The influence of land use and land cover change on landslide susceptibility in the Lower Mekong River Basin

The Lower Mekong River Basin in Southeast Asia experiences frequent rainfall-triggered landslides especially during the monsoon season. In this study, the influence of land use and land cover change and other causative factors on landslide susceptibility is evaluated in the Lower Mekong Basin. Frequency ratio analysis is performed to quantify the relationship between LULC change and susceptibility. Detailed landslide inventory maps are used for analysis with yearly LULC maps. The LULC change is used as a contributing variable in a logistic regression-based susceptibility model with other variables including distance to roads, slope, aspect, forest loss, and soil properties. The Receiver Operating Characteristic (ROC) curve and Area Under the Curve are estimated for the model trained by each landslide inventory. The models show good performance, with AUC values ranging from 0.697 to 0.958 and an average AUC equal to 0.820. Both the Frequency Ratio analysis and the Logistic Regression models indicate LULC change from agricultural land to forest has a positive correlation with landslide occurrence. The most significant factors in the models are found to be distance to roads, slope, and aspect. A better understanding of the effects of LULC on landslide susceptibility can be useful for local land and disaster management and for the implementation of LULC as a factor in future susceptibility models. Using datasets that are unique to the Lower Mekong region, this study provides additional insights into the relationship between causative factors and landslide activity to better inform regional and global landslide susceptibility modeling.


Introduction
A landslide encompasses a wide range of mass movements and can be defined as the downslope movement of soil, rock, or earth (Highland and Bobrowsky 2008). The movements can be triggered by various external activities such as intense rainfall, earthquakes, changes in water level, waves, or stream erosion, which lead to a decrease in shear strength and an increase in shear stress on the slope (Dai et al. 2002). Landslides can also be caused by anthropogenic activities such as excavation, road construction, and land use changes. These factors often induce small, shallow landslides, but abrupt changes to the slope surface such as poor construction and planning can result in larger, more dangerous landslides (Jaboyedoff et al. 2016). Watersheds that have been recently affected by wildfires can be highly susceptible to rainfall-triggered landslides that usually occur within a short time following the burn (Degraff et al. 2015;Kean et al. 2011). This study focuses on rainfalltriggered landslides as they are the most frequent and cause loss of life and destruction of property across the globe (Froude and Petley 2018). Rapid urbanization can increase the risk for landslides, especially along poorly constructed roads and deforested areas in mountainous regions (Forbes et al. 2012). However, better land management of forests and cultivated areas can produce a decrease in rainfall-triggered landslide susceptibility (Pisano et al. 2017).
In the Lower Mekong River Basin (LMRB) in Southeast Asia, the monsoon season brings an increase in flood and landslide disasters due to an increase in rainfall from large storms in combination with the complex topography of the region. The LMRB has experienced extensive changes from urban and agricultural expansion, deforestation, river damming, and natural disasters such as flood and drought. In this region, changes in Land Use and Land Cover (LULC) are largely influenced by agricultural prices, road accessibility, construction projects, and climate change (Spruce et al. 2020). Spruce et al. (2020) assessed LULC changes in the Lower Mekong using two maps from 1997 to 2010. In their analysis, 2.5% of the total area of permanent agriculture decreased which could be associated with the abandonment of crops or converting cropland to forests. They also identified a 6.7% increase of the total area categorized by scrub/shrub/herbaceous which could be attributed to abandoned cropland reverting back to forest. Looking at the changes between LULC classes, some cropland had changed to deciduous forest/scrub over time between 1997 and 2010. Changes in land cover can have variable impacts on landslide susceptibility. In some cases, human impacts such as deforestation and mining serve to exacerbate instability on slopes (Winter et al. 2010). In other examples, thoughtful engineering and planning can serve to stabilize slopes (Prastica et al. 2019;Yan et al. 2019).
The preconditions for landslides vary, but changes in land use and land cover (LULC) have been shown to have local impacts (e.g., Glade 2003;Hewawasam 2010;Mugagga et al. 2012;Reichenbach et al. 2014). Pisano et al. (2017) evaluated how land cover change affects slope stability over time in the Italian Southern Apennine Mountains by treating land cover as a dynamic variable, unlike many other landslides susceptibility studies that consider land cover as a static variable. This study found that a decrease in forest and cultivated land and increase in barren, pasture, and shrub land led to an increase in landslide susceptibility. A study by Persichillo et al. (2017) assessed shallow landslides in the northern Apennine Mountains in Italy in areas with land abandonment and changes in land management practices from 1954 to 2012. They found that cultivated lands that were abandoned and allowed to gradually recover naturally was the land cover change scenario most susceptible to landslides and land cover was the most predisposing factor in all study areas.
Similarly, Deng et al. (2018) investigated landslide distribution and agricultural abandonment in several provinces in China's mountainous areas. They concluded that more landslides occurred in areas with high incidence of agricultural abandonment. Furthermore, the effects of land use changes on landslides were analyzed in a landslide-prone region in Northeast Turkey with mountainous topography and high rainfall frequency by Karsli et al. (2009). Land cover changes and landslides were identified using aerial images taken in 1973 and 2002. Their results indicated that the land cover type played an important role in landslide occurrence as 95% of the landslides identified from the imagery were in areas with acidic soil weakened by fertilizer use in agriculture.
The effects of LULC on landslide susceptibility vary based on geographical location of the study area and methodology used for analysis. Remondo et al. (2003) applied various validation techniques for landslide susceptibility maps in the Deba Valley in northern Spain and state that while intense rainfall is the main triggering mechanism for shallow landslides, human activity affecting land cover changes are also influential as these alterations increase the sensitivity of the area to intense precipitation. Kafy et al. (2017) evaluated the association between LULC changes and landslides in the Chittagong Hill Tracts region, Bangladesh using a change detection technique to obtain the change over two decadal time periods from 1995 to 2005 and 2005 to 2015. They classified LULC as baresoil, hill forest, built-up area, and waterbody and found a positive relationship between landslide occurrence and decreasing hill forest. A study by Hewawasam (2010) investigated landslides in the Upper Mahaweli Catchment, Sri Lanka and stated that human activities were the primary cause of the increased landslide activity in the area, many of which occurred in agricultural areas on hillslopes. After the decline of the coffee industry in this region, many coffee lands that were deforested were restored as tea plantations, and out of 200 landslides the largest percentage, 35%, of landslides evaluated in their study occurred in LULC areas classified as tea plantation. Landslide occurrence in response to LULC change in New Zealand was assessed by Glade (2002), who found that forest clearance led to increased rates of landslides by evaluating sedimentation rates following land use modifications that triggered landslide activity. Lee and Sambath (2006) mapped landslide susceptibility in the Damrei Romel area in Cambodia and found higher landslide activity in areas with LULC classified as forest and shrubland than agriculture and grasslands due to landslides occurring in mountainous areas with steep slopes. Furthermore, Liu et al. (2021) examined changes in landslide susceptibility due to LULC changes in Lixian County, China and also found that landslide-prone areas were attributed to LULC classifications of rainfed cropland and rural land, and forested areas had low landslide susceptibility. The results of these studies show that LULC is an important factor influencing landslide activity, but the relationship varies between studies.
Frequency ratio (FR) analysis is a common method used to assess the relationship between susceptibility and the occurrence of a landslide event (Gariano et al. 2018;Pourghasemi et al. 2013). A study by Silalahi et al. (2019) used GIS mapping and FR analysis to assess the effects of contributing factors on landslides in Bogor, Indonesia. Their results indicated land cover as one of the most important factors contributing to landslides as well as lithology and soil type in this area. Additionally, Khan et al. (2019) used the FR to create a landslide susceptibility index which was used to produce a susceptibility map for northern Pakistan. Their study found barren land and irrigated agricultural land to have the highest FR values of the land cover classifications; however, distance to roads was found to have the highest overall FR value. These and other studies use FR to assess conditioning factors and create susceptibility maps but however, rarely incorporate land cover as a dynamic variable. Additionally, statistical methods like Logistic Regression 1 3 (LR) are effective in identifying input variable importance/significance and several studies have considered LULC within this framework (e.g., Reichenbach et al. 2018;Lee and Sambath 2006;Shahabi et al. 2014;Bai et al. 2010;Das and Lepcha 2019;Bornaetxea et al. 2018). LR has been an effective tool for developing landslide susceptibility maps and highlighting the significance of contributing variables; however, few studies have considered how LULC can be considered dynamically in these models to explain changes over time in a region. Hemasinghe et al. (2018) analyzed susceptibility in mountainous regions predisposed to landslides in Sri Lanka. Their study examined slope, aspect, lithology, land cover, distance to rivers and roads as predictor variables. Land cover was determined to be the most influential factor in the study area. Known landslide locations were used to validate their susceptibility map, and a majority (76%) of the landslide points were in high and extremely high susceptibility areas. Land cover is similarly used as a static predictor variable in many other susceptibility studies, uniquely this study will treat land cover change as a dynamic variable that changes over time.
The question posed in this study is-how do changes in land use and land cover (LULC) impact landslide susceptibility in the LMRB? We evaluate these interactions using several new landslide event inventories mapped between 2015 and 2018 that provide information within areas of Vietnam, Myanmar, and Laos (Amatya et al. 2021b) (Fig. 1). Additionally, this study closely examines the effects of LULC change on landslide occurrence, a dynamic which is not completely understood in the LMRB (Shu et al. 2019). Our work seeks to understand the relationship of changing LULC over time and how this impacts susceptibility. The relationship between LULC changes and landslide occurrence will be analyzed using Frequency Ratio (FR) analysis and Logistic Regression (LR) modeling. The FR will be used to closely examine the frequency of landslides within the various LULC change scenarios present in each of the landslide inventory locations. This study will use the LR models to compare LULC changes with other contributing factors like slope, forest loss, soil properties described in Table 1. The FR and LR results will be compared to determine any similarities regarding the significance of LULC changes on landslide occurrence. Results of this work are important as population expansion, road development and farming continue to increase (and hence changes in land cover) in the LMRB (Spruce et al. 2018). This work is part of a broader effort to characterize landslide susceptibility, hazard, and exposure within the LMRB for decision making at the country and municipal level using satellite remote sensing products.

Landslide inventory
The landslide inventories used in this research were mapped utilizing high-resolution satellite imagery from Planet (Planet Team, 2017) using the modified framework of Semi-Automatic Landslide Detection (SALaD) system (Amatya et al. 2021a). SALaD uses object-based image analysis and machine learning to map landslides. As we are focusing on rainfall event-based inventories, it is important we map landslides induced by that event only as pre-event landslides may have been triggered by phenomena such as earthquakes. A change detection-based approach was introduced to the SALaD framework (SALaD-CD) utilizing pre-and post-event imagery (Amatya et al. 2021b). The new framework incorporates image normalization, image co-registration and change detection. The landslide polygons produced by SALaD-CD were manually corrected and converted to initiation points using NASADEM (NASA JPL 2020). A total of 18 rainfall-triggered landslide inventories between 2015 and 2018 were used in this analysis: 13 in Vietnam, three in Myanmar, and two in Laos (Fig. 1). Details on the landslide inventories used in this study are provided in Table 2 including location information, year of event, satellite image source, number of landslides, total area of the inventory extent, and total area of the landslides within the inventory extent.

Digital elevation model (DEM)
NASADEM with 30 m spatial resolution was used to derive the input variables for slope and aspect (NASA JPL 2020). NASADEM is derived from the Shuttle Radar Topography Mission (SRTM), which was launched in 2000 (Van Zyl 2001) with processing Forest cover loss 30 m Hansen et al. (2013) improvements, elevation control, void-filling and merging with data that was unavailable at the time of the mission. NASADEM also provides an improved spatial resolution from the original three-arcsecond SRTM DEM to one-arcsecond, making it the finest resolution, global, freely-available DEM product. Slope is one of the major influencers of landslide activity, and therefore was selected based on its likely correlation with landslide susceptibility (Bruschi et al. 2013;Chen et al. 2019;Indhanu et al. 2020). The aspect is derived using GIS and reclassified to represent North, South, East, and West facing slopes. Data are publicly available from https:// doi. org/ 10. 5067/ MEaSU REs/ NASAD EM/ NASAD EM_ HGT. 001.

Land use/ land cover (LULC)
Land use and land cover (LULC) data from the Regional Land Cover Monitoring System (RLCMS) is available via the SERVIR-Mekong Land Use Portal and is presented at 30 m spatial resolution and yearly temporal resolution from 1987 to 2018 for the Lower Mekong River Basin (Saah et al. 2020). The RLCMS uses historical Landsat and MODIS data to create LULC maps. The methodology behind deriving the maps can be summarized, defining the classification typologies, creating the primitive layers using supervised classification and machine learning algorithms, combining the primitive layers into land cover maps, and lastly, an accuracy assessment. The typology classifications were determined by stakeholders from Cambodia, Laos PDR, Myanmar, Thailand, and Vietnam. LULC data from ten years prior to each landslide event, spanning from 2005 to 2018, were used for analysis in this study. Due to limited spatial representation of the original classifications, the 1 3 land cover was reclassified into three categories shown in Table 3, similarly to the methodology in Chen et al. (2019). For more information on how this dataset was created and its details, please review Saah et al. (2020). Data are publicly available from https:// landc overm apping. org.

Roads
Proximity to roads has been found influential on rainfall-triggered landslide occurrence in several areas in previous studies (Larsen and Parks 1997;Penna et al. 2014;McAdoo et al. 2018). The construction of roads changes the hydrologic response and surface and subsurface flow paths in the affected area which can influence landslide susceptibility (Penna et al. 2014). McAdoo et al. (2018) found that within 100 m of a road rainfall-triggered landslides were more than two times as likely to occur in Nepal. The Global Roads Inventory Project (GRIP) is compiled from publicly available national vector datasets from governments, research institutes, NGOs and crowd-sourcing initiatives and includes over 21 million km of roads. The GRIP dataset is available as a vector dataset for each region of the world. The Southeast Asia GRIP dataset was used to derive a distance to roads raster which represents the distance to nearby roads for each pixel in the selected extent using GIS software for each landslide inventory location. For more information on how this dataset was created and its details, please refer to Meijer et al. (2018). Data are publicly available from https:// www. globio. info/ downl oad-grip-datas et.

Soil properties
The World Soil Information Service (WoSIS) provides standardized soil profile data for various environmental applications at global scale. The most current WoSIS snapshot is a compilation of nearly 200,000 soil profiles from locations across the globe. These profiles are standardized and distributed using a database model, SoilGrids, which uses machine learning methods to map the spatial distribution of soil properties based on the soil profile observations (Hengl et al. 2017). Soil data are available for six standard depth intervals at 250 m spatial resolution. For a full description of this data, please refer to Batjes et al. (2020). Two soil layers representing bulk density and organic carbon were downloaded from SoilGrids for each study region at the at 0-5 cm depth. These variables were

Forest cover loss
Global Forest Change data are created from Landsat imagery at 30 m spatial resolution and characterizes forest extent, loss, and gain by Hansen et al and is available from 2000 to 2021. Forest cover loss is defined as a change from a forest no non-forest state. In this study, forest cover loss is estimated for the extent of each landslide inventory over 10 years prior to the event, spanning from 2005 to 2018 depending on the year the event occurred and is represented as a binary raster of zeros (no loss) and ones (loss). For example, Forest Cover Loss is estimated over the time period 2007-2017 for the landslide events occurring in 2017. For more information on the methodology and details of this dataset, please review Hansen et al. (2013). Data are publicly available from https:// glad. earth engine. app/ view/ global-forest-change.

LULC analysis
The landcover maps were reclassified from nine categories to three broader categories due to the limited spatial representation of the original land cover classifications in the small extents represented by the landslide inventory locations. These reclassification categories as well as the original categories represented in the study area described in Table 3 and shown in Fig. 2 for the greater Lower Mekong region. The reclassified landcover maps were analyzed to estimate the amount and type of change over the 10-year time period prior to the landslide event in each study region as well as the overall LULC patterns in the greater Lower Mekong region from 1998 to 2018. Overall, the time frame used to analyze LULC changes varies in the literature. Several studies chose time intervals that ranged from five to 30 years due to the availability of aerial imagery (Karsli et al. 2009;Persichillo et al. 2017). The 10-year time period was selected based on previous studies that assessed LULC changes over time (Chen and Huang 2013). Chen and Huang (2013) compared LULC changes over 10 years (1999 to 2009) to determine the relationship between LULC change and landslides triggered by Typhoon Morakot in Taiwan. They found that areas with a change in land cover had a higher frequency of landslides than non-changed areas.
Here, the LULC change over the ten-year time period was estimated by assigning a change scenario to each pixel based on its land cover classification in the two maps (i.e., map of landcover in 2018 compared to the map of landcover in 2008). There is a total of nine possible LULC scenarios for each pixel including three no-change scenarios where the type of land cover classification did not change over the specified time period. The percentage of total area of each LULC scenario is estimated and reported for each landslide inventory location and shown in Fig. 3a and averaged over all locations in Fig. 3b. Land cover classified as urban represents less than 1% in each study area except for one inventory, Nha Trang (NH), which is composed of 16% urban area. Only three of the inventories are characterized by greater than 25% agricultural area (Fig. 3a). Forests comprise over 50% of each area and 81% of the total area used for analysis. The LULC change scenario with the 1 3 largest area represented in the study locations is from agriculture to forest, representing 2% of the total area and less than 5% in each inventory extent (Fig. 3b).

Frequency ratio analysis
Frequency ratio (FR) was used in this study to quantitatively examine the relationship between landslide occurrence and each LULC change scenario. The FR is the ratio of the percentage of landslide occurrence in a factor class to the total percentage of that factor in the area, and the average FR value is one (Khan et al. 2019). Values greater than one indicate that landslides occur more frequently relative to the total distribution within the variable being considered where M i = number of pixels containing landslides in LULC class i, ∑ M i = total number of pixels containing landslides in the study area, N i = total number of pixels in the study area for that particular LULC class, and = ∑ N i total number of pixels in the study area. The FR of each LULC scenario is estimated for each study location.

Logistic regression
A logistic regression model is used in this study to further assess the relationship between changes in LULC and landslide occurrence as well as determine any similarities with the results of the frequency ratio analysis. The model is outlined in the workflow diagram shown in Fig. 4. Many susceptibility studies use logistic regression to evaluate binary response variables such as landslide occurrence (Lombardo and Mai 2018;Horafas and Gkeki 2017;Pourghasemi et al. 2013). Logistic regression models the probability of events based on the linear combination of independent contributing variables (Camilo et al. 2017). The model produces coefficients for each input variable, which are used to predict the probability of landslides over the study area. Positive coefficients indicate that there is positive correlation with the presence of this conditioning factor and landslide occurrence, and negative coefficients indicate a negative correlation between presence of the factor and landslide occurrence and there is an absence of this factor in locations where landslides occur. The predictions are used with raster data of each contributing variable to create a landslide susceptibility map. The equations used in logistic regression are as follows: where P = probability of occurrence of the event occurring, Z = linear combination, b 0 = intercept, b i = slope coefficients, and x i = independent variables (Lee and Sambath 2006). To create training and testing datasets, an equal number of non-landslide points are generated. The logistic regression model is used to create landslide susceptibility maps for each of the 18 study regions mapped in the landslide event inventory. The model is trained with 70% of the landslide inventory as well as randomly generated points that do not coincide with landslide locations. The following predictor variables are tested in the logistic regression model in this study: slope, aspect, LULC change, distance to roads, bulk density, organic carbon, and forest loss. These variables were selected based on their spatial and temporal availability and potential correlation with landslide occurrence as well as their significance in the logistic regression models determined by their corresponding p-values. A significance level of 0.01 implies there is less than 0.1% chance that the coefficient may be equal to zero and therefore be insignificant in determining landslide susceptibility. A p-value less than 0.05 is statistically significant and p-values greater than 0.05 are determined as not statistically significant. All of the model input variables were reclassified into categorical factors for simplified comparison between the categories. The classifications of each factor are shown in Table 4. These maps were validated using the receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) analyses, common statistics for assessing the predictive capacity of susceptibility models (Felicísimo et al. 2013). The ROC was derived using a testing dataset that is comprised of 30% of the landslide inventory as well as randomly generated non-landslide points. These metrics are often considered to evaluate the performance of susceptibility models and the model with a larger AUC value is considered better predictive model (Gorsevski et al. 2006;Zhou et al. 2018;Lombardo and Mai 2018).

Results
Eighteen landslide inventories were used in this study to assess the relationship between land cover change and landslide occurrence. The percentage of total area represented by each LULC scenario averaged over all inventory locations is shown in Table 5. The number of total landslides occurring in each LULC scenario from the 18 inventories are shown in Table 6. The frequency ratio value averaged over all inventories for each LULC scenario 1 3 is shown in Table 7. The category forest with no change in LULC over ten years prior to the landslide event made up about 82% of the total area when averaging over all 18 inventory locations (Table 5). The majority of landslides (15,568) also occurred within this LULC category (Table 6). The average FR for forest with no change was 2.06 (Table 7). The category making up the second largest percentage of total area is agricultural land with no change over ten years prior to the landslide event, which composed about 15% of the total area (Table 5). This LULC category experienced 2530 landslides (Table 6) and had an average FR of 1.64 when averaged over all locations (Table 7). The category of urban with no change represented about 1% of the total area within the 18 inventory locations (Table 5) and experienced no landslides. The LULC change scenario from agriculture to forest in the ten years prior to the landslide event comprised only 2.1% of the total area within all inventory locations (Table 5). However, this area experienced 614 landslides (Table 6) and had an average FR of 2.83 (Table 7) for all locations. The results of the FR analysis could indicate that areas experiencing a change from agriculture to forest are more susceptible to landslides than other land cover change scenarios.
The frequency ratio (FR) aids in understanding location specific conditions for landslide occurrence by providing the ratio of the landslide area to the total area. Figure 5 shows a matrix diagram of the FR for each landslide inventory and LULC change scenario. FR values less than one are represented by dark blue. From Fig. 5, we can see that all categories including urban produced FR values less than one. The scenario from forest to agriculture was only found significant in the Sin Ho, Vietnam with a FR of 1.63 and Hakha, Myanmar with a FR of 1.94. The scenario agriculture with no change in ten   Spruce et al. (2020) observed similar changes from agriculture to forest in the Lower Mekong between the period from 1997 to 2010. These high FR values could be explained by agricultural abandonment practices having an impact on landslide occurrence which would be consistent with the results found by Deng et al. (2018) and Persichillo et al. (2017). In addition to assessing the LULC within the extents of the 18 landslide event inventories, land cover changes were also analyzed over the greater Lower Mekong region from 1998 to 2018. Figure 6 shows the area in square km represented by each LULC category over the entire Lower Mekong region and a noticeable change in land use between 1998 and 2001. The urban areas increased by 11.3% from 1998 to 2001 and increased constantly at an average rate of 1.2% per year. The area represented by forests decreased by 5.5% of the total area from 1998 to 2001 with an average decrease of 0.3% per year. In contrast, the amount of agricultural area increased drastically by 11.6% from 1998 to 2001 with an average increase of 0.6% per year. The percentage of total area represented  Figure 7 shows that the majority of the Lower Mekong region is forested ranging from 67 to 63% of the total area for the time period from 1998 to 2018. About a third of the area is classified as agricultural land, ranging from 30 to 34%, and the smallest percentage of total area of the region is classified as urban, ranging from 0.9 to 1.1%. Figure 8 shows the percentage of total area represented by each LULC change scenario over the greater Lower Mekong region between 1998 and 2018. The LULC can be summarized as 62.1% forest unchanged, 29.1% agricultural area unchanged, 5.8% changed from agriculture to forest, 1.7% changed from forest to agriculture, 0.9% urban unchanged, and the combined land cover change scenarios of urban to forest, urban to agriculture, and agriculture to urban constitute merely 0.3% of the total area. A logistic regression model was trained and validated for the extents of 18 landslide event inventories. Table 8 presents a summary of the logistic regression coefficients and p-values of all models for the landslide inventory locations with greater than 1000 landslides to avoid potential overfitting from the smaller observation counts. The logistic regression model results for each of the 18 inventories are provided individually in Online Resource 1. In Table 8, the inventory abbreviations are shown next to the corresponding values for the minimum and maximum coefficient and p-value for each category within the factors. From this table, Muong Lat (MT) represents the highest coefficients in the land cover change corresponding to forest (no change), agriculture to forest, and agriculture (no change). When examining the LULC over the inventory location, the landslides largely occur in a forested area with no variation in land cover. The land cover in the Muong Lat (MT) area is fairly homogenous with minimal area represented by the land cover change scenarios other than forest (no change). However, several landslides did occur within areas that underwent deforestation for agriculture and in invariable agricultural areas. The model trained with the Da Bac (DB) inventory produces minimum coefficients for all distance to roads categories which indicates that in this area, the presence of roads is not very influential on landslide occurrence. When examining the Da Bac (DB) location, there is an ample road network throughout the province, but the landslide distribution does not appear to be influenced by the presence or absence of roads and the landslides do not cluster around roads.
Furthermore, Mu Chang Chai (MC) contributes to the minimum coefficients for the slope categories which could indicate that slope is less influential in this location, even though the p-values corresponding to the slope factor were not significant. The inventory mapped in Mu Chang Chai (MC) occurred in a mountainous forested area with varying slope. The inventory in Mu Chang Chai (MC) produced the highest coefficient for forest loss and the minimum p-value which could be explained by a large cluster of landslides occurring around an area that had experienced substantial forest loss in the years prior to the event. Mu Chang Chai (MC) also represents the maximum coefficients for proximity to roads indicating that roads are more influential on landslides in this location compared to other inventories. When looking closely at the area, the landslide distribution appears to be influenced by proximity to the road network. However, the p-values corresponding to distance to roads in the Mu Chang Chai (MC) inventory indicate this factor was not found to be statistically significant in the logistic regression model. Overall, the coefficients for distance to roads decrease with distance indicating that road proximity has an influence on landslide occurrence. This would agree with the findings of Larsen and Parks (1997) which identified that landslide frequency decreased as distance to roads increased in Puerto Rico.
The logistic regression models were validated using 30% of the landslide inventory available in each of the 18 study locations, which consists of one-third of the number of landslides in each inventory shown in Table 2, as well as an equal number of randomly generated non-landslide points. The number of points used in the validation varies for each model according to the number of available landslide points in that event inventory, and the points selected were chosen at random. The models were validated by plotting the true positive rate versus the false positive rate (ROC) and calculating the area under the resulting curve (AUC). The AUC values from the validation of the logistic regression model, the number of landslides within each landslide inventory, and the number of landslides used for validation are presented in Table 9. The model validation results found all AUC values to be estimated above 0.7 except for one inventory in Hakha, Myanmar with an AUC value equal to 0.697. This was the lowest AUC among the model results indicating the model  Table 9, the number of landslides and AUC values are not correlated indicating the number of landslides within each location used to train the models did not affect the model performance.

Discussion
The land cover of the inventories used in this study was dominated by forests with minimal LULC changes present. According to Table 5, the scenarios of LULC change from forest to urban, agriculture to urban, urban to forest, and urban to agriculture combined make up only 0.082% of the total area averaged over the 18 inventory locations. Looking at the LULC analysis over the LMRB in Figs. 6 and 7, the relatively small area occupied by urban areas may make it difficult to see any conclusive patterns in landslide activity. The LULC scenarios involving urban activities were less prominent in the landslide inventories used in this study, but also occupy the lowest amount of total land area in the greater Lower Mekong region so this category is not only less represented in the 18 landslide inventory area locations, but overall in the Lower Mekong River Basin. The amount of area occupied by urban and built-up classification is not comparable to the amount of agricultural and forested areas in the LMRB. Additionally, the scenario involving urban activities did not experience any landslides within the 18 inventories used in this study. Therefore, the impact of LULC change scenarios to or from the urban classification cannot be fully determined using the 18 inventories presented in this study. *Colors correspond to landslide inventories highlighted in Table 8 1 3 The frequency ratio (FR) reveals potential correlation between landslide locations and causative factors in the area. Overall, the scenario from agriculture to forest had the highest average FR value followed by the scenario of forest with no change (Fig. 5). The forested areas that experienced no change had an average FR value of 2.06 when averaged over all 18 study locations, and this is most likely due to the prevalence of forested mountainous regions with steep slopes where most landslides occur in this region. The results of this study indicate that the land cover change scenario from agriculture to forest could have an impact on landslide occurrence and that areas changing from agriculture to forest may be more susceptible to landslides in the years following the land cover change. This could indicate that abandoned agricultural lands left to naturally recover are more susceptible to landslide activity than other land cover change scenarios, which would be similar to the conclusions drawn by Deng et al. (2018) and Persichillo et al. (2017). Further evaluation of land cover change from agriculture to forest with landslide inventories where this LULC change scenario is more prevalent than the inventories used here needs to be executed to further understand the impact of land use changing from agriculture to forest as invariable forests dominate the areas used in this study. Additionally, the FR being a univariate analysis must be taken into consideration. The areas where landslides occurred that were deforested for agricultural purposes could be influenced by additional factors. Other limitations include the spatial resolution and the availability of remotely-sensed datasets for the region. For example, the finest resolution for publicly available DEM data for the Lower Mekong River Basin is 30 meters, which is much coarser than other regions such as USA with one-meter DEM available. With landslides ranging in size, coarser spatial resolutions may not accurately represent the environment of very small, shallow slope failures. Enhanced spatial resolution of remotely-sensed data would greatly benefit modeling efforts for landslide susceptibility in the Lower Mekong River Basin.
We can make generalized conclusions for the 18 inventory locations and models regarding landslide causative factors using the coefficients and p-values from Table 8. When averaged over all 18 models the causative factors associated with distance to roads, forest loss, and slope display positive coefficients, which is expected considering these variables are known to be influential on landslide occurrence. Overall, the most significant variables based on the mean p-values were slope and aspect. The high significance of aspect could be explained by the models using flat as the basis factor. When considering all 18 inventories, the logistic regression models did not indicate that areas changing from agriculture to forest were as influential on landslide occurrence as the FR analysis did, which could be explained by the FR being a univariate analysis and the logistic regression being multivariate. However, the average coefficient for the LULC change category of agriculture to forest has the only positive coefficient of any of the change scenarios besides forest and agriculture with no change. Forest with no LULC change had the highest and only positive average coefficient relative to the other land cover change scenarios which is most likely due to the majority of landslides in this study occurring in steep forested areas and the considerable amount of forested area throughout the study locations. Areas that changed from forest to urban have the lowest average coefficient which could be explained by the limited number of landslide points within the area represented by this scenario in the inventories. Alternatively, this could mean that areas being developed are not susceptible to landslides due to other factors such as low slope which could indicate informed land management decisions in the region.
The Lower Mekong River Basin is an area where people are dependent on the Mekong and its branches for livelihood. There have been numerous studies on the hydrology (Mohammed et al. 2018b; Le et al. 2020Le et al. , 2022Mondal et al. 2022), use of satellite data (Fayne et al. 2017;Le et al. 2018;Mohammed et al. 2018c;Dandridge et al. 2019Dandridge et al. , 2020 to characterize the basin and hence provide better management strategies (Mohammed et al. 2018a) for the water resources. Landslides play an important role in the watershed and can cause partial blockages of the flow if they occur close to a stream channel.

Conclusions
Land use and land cover (LULC) changes can affect slope stability and geological conditions that may influence the occurrence of landslide activity. This study assessed the influence of LULC changes on landslide susceptibility for 18 locations throughout the Lower Mekong River Basin (LMRB). The majority of landslides occurred in the LULC change scenario with the largest land coverage percentage, which is forested areas that did not experience any change in land cover. However, the LULC change scenario from agriculture to forest had the highest FR values overall. Both the FR analysis and LR models indicated that the LULC change scenario from agricultural land to forest could positively correlate with landslide occurrence. However, for real-time analysis the LULC data available on the SERVIR-Mekong Land Cover Portal would be needed to be updated each year as currently, maps are only available only up to 2018.
There were data limitation issues regarding the representation of land cover changes in the extent of the landslide inventories. Not all land cover change scenarios are represented in these study areas, more locations with landslide inventories that experienced land cover changes are needed to fully analyze the influence of LULC on landslide susceptibility. This method can further be applied to new landslide inventories for other locations in the Lower Mekong River Basin as the data become available to discover more consistent correlations between LULC change and landslide susceptibility. Specifically, the relationship between agricultural abandonment and landslide occurrence could be further analyzed to determine how this LULC change influences susceptibility in the Lower Mekong Region. The inclusion of dynamic LULC in landslide susceptibility models could greatly improve hazard assessment and should be investigated as the combination of land use changes over time due to population expansion and disturbances caused by climate change. This analysis is done primarily using remote sensing products, making it transferable to other landslide-prone regions around the world. Additional research efforts could further investigate the role of agricultural abandonment and natural recovery on rainfall-triggered shallow landslides.