean of 8.089°. In terms of slope angle, the Transfer region has more similarities to Target 2, in contrast to Target 1. Thus, in can be concluded that topographically, the Transfer region is shares more in similarity to Target 2 in comparison to Target 1. The location of the study areas is depicted in Figure 1.

1.png

Figure 1. Location of study area

2.2 Methodology

This study was conducted in four primary phases (Figure 2).

2.png

Figure 2. Study flowchart

The first phase was conducted to collect the developmental data required by the ANN models. The second phase was to check for multicollinearity amongst the variables in the primary dataset. The third phase was the ANN model developmental and evaluation phase where hyperparameters optimisation (HPO) and metrics evaluation was conducted. The final phase was to plot the LSMs.

2.2.1 Data collections and primary data preparation

To develop the ANN models, both landslides inventory and landslides factors were required. The data was detrimental for the ANN model as it is a supervised approach to predict landslide susceptibility.

Landslide inventory consists of landslides and non-landslides points. These variables are important at it helps the ANN model to make a distinction between the high risk and low risk areas in conjunction with the landslide factors [33]. For the target areas, the landslide points data were provided by NASA through the Global Landslide Catalogue for landslide triggered by rainfall [34]. As for the Transfer region, the landslide points data were provided by the local authorities. The non-landslide points on the other hand were obtained by randomly sampling points with slope angle less than or equal to 5° [7]. For the Transfer region, there were 53 landslides points, and 977 non-landslides points. For Target 1, there were 34 landslide points, and 199 non-landslide points. For Target 2, there were 41 landslide points, and 154 non-landslide points.

Landslide factors on the other hand are variables that causes landslides to occur which can be divided into triggering factors and conditioning factors. It was crucial to ensure that all of the variables selected are present in all study areas to enable the ANN model transfer, thus geological variables were not included [35]. The triggering variable in this study was rainfall, where the data was provided by CHIRPS for the average rainfall in mm pentad during the wet season. Rainfall is a significant triggering variable of landslides especially in the tropics where it is abundant during the wet season. It is responsible for the increase in pore water pressure that can greatly destabilise a slope forming material stability [36].

For the conditioning factors, it was represented through aspect, curvature, distance from roads (DFR), elevation, Normalised Difference Vegetation Index (NDVI), and slope angle.

The aspect data for the study areas were provided by NASA SRTM’s processed from the digital elevation model (DEM) in Google’s Earth Engine (GEE) [32]. Aspect refers to orientation of a slope face ranging from 0° North circling clockwise back to its origin. Different aspect has different levels of landslide susceptibility depending on the area of interest. This is due to the localised climate conditions of each aspect, that causes the weathering rate for the slope forming materials to be different from one to the other [37]. The difference in weathering conditions will ultimately lead to differing landslide susceptibility rate.

The curvature data was obtained by conducting a terrain analysis on the DEM through the System for Automated Geoscientific Terrain Analyses extension in Quantum Geographical Information System (QGIS) [38]. Curvature ranges from concave to flat to convex as it goes from the negative values to 0 to positive values [39]. As curvature goes from the positive to negative, the susceptibility towards landslide increases due to the tendency of ponding on concave curvatures, which allows rainfall to better penetrate the surface soil [30].

DFR in this study was a measure of anthropogenic activity as not only does it correlate with the disturbance of the original slope to make way for the construction, but human settlement has tendency to be nearby a road for easier access [40]. The expansion of human settlements increases the area of impervious surface, which leads to an increase in runoff that could trigger landslide [41]. Thus, being near a road generally increases the likelihood of its surroundings to be more susceptible to landslides. The DFR data was obtained by querying road data using the Quick Open Stream Map service in QGIS, before converting it into a proximity map using QGIS’s proximity buffer function [42].

The elevation data was supplied by NASA SRTM through GEE. As elevation increases the likelihood of landslides also increases due to the influence it has on the other landslide conditioning factors [43]. Furthermore, shallow landslides have been observed to occur more in areas with variations in the elevation [44].

NDVI is a measure of vegetation density, where areas with high NDVI tends to have lower susceptibility towards landslides due the support of roots which reinforce the slope forming materials [45]. Furthermore, areas with a low NDVI values are known to be more susceptible to landslides as not only does it lacks the root support, it also increases the rate of rainfall infiltration due to the lack of surface cover, which leads to saturated soils that are more likely to experience slope failure [46]. The NDVI data of this study was supplied by the United States Geological Survey Land Satellite [47].

Slope angle is an important variable to consider in a landslide susceptibility study. As slope angle increases, the stability of the slope forming material decreases making it more prone to landslide in comparison to areas with lower slope angle [48]. Furthermore, slope angle influences landslide motion characteristics with an increase in slope angle leading to a greater runout [49]. The slope angle data in this study was obtained from NASA SRTM through GEE.

After the landslide inventories, and landslide factors have been collected, the primary data was prepared. It was done by sampling the values of each landslide factors at each landslide and non-landslide points using the raster sampling tool in QGIS [50, 51].

2.2.2 Multicollinearity assessment and data preparation

Multicollinearity is not a favourable condition when it comes to developing an ML model. Input variables with multicollinearity amongst each other exceeding 0.8 is highly correlated [52]. When two input variables are highly correlated, it introduces redundancy issues to the ANN model, as it considers two input variables as a single variable, thus increasing the variable influence on the output unjustifiably [53]. In this study, the multicollinearity was accessed on the primary data through Microsoft’s Excel data analysis plugin for correlation, where if a correlation amongst two input variables exceeds 0.8, one of it were to be removed from the dataset [54]. After no multicollinearity issue has been determined, the primary data undergoes data preprocessing. In this study, the landslide factors were processed through min-max scaling (Eq. (1)). Min-max scaling ensures the landslide factors to be scaled from 0 to 1, to avoid issues where a single variable would overpower the other variables that have smaller values, such as the values of elevation in comparison to slope angle [55, 56].

$\hat{x}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}$          (1)

where, $\hat{x}$ is the scaled x value, $x_{\text {min }}$ is the minimum value of $x$, $x_{\max }$ is the maximum value of $x$, and $x$ is the current value of $x$.

2.2.3 ANN model development and evaluation

The ANN models were developed in R through the “NeuralNet” package, that allows hyperparameter optimisation (HPO) such as architecture, maximum steps, and learning algorithm [57]. The training data was used to develop the ANN model whereas the testing was used to evaluate its predictive performance based on never-before-seen data. The evaluation metrics for the ANN models were Area Under the Receiver Operating Curve (AUC) and recall. AUC provides information on how well the ANN model distinguishes landslide cases and non-landslide cases based on the landslide factors [58]. It is obtained by evaluating the area under the curve for true positive rate (TPR) versus the false positive rate (FPR). Recall on the other hand is TPR, in a ML development where the 1s (landslide) data pool is at an imbalance to the 0s (non-landslide) as it is smaller, it is more important for the ML model to be able to predict more of 1s accordingly (Eq. (2)) [59]. Both AUC and Recall minimum score were set at 0.8, where if the scores were lower, HPO would be conducted. The importance of each landslide factors on landslide occurrence in the ANN models were determined through Garson’s algorithm from the “NeuralNetTools” package in R.

Recall $=\frac{T P}{T P+F N}$           (2)

where, TP is true positive, and FN is false negative.

2.2.4 LSM plotting

To plot the LSM for each study area, the landslide factors data was converted into a numerical data frame using the “terra” package in R [60]. Then, the numerical data frame underwent the same min-max scaling process as the training data and the testing data. The ANN model was then deployed to plot the LSM where each target areas landslide susceptibility was predicted by its local ANN model and the Transfer region ANN model. The LSMs were important as it visualised the sensitivity of the ANN models to the variable importance [5].

This study aimed to conduct a comparative analysis of local and transferred ANN models in landslide susceptibility prediction in a tropical region. There were similarities and differences between the transfer region and Target 1 and Target 2. In all three regions, the slope angle was determined to have the most influence on landslide occurrence. The geological condition was the primary difference for all of the study areas.

For Target 2 statistics, the local ANN model achieved a mean of 0.068 and a median of 0.007 for landslide susceptibility. When deploying the Transfer ANN model on Target 2, the mean was 0.136 whereas the median was 0.0001. Both ANN models showed that landslide predictions were skewed to the right; however, a stronger skew was observed again in the Transfer ANN model predictions. This can be attributed to differences in variable importance between the target ANN models and the Transfer ANN model.

In Target 1, the Transfer ANN model achieved a recall of 0.89 and an AUC score of 0.988, in comparison to the local ANN model, which achieved a recall and AUC score of 1, respectively. In Target 2, the Transfer ANN model has achieved a recall of 0.86 and an AUC score of 0.995, in comparison to the local ANN model with a recall and AUC score of 1 respectively. The difference was explained through differences in variable importance for each region, as determined by Garson’s algorithm. For the LSMs, the Transfer ANN model predictions in each target areas were heavier skewed to the right in comparison to the local ANN models predictions. This study showed that the transfer was successful, as indicated by high performance metrics on the test data. The transferred ANN model indicated that in the tropical region, a region with scarce landslide related data can be predicted by a transferred ML model developed with a more comprehensive dataset. However, its applicability is limited to a preliminary assessment of landslide susceptibility, as the generalization of variable importance differs. To further improve performance, it is suggested that more data be obtained from local agencies or to combine datasets, as observed in studies on the Himalayas. It is recommended that future research explore the transferability of ML models to different tropical regions outside of Southeast Asia.