Global landslide susceptibility prediction based on the automated machine learning (AutoML) framework

Figures & data

Figure 1. The study area and its DEM.

Figure 2. Thematic map of the lithological factor.

Table 1. Lithology abbreviation in Figure 2.

Abbreviation	Full name
vi	Intermediate Volcanic R.
vb	Basic volcanic Rocks
pa	Acid Plutonic Rocks
mt	Metamorphic Rocks
su	Unconsolidated Sediments
ss	Siliciclastic Sedimentary Rocks
pb	Basic Plutonic Rocks
pi	Intermediate Plutonic R.
wb	Water Bodies
py	Pyroclastics
sc	Carbonate Sedimentary Rocks
va	Acid volcanic Rocks
nd	No Data
ev	Evaporites
ig	Ice and Glaciers
sm	Mixed Sedimentary Rocks

Figure 3. Thematic map of the precipitation factor.

Figure 4. Thematic map of the faults distance factor.

Figure 5. Global landslide inventory map.

Table 2. Datasets for global LSP.

Landslide database source	Database level	Number of landslides	Source and description
COOLR (Juang et al. 2019)	Global	6343	It is based on the in-depth development of the Global Landslide Catalog initiated by NASA, which records the source and source link of the landslide, when the landslide occurred, where it occurred, the triggering cause, the specific description of the triggering event, the scale of landslide and the data accuracy of landslides since 2007.
GFLD (Froude and Petley 2018)	Global	277	It records the time, location, trigger, casualty reports and data accuracy of life-threatening landslides from 2004 to 2017. The global-scale dataset was filtered based on its spatial resolution, and landslide data with a spatial resolution of less than 1000 m were finally selected.
FraneItalia (Calvello and Pecoraro 2018)	National	6708	It records landslide data in Italy from 2010 to 2019, with additional information including the type of induced landslide and the level of loss of life and property after the disaster. The landslide data accuracy levels are classified as deterministic, approximate and municipally provided landslide data.
Australian Landslide Database (Osuchowski and Atkinson 2008)	National	342	It records the time intervals from 2008 to 2018, and this landslide data accuracy grade is classified as GPS surveys, GIS location, map location, satellite imagery, local reports, report-based landslide locations, and unknown, depending on the credibility of the data source.
NZLD (Rosser et al. 2017)	National	3480	It is also a national dataset, but this dataset lacks accuracy classification.
Washington State Landslide Database, USA (Slaughter et al. 2017)	Provincial	23879	It was classified as low, medium and high. High-precision landslide data were uniformly selected for data screening.
Landslide Database in Utah, USA (Elliott and Harty 2010)	Provincial	1725	It was classified as low, medium and high. High-precision landslide data were uniformly selected for data screening.
90m spatial resolution total landslide data volume		42754
Amount of landslide data after 1000 m resampling		14290

Table 3. Datasets for regional LSP.

Landslide database source	Database level	Number of landslides	Source and description
Ireland (Mckeon 2016)	National	1542	It is provided by the Geological Survey of Ireland, which provides information such as the occurrence time, location, triggering causes and data accuracy status judgment of landslide data, and can be screened according to the data accuracy status, which is divided into certain and uncertain. Due to the long time span of landslide data cataloged in this landslide database, landslide data from definite sources are finally selected as the test set.
Denmark (Luetzenburg et al. 2022)	National	4771	It is of high quality, and the data source paper contains the comparative map of the mountain shadow model and the orthophoto image of the landslide data with high reliability, and all of them are included in the validation set.
Norway	National	13458	It is provided by the Norwegian Water Resources and Energy Authority, and includes information such as the occurrence time, location and cause, as well as the determination of the data accuracy status. Landslide sites with a spatial resolution of 1000 m or less were selected into the test set.
Veneto Province, Italy	Provincial	33770	The landslide database of Veneto Province, Italy. The database of hydrological hazards in Apulian Province, Italy records landslides caused by geo-hydrological hazards, including information on the occurrence time, location, and induced causes of landslide data.
Apulia Province, Italy (Vennari et al. 2022)	Provincial	2023	The hydrological hazard database of Apulia Province, Italy. It records landslides caused by geo-hydrological hazards, including information on the occurrence time, location, and induced causes of landslide data.
Shaanxi Province, China	Provincial	10158	The data is a geological disaster that caused loss of life and property in Shaanxi Province from 2000 to 2016. It was provided by the Shaanxi Provincial Geological Survey Institute of China. In the data screening, the landslide data with landslide time description was selected as the research target data.
Hokkaido, Japan	Prefectural	21809	The Japanese Hokkaido Earthquake Landslide Database records landslide data triggered by the Ibri earthquake, including new triggered landslides after this earthquake, and historical landslide data.
Yongxin County, China	Prefectural	364	This data were selected from Yongxin County, China, and all available datasets were from Jiangxi Meteorological Bureau and Jiangxi Department of Land and Resources.

Figure 6. Schematic diagram of the proposed framework.

Table 4. Parameter settings for different methods of global landslide susceptibility.

Model	Parameter	Value
Auto-PyTorch	optimize_metric	‘accuracy’
	total_walltime_limit	600
	func_eval_time_limit_secs	600
	memory_limit	24576
	run time	10 mins
TPOT	generations	4
	population_size	10
	cv	2
	random_state	42
	verbosity	2
	run time	10 mins
RF	n_estimators	Default
	criterion
	max_depth
	max_features
NB	prior	Default
	var_smoothing

Table 5. Parameter settings for Auto-PyTorch of regional landslide validation.

Model	Parameter	Value
Auto-PyTorch	optimize_metric	‘accuracy’
	total_walltime_limit	600
	func_eval_time_limit_secs	1800
	memory_limit	24576
	run time	10 mins

Table 6. Parameter settings for different methods of performance comparison at different training times.

Model	Parameter	Value
Auto-PyTorch	optimize_metric	‘accuracy’
	total_walltime_limit	600
	func_eval_time_limit_secs	600, 1200, 1800
	memory_limit	24576
	run time	10 mins, 20 mins, 30 mins
TPOT	generations	4, 8, 10
	population_size	10
	cv	2
	random_state	42
	verbosity	2
	run time	10 mins, 20 mins, 30 mins
RF	n_estimators	Default
	criterion
	max_depth
	max_features

Figure 7. ROC curves of different prediction methods.

Table 7. Evaluation indications of different prediction methods.

Models	ACC	AUC	RMSE	MAE
Auto-PyTorch	0.9014	0.9632	0.3139	0.0985
TPOT	0.8911	0.9564	0.3298	0.1088
RF	0.8967	0.9585	0.3212	0.1032
NB	0.7465	0.8393	0.5034	0.2534

Figure 8. Global landslide susceptibility map by Auto-PyTorch.

Figure 9. ROC curves of Auto-PyTorch.

Table 8. Performance of Auto-PyTorch at different spatial resolution with a training time of 30 min.

Models	ACC	AUC	RMSE	MAE
Auto-PyTorch (90 m)	0.9105	0.9740	0.2991	0.0895
Auto-PyTorch (1000 m)	0.9028	0.9624	0.3116	0.0971

Table 9. AUC values of different national datasets.

National datasets	AUC
	Non-landslide samples from outside the buffer zones	Non-landslide samples from low susceptible areas in Figure 8
Ireland	0.5431	0.9746
Denmark	0.9383	0.9925
Norway (part)	0.7416	0.9564

Figure 10. National landslide susceptibility maps by Auto-PyTorch. (a) and (d) Ireland, (b) and (e) Denmark, (c) and (f) Norway (part). the first line shows the results of Auto-PyTorch when the specified area is selected in the global LSP map., and the second line shows the results of Auto-PyTorch when non-landslide samples are selected from the low susceptible areas of the obtained global landslide susceptibility maps.

Figure 11. Provincial landslide susceptibility maps by Auto-PyTorch. (a) and (d) Veneto Province, Italy, (b) and (e) Apulia Province, Italy, (c) and (f) Shaanxi Province, China. The first line shows the results of Auto-PyTorch when the specified area is selected in the global LSP map, and the second line shows the results of Auto-PyTorch when non-landslide samples are selected from the low susceptible areas of the obtained global landslide susceptibility maps.

Table 10. AUC values of different provincial datasets.

Provincial datasets	AUC
	Non-landslide samples from outside the buffer zones	Non-landslide samples from low susceptible areas in Figure 8
Province of Veneto, Italy	0.7815	0.9762
Province of Apulia, Italy	0.7416	0.8873
Shaanxi Province, China	0.7898	0.9512

Figure 12. Prefectural landslide susceptibility maps by Auto-PyTorch. (a) and (c) Yongxin County, China, (b) and (d) the earthquake zones of Hokkaido, Japan. The first line shows the results of Auto-PyTorch when the specified area is selected in the global LSP map, and the second line shows the results of Auto-PyTorch when non-landslide samples are selected from the low susceptible areas of the obtained global landslide susceptibility maps.

Table 11. AUC values of different prefectural datasets.

Prefectural datasets	AUC
	Non-landslide samples from outside the buffer zones	Non-landslide samples from low susceptible areas in Figure 8
Yongxin County, China	0.6674	0.9603
the earthquake zones of Hokkaido, Japan	0.5842	0.9820

Figure 13. 1000m spatial landslide susceptibility map (Yongxin County, China).

Figure 14. Evaluation indicators values over time over time for the two AutoML models and the RF model. (a) ACC, (b) AUC, (c) MAE and (d) RMSE.

Table 12. Performance of two AutoML models under different training time.

Model	Training time (min)	ACC	AUC	RMSE	MAE
Auto-PyTorch	10	0.9014	0.9632	0.3139	0.0985
	20	0.9014	0.9632	0.3139	0.0985
	30	0.9028	0.9624	0.3116	0.0971
TPOT	10	0.8911	0.9564	0.3298	0.1088
	20	0.8937	0.9600	0.3259	0.1063
	30	0.8981	0.9612	0.3193	0.1019
RF	\	0.8967	0.9585	0.3212	0.1032

Juang CS, Stanley TA, Kirschbaum DB. 2019. Using citizen science to expand the global map of landslides: introducing the Cooperative Open Online Landslide Repository (COOLR). PLoS One. 14(7):e0218657. doi: 10.1371/journal.pone.0218657.

 PubMed Web of Science ®Google Scholar

Froude MJ, Petley DN. 2018. Global fatal landslide occurrence from 2004 to 2016. Nat Hazards Earth Syst Sci. 18(8):2161–2181. doi: 10.5194/nhess-18-2161-2018.

 Web of Science ®Google Scholar

Calvello M, Pecoraro G. 2018. FraneItalia: a catalog of recent Italian landslides. Geoenviron Disasters. 5(1):1–16. doi: 10.1186/s40677-018-0105-5.

 Google Scholar

Osuchowski M, Atkinson R. 2008. Connecting diverse landslide inventories for improved information in Australia. Proceedings of the First World Landslide Forum, Tokyo, Japan. p. 67–70.

 Google Scholar

Rosser B, Dellow S, Haubrock S, Glassey P. 2017. New Zealand’s national landslide database. Landslides. 14(6):1949–1959. doi: 10.1007/s10346-017-0843-6.

 Web of Science ®Google Scholar

Slaughter S, Burns W, Mickelson K, Jacobacci K, Biel A, Contreras T. 2017. Protocol for landslide inventory mapping from lidar data in Washington State. Washington Geological Survey, USGS. p. 1–26.

 Google Scholar

Elliott A, Harty K. 2010. Landslide maps of Utah, Ogden 30 ‘X 60’ quadrangle: Utah Geological Survey Map 246DM, 14. p. 46 plates, scale 1: 100,000.

 Google Scholar

Mckeon C. 2016. National landslide susceptibility mapping project summary. Final Report. Geological Survey Ireland.

 Google Scholar

Luetzenburg G, Svennevig K, Bjørk AA, Keiding M, Kroon A. 2022. A national landslide inventory for Denmark. Earth Syst Sci Data. 14(7):3157–3165. doi: 10.5194/essd-14-3157-2022.

 Web of Science ®Google Scholar

Vennari C, Salvati P, Bianchi C, Casarano D, Parise M, Basso A, Marchesini I. 2022. AReGeoDatHa: Apulian regional GeoDatabase for geo-hydrological hazards. J Environ Manage. 322:116051. doi: 10.1016/j.jenvman.2022.116051.

 Web of Science ®Google Scholar

Data availability statement

The data and codes are available upon request to the authors.