Rainfall‐induced landslide inventories for Lower Mekong based on Planet imagery and a semi‐automatic mapping method

Abstract

Fatal landslides occur every year during the rainy season (June–November) in the Lower Mekong Region (LMR). There is an urgent need to develop a landslide early warning system in the LMR. In collaboration with the Asian Disasters Preparedness Center and NASA’s SERVIR Programme, we are regionalizing the global Landslide Hazard Assessment System for Situational Awareness model for the LMR (LHASA‐Mekong). A robust set of landslide inventories are needed to effectively train the machine learning‐based LHASA‐Mekong model. In this study, the Semi‐Automatic Landslide Detection (SALaD) system was modified by incorporating a change detection module (SALaD‐CD) to produce rainfall event‐based landslide inventories using pre‐ and post‐imagery from RapidEye and PlanetScope for various locations in the LMR that were identified based on media and government reports. These rainfall‐induced landslides are published as initiation points for ease of use. In total, we created 22 inventories: 2 in Laos, 4 in Myanmar, 1 in Thailand and 15 in Vietnam. These inventories are being used to train the LHASA‐Mekong model and quantify the effects of Land use/Land cover change on landslide susceptibility. These open data will be a valuable resource for advancing landslide studies in this region.

INTRODUCTION

Cambodia, Laos, Myanmar, Thailand and Vietnam make up the Lower Mekong Region (LMR). Rainfall-induced landslides are a pervasive issue across the LMR, causing severe economic losses and hundreds of deaths annually (Figure 1). A landslide inventory detailing the spatial characteristics of past landslides is the most important information for formulating effective landslide hazard and risk mitigation strategies. Traditionally, these inventories are produced by manual mapping, a very cumbersome process. Inventories are also limited in mapped landslide size and mapping area coverage due to the coarse spatial resolution of freely available satellite imagery and infrequent acquisitions over the region by very high-resolution satellites.

FIGURE 1. Distribution of reported landslides across the LMR based on NASA’s Global Landslide Catalog (Kirschbaum et al., 2015) and susceptible areas to landslides according to NASA’s global landslide susceptibility map (Stanley & Kirschbaum, 2017)

The recent launch of constellations comprised of 180+ CubeSats such as PlanetScope and SkySat (www.planet.com) has made daily imaging of any location on the earth's surface at very high-resolution possible. Development of advanced automated and semi-automated landslide mapping techniques have enabled efficient handling of large volumes of images for generating landslide inventories with large areal coverage. Pixel-based (Borghuis et al., 2007; Nichol & Wong, 2005), Object-based image analyses (OBIA) (Amatya et al., 2019; Comert, 2021; Hölbling et al., 2012, 2015; Lu et al., 2011; Martha et al., 2010, 2016; Stumpf & Kerle, 2011) and Deep learning-based (Ghorbanzadeh et al., 2019; Meena et al., 2021; Sameen & Pradhan, 2019) approaches have been extensively used for landslide mapping all around the world.

In collaboration with the Asian Disaster Preparedness Center (ADPC) and the NASA SERVIR Programme (herein SERVIR-Mekong; https://servir.adpc.net), we are developing a regional version of our global Landslide Hazard Assessment System for Situational Awareness (LHASA) model (Kirschbaum & Stanley, 2018; Stanley et al., 2021) for the LMR (LHASA-Mekong). LHASA version 1 (Kirschbaum & Stanley, 2018) used a decision tree approach combining a global susceptibility map with satellite precipitation estimates to issue near real time hazard warning. Recently, LHASA was upgraded to version 2 (Stanley et al., 2021). LHASA version 2 follows a data-driven approach using machine learning. LHASA-Mekong adopts the framework of LHASA version 2. The first requirement for developing this system is landslide inventories from the region. Although open landslide data such as the Global Landslide Catalog (GLC) (Kirschbaum et al., 2015) and SERVIR-Mekong Myanmar Mapathon Landslides (SMMML) (NASA SERVIR Science Coordination Office & SERVIR Mekong, 2018) are available for this region, the coverage and spatial distribution of landslides in these data sets are not sufficient to properly train the LHASA-Mekong model. Since the LHASA-Mekong follows a data-driven approach, inventories used for training need to be dated. Although the GLC is dated, it is based on media reports and has some positional inaccuracies related with manual entry of landslides and uncertainties in reporting. SMMML is an inventory of landslides from Myanmar mapped during a Mapathon organized by the NASA/SERVIR-Mekong. Landslides were manually mapped in point form using Google Earth imagery. Although some of the landslides can be attributed to a specific event relying on media reports, most landslides could not be dated; thus, it cannot be used for training dynamic models like LHASA-Mekong. There are other country wide landslide databases from the region such as the one created during the State Funded Landslide Project (SLFP) for Vietnam (Hung et al., 2017) and for Thailand maintained by the Department of Mineral Resources (DMR). However, these inventories are not publicly available.

The NASA Commercial Smallsat Data Acquisition (CSDA) Programme (https://earthdata.nasa.gov/esds/csdap) provides access to the Planet archives for NASA-funded researchers. Based on media and government reports of rainfall-induced landslide events, we identified landslide mapping locations. An OBIA-based Semi-Automatic Landslide Detection (SALaD) system (Amatya et al., 2021) was modified by incorporating a change detection module (SALaD-CD) to create rainfall event-based inventories for the LMR in places where landslides were present on the Planet imagery. These inventories will be used to train and validate the LHASA-Mekong model and investigate the effects of changes in Land Use/Land Cover (LULC) on landslide susceptibility. As an open record of landslides in the region, this data set may be valuable for further landslide-related studies in this region.

DATA DESCRIPTION AND DEVELOPMENT Event identification and satellite data

We searched for media and government reports of rainfall-induced landslides in the LMR from 2009 to 2020. When we found reports, we searched for the specified dates and locations in Planet archives. If we found evidence of landslides, we downloaded pre- and post-event imagery for landslide mapping. In total, we were able to find 22 locations in the region. Identified locations are shown in Figure 2 with detailed media and government report links on Table 1. We were able to find 2 locations in Laos, 4 in Myanmar, 1 in Thailand and 15 in Vietnam. There were other landslide events in this region during the time period. We were unable to map at four locations due to unavailability of a pre- or post-event imagery. We also could not pinpoint 12 locations in Planet archives.

FIGURE 2. Landslide mapping locations identified from media and government reports in the LMR. Mapping area extent at identified locations are available in Footprint.shp file in the data set

TABLE 1 Media and government reports used to identify rainfall-induced landslide events sorted by date

District	Country	Event date	Source
Khao Phanom	Thailand	30 March, 2011	https://archive.iges.or.jp/files/research/bangkok/PDF/20190207/Presentation%202-Thailand%20Landslide%20Risk%20Management.pdf
Falam	Myanmar	30–31 July, 2015	https://reliefweb.int/report/myanmar/landslides-hit-chin-state
Hakha	Myanmar	30–31 July, 2015	https://reliefweb.int/report/myanmar/landslides-hit-chin-state
Thaphabath	Laos	11 September, 2015	https://www.nectec.or.th/sectionFileDownload/4946
Mu Chang Chai	Vietnam	2–3 August, 2017	http://floodlist.com/asia/vietnam-floods-landslides-yen-bai-son-la-august-2017
Muong La	Vietnam	2–3 August, 2017	http://floodlist.com/asia/vietnam-floods-landslides-yen-bai-son-la-august-2017
Bat Xat	Vietnam	23–28 August, 2017	https://mt.gov.vn/moitruong/quy-chuan-chat-luong/50541/lao-cai--nhieu-diem-sat-lo-tren-cac-tuyen-duong-tai-huyen-bat-xat-da-duoc-khac-phuc.aspx
Da Bac	Vietnam	10–11 October, 2017	http://www.baohoabinh.com.vn/PrintPreviewE/594/
Phu Yen	Vietnam	10–11 October, 2017	http://floodlist.com/asia/vietnam-tropical-depression-floods-landslides-october-2017
Tram Tau	Vietnam	10–11 October, 2017	https://en.vietnamplus.vn/yen-bai-mountainous-tram-tau-district-faces-landslide-risk/119774.vnp
Sin Ho	Vietnam	23–24 June, 2018	http://floodlist.com/asia/vietnamlai-chau-ha-giang-thai-nguyen-june-2018
Tam Duong	Vietnam	23–24 June, 2018	http://floodlist.com/asia/vietnamlai-chau-ha-giang-thai-nguyen-june-2018
Than Uyen	Vietnam	23–24 June, 2018	http://floodlist.com/asia/vietnamlai-chau-ha-giang-thai-nguyen-june-2018
Vi Xuyen	Vietnam	23–24 June, 2018	http://floodlist.com/asia/vietnamlai-chau-ha-giang-thai-nguyen-june-2018
Hpa-An	Myanmar	28–30 July, 2018	http://floodlist.com/asia/myanmar-floods-july-2018
Phong Tho	Vietnam	3 August, 2018	https://en.nhandan.org.vn/society/item/6466802-localities-strengthen-efforts-to-cope-with-floods-and-landslides.html
Xieng Ngeun	Laos	30 August, 2018	https://reliefweb.int/report/lao-peoples-democratic-republic/death-toll-rises-floods-landslides-laos
Muong Lat	Vietnam	27 August–1 September, 2018	https://vietnamnews.vn/society/465078/muong-lat-isolated-after-deadly-flash-floods-and-landslides.html
Nha Trang	Vietnam	18 November, 2018	http://floodlist.com/asia/vietnam-floods-landslides-toraji-november-2018
Thaton	Myanmar	9 August, 2019	http://floodlist.com/asia/myanmar-landslide-paung-mon-state-august-2019
Phong Dien	Vietnam	12 October 2020	https://blogs.agu.org/landslideblog/2020/10/14/rao-trang/
Huong Hua	Vietnam	18 October 2020	https://blogs.agu.org/landslideblog/2020/10/19/huong-phung-landslide-1/

We used imagery from PlanetScope and RapidEye (Planet Team, 2017) to map landslides. The choice between RapidEye and PlanetScope imagery is dictated by availability of pre- and post-imagery, cloud cover and areal coverage of the location. The imagery were available through the CSDA Programme (https://earthdata.nasa.gov/esds/csdap). RapidEye was a constellation of five satellites launched in 2008 and retired in 2020. PlanetScope is a constellation of 180+ CubeSats that started imaging the earth's surface from 2016. We used the RapidEye Ortho Tile, PlanetScope Ortho Tile and Ortho Scene products in this study. RapidEye and PlanetScope Ortho Tiles are available in 25 × 25 km tiles with spatial resolution of 5 and 3.125 m, respectively. Planet Ortho Scene is available in various tile sizes with minimum size of 24 × 8 km and maximum size of 32.5 × 19.6 km with spatial resolution of 3 m, depending upon the generation of satellites. Four bands (blue, green, red and near-infrared) with top of atmosphere radiance values were used for landslide mapping. Detailed specifications can be found in latest version of Planet imagery product specifications (https://assets.planet.com/docs/Planet_Combined_Imagery_Product_Specs_letter_screen.pdf). Planet ortho scenes were mosaiced to form tiles comprised of a maximum of four scenes and clipped to match the extent of pre- and post-tile overlap. Table 2 lists the types of satellites and products used to generate the landslide inventories.

TABLE 2 Identified landslide mapping locations and imagery used sorted by country

District	State/Province	Country	Satellite	Product	Pre-image (mm/dd/year)	Post-image (mm/dd/year)
Thaphabath	Bolikhamsai	Laos	RapidEye	Ortho Tile	12/24/2014	12/12/2015
Xieng Ngeun	Luang Prabang	Laos	PlanetScope	Ortho Scene	2/15/2018	11/2/2018
Falam	Chin	Myanmar	RapidEye	Ortho Tile	12/14/2014	2/18/2016
Hakha	Chin	Myanmar	RapidEye	Ortho Tile	12/14/2014	2/18/2016
Hpa-An	Kayin	Myanmar	PlanetScope	Ortho Scene	12/22/2017	12/5/2018
Thaton	Mon	Myanmar	PlanetScope	Ortho Scene	2/3/2019	11/19/2019
Khao Phanom	Krabi	Thailand	RapidEye	Ortho Tile	1/20/2010	3/21/2013
Vi Xuyen	Ha Giang	Vietnam	PlanetScope	Ortho Scene	5/28/2017	11/2/2018
Da Bac	Hoa Binh	Vietnam	RapidEye	Ortho Tile	2/8/2016	12/19/2017
Nha Trang	Khanh Hoa	Vietnam	PlanetScope	Ortho Scene	11/2/2018	11/21/2018
Phong Tho	Lai Chau	Vietnam	PlanetScope	Ortho Scene	3/2/2018	11/29/2018
Sin Ho	Lai Chau	Vietnam	RapidEye	Ortho Scene	2/8/2016	2/7/2019
Tam Duong	Lai Chau	Vietnam	PlanetScope	Ortho Scene	2/16/2018	11/3/2018
Than Uyen	Lai Chau	Vietnam	PlanetScope	Ortho Scene	4/8/2018	11/3/2018
Bat Xat	Lao Chai	Vietnam	PlanetScope	Ortho Scene	5/28/2017	12/20/2017
Huong Hua	Quang Tri	Vietnam	PlanetScope	Ortho Tile	3/7/2020, 3/8/2020	12/28/2020, 12/29/2020
Muong Lat	Tanh Hoa	Vietnam	PlanetScope	Ortho Tile	12/19/2017	11/1/2018
Phong Dien	Thua Thien-Hue	Vietnam	PlanetScope	Ortho Scene	1/14/2020	1/25/2021
Muong La	Son La	Vietnam	RapidEye	Ortho Tile	2/9/2016	12/22/2017
Phu Yen	Son La	Vietnam	RapidEye	Ortho Tile	2/8/2016, 2/11/2016	12/19/2017, 12/22/2017
Mu Chang Chai	Yen Bai	Vietnam	RapidEye	Ortho Tile	2/9/2016	12/22/2017
Tram Tau	Yen Bai	Vietnam	RapidEye	Ortho Tile	2/8/2016, 2/9/2016, 2/11/2016	12/19/2017, 12/22/2017

In addition to the optical data, we also utilized the 30 m NASADEM (Crippen et al., 2016), a reprocessed Shuttle Radar Topography Mission (SRTM) data set with improved height accuracy and void filling.

Landslide mapping

SALaD (Amatya et al., 2021), a Python-based open-source system, uses OBIA and machine learning to map landslides. Inputs to SALaD are a post-event image and a DEM. As it only uses a post-event image, it does not have the ability to separate old landslides from new event-induced ones. As we are focusing on rainfall event-based inventory, it is important we map landslides induced by that event only. A change detection-based approach utilizing pre- and post-event imagery can achieve this. In this study, we introduced a change detection-based module to the SALaD framework. We will refer to the new system as SALaD-CD. The new framework is shown in Figure 3. We follow a change detection approach proposed by Mondini, Guzzetti, et al. (2011), which has been successfully utilized in other studies (Lu et al., 2019; Mondini, Chang, et al., 2011). The new framework incorporates image normalization, image co-registration and change detection. All new steps were also implemented in Python. We first tested SALaD-CD in Hpa-An, Kayin state, Myanmar.

FIGURE 3. Flowchart showing each step of the SALaD-CD system. The blue boxes are inputs to the system; yellow boxes are manual steps; red boxes are the Python processes and black boxes are outputs

As pre- and post-imagery are taken at different times, each have different atmospheric, illumination and sensor conditions. A radiometric normalization must be done to match spectral characteristic between pre- and post-imagery. The spectral properties of the post-image was adjusted to the pre-image using the relative image normalization technique (Hong & Zhang, 2008).[Image Omitted. See PDF]where $S$ is the radiance of the post-image, $N$ is the radiance of new image normalized to pre-image, µ and $\sigma$ are the mean and variance of radiance values pre- and post-image.

A small spatial shift exists between two images acquired at different times. Image co-registration must be done to correct for this shift. Python package AROSICS (Scheffler et al., 2017) was used to co-register pre- to post-image. The co-registration is done using phase correlation for sub-pixel shift estimation in the frequency domain utilizing the Fourier shift theorem.

Three change detection techniques: NDVI difference ($\delta NDVI)$, Principal Component Analysis (PCA) and Independent Component Analysis (ICA) were used to map event-induced landslides. Landslides are synonymous with change in vegetation. There is a decrease in NDVI pre- and post-landslides. The $\delta NDVI$ is calculated as:[Image Omitted. See PDF]where ${\mathrm{NDVI}}_{\mathrm{pre}}$ is pre-image NDVI and ${\mathrm{NDVI}}_{\mathrm{post}}$ is post-image NDVI.

PCA and ICA are dimensionality reduction methods that transforms a set of variables to new components. PCA and ICA were conducted using a single stack image composed of four bands, red and near-infrared bands from pre- and post-images. PCA transforms variables into uncorrelated components whereas in ICA variables are transformed into statistically independent components. PCA and ICA produces set of four principal and four independent components. PCA and ICA were conducted using Scikit-learn package (Pedregosa et al., 2011). We visually selected the most suitable component each from four principial and four individual components that highlighted changes related to event-induced landslides. At last, we end up with three image derivatives comprising of $\delta \mathrm{NDVI}$, one principal component and one independent component. Figure 4 shows change detection-based image derivatives for Hpa-An, Kayin State, Myanmar.

FIGURE 4. Change between (a) pre-event (22 December 2017) and (b) post-event (5 December 2018) PlanetScope imagery highlighted by (c) 4th principal component, (d) 3rd independent component and e. NDVI difference for Hpa-An, Kayin state, Myanmar

The rest of the steps follow SALaD’s framework. In OBIA, similar pixels are grouped to form homogenous objects using image segmentation. The post-event image was segmented using Mean-Shift segmentation (Comaniciu & Meer, 2002). The segmentation parameters were calculated in the training area (Figure 5) using Plateau Objective Function (Martha et al., 2011). The mean of NDVI difference, principal component and independent component of each object were used for landslide classification using a Random Forests (RF) model (Breiman, 2001). The training data set for RF was created in a 25 km² training area within the post-image using a manual inventory (Figure 5). The segmented objects that had an overlap of 75% or more with the manual landslides were chosen as landslide objects, and all other objects within the training area were set as non-landslide objects. Based on this data, a RF classifier with 500 trees was trained and remaining objects were classified. The classified landslide objects were dissolved to form final landslide areas.

FIGURE 5. Training and validation areas for a tile of PlanetScope image acquired on 5 December, 2018 at Hpa-An, Kayin state, Myanmar. The red tile highlights the subset area used for training and yellow tile highlights the validation area

We had to use multiple image tiles to map in an area. For example, the Hpa-An landslide mapping area was covered by three image tiles. Since the PCA and ICA values change between tiles, the process described above was repeated for each image tile used in mapping.

Validation

Figure 6 shows the landslides mapped manually and detected by SALaD-CD. The SALaD-CD detected landslides were compared with manually mapped landslides in a 25 km² validation area (Figure 5). Three metrics based on overlapping area were calculated: true positive (TP), false negative (FN) and false positive (FP). TPs are correctly detected landslides; FPs are detected landslides that have not been mapped manually and FNs are manually mapped landslides not detected by the SALaD-CD. Based on these metrics, the two accuracy indices, producer accuracy (PA) and user accuracy (UA) were calculated as follows:[Image Omitted. See PDF][Image Omitted. See PDF]

FIGURE 6. Insets highlighting landslides: (a) mapped manually; (b) detected by the SALaD-CD

The PA denotes how much of the manual inventory was detected. The UA denotes how much of the detected landslides are actual landslides. It was observed that the SALaD-CD was successful in detecting 83.14% of the area of the manually mapped inventory (Table 3). When evaluating based on the intersection of any portion of the manual and SALaD-CD detected landslides, a PA of 97% was obtained. This suggests that most landslides were detected using SALaD-CD, but the areas were not always accurate.

TABLE 3 Comparison of SALaD-CD and manually mapped landslides based on overlapping landslide area

True positive (m2)	False positive (m2)	False negative (m2)	Producer accuracy (%)	User accuracy (%)
384,206.6	201,843.9	778,88.51	83.14	65.55

Post processing

The final output from SALaD-CD contained false positives such as agricultural areas, built-up areas, river channels, barren areas, etc. These types of false positives were manually removed. Amalgamation due to dissolving of landslide objects during production of the final landslide areas were also resolved manually. As we are using this inventory to train LHASA-Mekong, which takes inventories in point form, and to avoid time consuming area corrections, we used NASADEM to calculate initiation points from each landslide polygons. The initiation point was assumed to be the highest elevation on the landslide boundary. The final inventory was obtained in the form of initiation points (Figure 7). We also manually mapped some initiation points for obviously missed landslide areas. 992 initiation points were obtained at Hpa-An, Kayin state, Myanmar.

FIGURE 7. Initiation points obtained from landslide areas and NASADEM

Final landslide inventories

The steps described above were used to produce landslide inventories in the 21 remaining identified locations to form final initiation point-based inventories.

DATA ACCESS

The data are available through Figshare (https://doi.org/10.6084/m9.figshare.14199227) as a shapefile. Landslides_LMR.shp contains all initiation point-based landslides mapped during this study. We have also provided footprints of Planet imagery used for landslide mapping in Footprint.shp. Metadata related to the shapefiles is provided in Metadata.pdf. We will continue to monitor this region for future landslide events. Any new inventories produced in the future will be hosted in NASA’s Landslide Viewer (https://landslides.nasa.gov/viewer).

POTENTIAL DATA USE AND REUSE

We created 22 new rainfall-induced landslide inventories for the LMR. These inventories are released as initiation points. We did not release them as polygons for two main reasons. First, SALaD-CD had trouble mapping complete landslide areas. In our case, we underestimated landslide area by 16.86% in comparison to a manual inventory (Table 3). Second, due to a long-time gap between some of the pre- and post-event imagery at a few locations some landslide areas had recovered due to revegetation. During visual checks, we found landslide scarps and large landslides were mapped well but long narrow runouts were not detected or had already been revegetated. Hence, to avoid release of polygons with missing areas, we resolved amalgamation manually and converted them to initiation points, which are more robust and complete. Hao et al. (2020) also converted landslide polygons obtained by OBIA to initiation points and highlighted various other challenges associated with automatic mapping. These initiation points have the potential to support a wide range of research locally, regionally and globally.

At the local scale, these data are very useful in quantifying local susceptibility (Bui et al., 2011; Lee & Min, 2001), exploring the relationship between landslides and causative factors such as slope, land cover, lithology, etc. (Zhou et al., 2002) and establishing local rainfall thresholds by examining relationship between rainfall intensity and duration for landsliding (Caine, 1980). The Regional Land Cover Monitoring System (RLCMS) developed by SERVIR-Mekong provides yearly land cover maps at 30 m resolution from 2000 to 2018 (https://www.landcovermapping.org/en/landcover/). Land cover data from RLCMS can be used to study the effects of LULC change on landslide susceptibility (Chen et al., 2019; Shu et al., 2019). The data may also be useful for guiding further data collection efforts, either through fieldwork or further image analysis. In the former case, these data would enable thorough planning of site investigations that could further characterize landslide hazard or identify impacts on local communities. In the latter case, analysts could search near mapped landslides for impacted buildings and infrastructure or better describe the characteristics of individual landslides. The initiation points can also be used to model runout using empirical models (Guthrie & Befus, 2021; Horton et al., 2013; Schilling, 1998).

At the regional scale, these data enable regionalization of global landslide hazard models such as LHASA. These data in combination with landslide databases maintained by countries in the LMR and other open data such as GLC can be used to train and test a dynamic hazard model of rainfall-triggered landslides like LHASA-Mekong, which will provide near real time estimates of landslide hazard across the region.

At the global scale, these public inventories with dates can be combined with other publicly available dated rainfall-induced landslide inventories (Emberson et al., 2021a) to further characterize the influence of extreme rainfall on landsliding (Marc et al., 2018). These inventories can also be used as additional data to train and test global modelling efforts such as LHASA (Stanley et al., 2021) and can, in turn, improve quantification of population, infrastructure and roads exposed to landslide hazard (Emberson et al., 2020, 2021b). These inventories can also contribute to the global data set used to investigate the potential of a global precipitation forecast (Khan et al., 2021) for use in a landslide forecasting system. These data will also be a valuable addition to the GLC (Kirschbaum et al., 2010, 2015), which has been used for various purposes globally (Benz & Blum, 2019; Jia et al., 2020, 2021; Stanley & Kirschbaum, 2017; Stanley et al., 2021).

SALaD-CD uses a change detection approach to map landslides. Hence, pre-event or old landslides are not mapped as they are not highlighted by PCA, ICA and NDVI difference. During our visual inspection, we did not find large number of pre-event landslides at the locations we identified. However, since we cannot release the Planet imagery used to create these inventories, we advise users to exercise caution particularly for use of these data in landslide susceptibility and hazard modelling where data need to be divided into landslide and non-landslides for training. We advise users to utilize Google Earth to check whether partitions were appropriate.

CONCLUSION

In this study, we introduce 22 new rainfall-induced landslide inventories: 2 in Laos, 4 in Myanmar, 1 in Thailand and 15 in Vietnam, which were created using Planet imagery and SALaD-CD. Locations for producing these inventories were identified by media and government reports. We acknowledge that other landslides have occurred in this region but were not able to locate or map them using Planet imagery. These inventories, in combination with GLC, are being used to train the LHASA-Mekong model and to quantify the effects of LULC change on landslide susceptibility. These open inventories will be a valuable resource for advancing landslide science and a starting point towards quantifying landslide hazard and risk in the LMR.

ACKNOWLEDGEMENTS

This research was funded by the NASA SERVIR Science Team (NNH18ZDA001N-18-SERVIR18_2-0036) and the collaboration with ADPC was supported via the joint US Agency for International Development (USAID) and NASA initiative SERVIR-Mekong. We would like to thank the NASA CSDA Programme for providing access to Planet data. We would like to thank the SERVIR-Mekong team at ADPC, SERVIR Coordination Office team and Spatial Informatics Group for helpful discussion during production of the data. We would like to thank Dr. Nishan Kumar Biswas for processing NASADEM of the LMR. We would also like to thank Dr. Robert Emberson for helpful discussions on data use and reuse.

CONFLICT OF INTEREST

The authors declare that they have no conflict of interest.

AUTHOR CONTRIBUTIONS

Pukar Amatya: Conceptualization (equal); Data curation (equal); Formal analysis (lead); Funding acquisition (supporting); Investigation (lead); Methodology (lead); Software (lead); Supervision (equal); Validation (lead); Visualization (lead); Writing – original draft (lead); Writing – review & editing (equal). Dalia Kirschbaum: Conceptualization (equal); Funding acquisition (lead); Supervision (equal); Writing – review & editing (equal). Thomas Stanley: Data curation (equal); Funding acquisition (supporting); Writing – review & editing (equal).

DATA AVAILABILITY STATEMENT

Data are available in Figshare (https://doi.org/10.6084/m9.figshare.14199227).