Performance comparison of landslide susceptibility mapping under multiple machine-learning based models considering InSAR deformation: a case study of the upper Jinsha River

Figures & data

Figure 1. Location of the study area and distribution of landslides along the Jinsha River.

Figure 2. Examples of landslides identified by geomorphic features and InSAR deformation results.

Table 1. Landslide-related environmental factors categories in the study area.

Factor	Resolution	Date	Source
Dem/m	30 m	2000	SRTM. http://gdex.cr.usgs.gov/gdex
Slope/°	30 m	–	DEM products
Aspect	30 m	–	DEM products
Curvature (plan and profile)	30 m	–	DEM products
Slope length/m	30 m	–	DEM products
TPI	30 m	–	DEM products
Lithology	1:500,000	–	China Geological Survey (CGS)
Distance to earthquake/m	–	2022	https://www.cea.gov.cn/
Distance to faults/m	–	2020	China Geological Survey (CGS)
Distance to rivers/m	–	2020	China Geological Survey (CGS)
Rainfall/mm	5 km	2010–2022	NASA. https://pmm.nasa.gov
Temperature/°C	5 km	2010–2022	NASA. https://pmm.nasa.gov
SPI	30 m	–	DEM products
TWI	30 m	–	DEM products
LULC	10 km	2015–2020	https://www.resdc.cn
NDVI	5 km	2010–2020	http: //www.gscloud.cn
Distance to roads/m	30 m	2020	–
Sentinel-1 SAR data	Range 2.3 m, azimuth 14 m	2017–2020	ESA. https://scihub.copernicus.eu

Table 2. Detail parameters of SAR images.

SAR Data	Time period	Scenes	Path	Frame	Incidence	Azimuth
Sentinel-1 (Ascending)	20170120–20201227	97	99	1280–1285	36.9°	−12.75°
Sentinel-1 (Descending)	20170115–20201222	94	33	492–487	43.9°	192.77°

Figure 3. Research methodological flow of the study.

Figure 4. Landslides-influencing factors. (a) dem, (b) aspect, (c) slope, (d) plan curvature, (e) plan curvature, (f) slope length, (g) TPI, (h) lithology, (i) distance to earthquakes, (j) distance to faults, (k) distance to rivers, (l) rainfall, (m) temperature, (n) SPI, (o) TWI, (p) LULC, (q) NDVI, (r) distance to roads, (s) LOS deformation acquired from ascending images, (t) LOS deformation acquired from descending images.

Figure 5. Correlation coefficient of each covariate based on Wilcoxon signed-rank test.

Table 3. Detail feature importance’s calculated by four ranking algorithms.

Methods
Chi -squared		Information gain		Rank correlation		Random forest importance
Aspect	0.2609	Aspect	0.1028	Aspect	0.2517	Aspect	115.1074
Earthquake	0.2346	Deformation	0.0989	Deformation	0.2363	Rainfall	95.1158
Deformation	0.2317	Rainfall	0.0906	Elevation	0.2067	Temperature	90.5271
Elevation	0.2184	Temperature	0.0577	Earthquake	0.2035	Earthquake	78.3104
River	0.1979	Earthquake	0.0493	River	0.1966	Deformation	66.4378
Rainfall	0.1872	NDVI	0.0332	Rainfall	0.1932	Road	63.0743
Temperature	0.1808	Fault	0.0306	Temperature	0.1903	NDVI	61.8288
NDVI	0.1750	Road	0.0301	NDVI	0.1448	Fault	59.9211
Fault	0.1343	River	0.0272	Fault	0.1380	LULC	54.1465
Road	0.1209	Elevation	0.0271	Road	0.1280	River	54.1271
LULC	0.0995	Lithology	0.0160	LULC	0.1008	Lithology	42.6985
Lithology	0.0852	LULC	0.0066	Lithology	0.0952	Elevation	40.9675
Plan curvature	0.0000	Plan curvature	0.0000	Plan curvature	0.0000	TPI	39.3102
Profile curvature	0.0000	Profile curvature	0.0000	Profile curvature	0.0000	Slope	33.4503
Slope	0.0000	Slope	0.0000	Slope	0.0000	Slope length	32.5726
Slope length	0.0000	Slope length	0.0000	Slope length	0.0000	Profile curvature	24.6031
SPI	0.0000	SPI	0.0000	SPI	0.0000	Plan curvature	19.3962
TPI	0.0000	TPI	0.0000	TPI	0.0000	TWI	9.2388
TWI	0.0000	TWI	0.0000	TWI	0.0000	SPI	4.8004

Figure 6. Landslide susceptibility mapping generated by logistic regression, random forest, SVM, and XGBoost models. (a–d) LSM without considering deformation, (e–h) LSM considering deformation.

Table 4. Statistical pixel values of landslide susceptibility before and after the introduction of deformation.

Mode	Class	No def		Def		Change rate (%)
		Grids	Percentage	Grids	Percentage
LR	Very low	23,696	9.83	27,321	11.34	15.30
	Low	53,390	22.16	58,237	24.17	9.08
	Medium	65,434	27.15	70,726	29.35	8.09
	High	61,663	25.59	56,279	23.36	−8.73
	Very high	36,788	15.27	28,408	11.79	−22.78
RF	Very low	44,329	18.40	43,145	17.90	−2.70
	Low	87,517	36.32	88,012	36.52	0.55
	Medium	71,886	29.83	70,697	29.34	−1.65
	High	25,084	10.41	26,407	10.96	5.29
	Very high	12,155	5.04	12,710	5.27	4.48
SVM	Very low	12,260	5.09	11,305	4.69	−7.82
	Low	31,225	12.96	32,525	13.50	4.18
	Medium	167,090	69.34	169,873	70.50	1.67
	High	20,551	8.53	18,911	7.85	−7.95
	Very high	9845	4.09	8357	3.47	−15.07
XGBoost	Very low	177,536	57.18	160,940	51.83	−9.36
	Low	48,453	15.60	51,111	16.46	5.51
	Medium	26,356	8.49	30,570	9.85	16.02
	High	20,999	6.76	26,121	8.41	24.41
	Very high	37,164	11.97	41,766	13.45	12.36

Table 5. Statistical pixel values of landslide susceptibility in landslide area before and after deformation introduction.

Mode	Class	No def		Def
		Grids	Percentage	Grids	Percentage
LR	Very low	0	0	0	0
	Low	2	0.04	57	1.33
	Medium	248	5.81	688	16.10
	High	623	14.58	3297	77.18
	Very high	3399	79.56	473	11.07
RF	Very low	0	0	0	0
	Low	8	0.18	6	0.17
	Medium	266	5.89	134	2.97
	High	671	14.86	498	11.03
	Very high	3570	79.07	3877	85.87
SVM	Very low	181	4.66	177	4.56
	Low	69	1.78	68	1.75
	Medium	425	10.94	513	13.20
	High	700	18.02	743	19.12
	Very high	3140	80.82	3014	77.58
XGBoost	Very low	286	4.71	254	4.18
	Low	60	0.99	69	1.14
	Medium	111	1.83	90	1.48
	High	301	4.96	175	2.88
	Very high	5313	87.51	5483	90.31

Figure 7. Typical regional susceptibility optimization results. (a–d) the susceptibility results and deformation information of L1 landslide (marked in Figure 6b), a random forest susceptibility results without considering deformation, (b) random forest susceptibility results considering deformation. (e–h) The susceptibility results and deformation information of L2 landslide (marked in Figure 6b), (e) random forest susceptibility results without considering deformation, (f) random forest susceptibility results considering deformation.

Figure 8. ROC curves of four susceptibility models. (a) ROC test without considering deformation characteristic factors, (b) ROC test considering deformation characteristic factors.

Table 6. Statistical confusion matrix of the four susceptibility models.

Model		Accuracy	Kappa	Precision	Recall	F1
Logistic Regression	No def	0.666	0.318	0.612	0.971	0.751
	Def	0.657	0.300	0.606	0.958	0.743
Random Forest	No def	0.793	0.582	0.747	0.907	0.819
	Def	0.878	0.600	0.746	0.944	0.824
SVM	No def	0.711	0.417	0.684	0.819	0.745
	Def	0.734	0.463	0.707	0.828	0.762
XGBoost	No def	0.776	0.554	0.846	0.692	0.761
	Def	0.812	0.627	0.940	0.699	0.789

Figure 9. Significance of landslide-related environmental factors of the study area predicted by XGBoost and random forest models.

Figure 10. Summary plot of SHAP values derived from RF model.

Data Availability Statement

Sentinel-1A data used in this study were provided by European. Space Agency (ESA) through the Sentinel-1 Scientific Data Hub.