Multi-Regional landslide detection using combined unsupervised and supervised machine learning

Figures & data

Figure 1. An example Random Forest with three Decision Trees.

Table 1. Landslide events used in this study.

ID	Event date	Longitude	Latitude	Landslide trigger	Country	# Landslide segments
L0	25-3-2017	−76.6625	1.168792	Downpour	Colombia	2
L1	27-6-2017	103.6529	32.06849	Continuous rain	China	28
L2	14-8-2017	−13.22985	8.43611	Unknown	Sierra Leone	3
L4	18-5-2016	80.431975	7.160808	Monsoon	Sri Lanka	5
L7	27-11-2016	101.613	3.366	Rain	Malaysia	1
L9	20-4-2017	−119.749	37.6786	Unknown	USA	1
L15	20-5-2017	−121.43238	35.865628	Rain	USA	13
L16	2-4-2017	−118.16398	47.88398	Downpour	USA	1
L17	18-1-2017	13.77821	42.42921	Earthquake	Italy	2
L18	3-7-2017	167.9676	−44.6861	Rain	New Zealand	2
L20	9-4-2017	111.82664	−7.723228	Downpour	Indonesia	1
L21	6-7-2017	130.93551	33.38917	Tropical cyclone	Japan	2
L22	6-7-2017	130.9002	33.38442	Tropical cyclone	Japan	6
L23	6-7-2017	130.76208	33.39482	Tropical cyclone	Japan	8
L25	6-7-2017	130.820741	33.403127	Tropical cyclone	Japan	1
L26	24-5-2017	130.811309	33.403094	Tropical cyclone	Japan	2
L27	24-5-2017	131.1126	33.4422	Tropical cyclone	Japan	4
L28	24-5-2017	131.089613	33.48504	Tropical cyclone	Japan	1
L40	11-4-2018	130.87977	33.52926	Unknown	Japan	6
L41	11-4-2018	130.896605	33.430754	Unknown	Japan	2
L43	6-9-2018	141.92972	42.71625	Earthquake	Japan	23
L47	6-9-2018	141.99075	42.716242	Earthquake	Japan	6
L48	6-9-2018	142.05179	42.71693	Earthquake	Japan	9
L55	15-11-2016	173.81687	−42.202247	Earthquake	New Zealand	15
L56	15-11-2016	173.81687	−42.157233	Earthquake	New Zealand	6
L57	15-11-2016	173.756415	−42.112203	Earthquake	New Zealand	2
L58	15-11-2016	173.877412	−42.202231	Earthquake	New Zealand	15
L59	15-11-2016	173.87741	−42.157217	Earthquake	New Zealand	17
L75	19-6-2018	85.890295	27.908346	Rain	Nepal	1

Figure 2. Methodology Overview. (a) Pre-processing. (b) Image segmentation and feature extraction. (c) Segment classification and landslide detection (modified after Herrera Herrera 2019).

Figure 3. Image pre-processing stages (landslide L0, see Table 1 for details) (modified from Herrera Herrera 2019).

Figure 4. Elbow plots for the landslides L0, L1, L7, L17 (from Herrera Herrera 2019).

Table 2. Landslide diagnostic features at segment level.

Features	Feature type	Statistics per segment	Target Image	Source
B2, B3, B4, B8 bands	Spectral	mean	post-event	Sentinel-2
NDVI	Spectral	mean	post-event	Sentinel-2
GNDVI	Spectral	mean	post-event	Sentinel-2
Brightness	Spectral	mean	post-event	Sentinel-2
RGD	Spectral	mean	pre- and post-event	Sentinel-2
VID	Spectral	mean	pre- and post-event	Sentinel-2
BRD	Spectral	mean	pre- and post-event	Sentinel-2
NDVIstd	Contextual	mean	post-event	Sentinel-2
Slope	Spatial	mean, maximum	pre-event	AW3D30(2011)
Elevation	Spatial	mean, maximum, minimum	pre-event	AW3D30(2011)

Figure 5. Image pre-processing and segmentation (from Herrera Herrera 2019); sample in a remote area in Italy (L17 in Table 1). (a) Cloud-free pre-landslide image. (b) Cloud-free post-landslide image. (c) Image difference using RGD. (d) Image segmentation.

Figure 6. Correlation map of features used in this study for the selected landslide and non-landslide segments.

Figure 7. Train and test sets used for RF1 and RF2.

Table 3. RF candidate hyper-parameters in this study.

Hyper-parameter	Candidates	RF1	RF2
'bootstrap'	True	True	True
'max_depth'	10, 20, 50, 100, 500	10	10
'max_features'	2	2	2
'min_samples_leaf'	1, 2, 3, 4, 5	1	2
'min_samples_split'	2, 5, 10	5	10
'n_estimators'	10, 50, 100, 200, 500, 1000	50	10

Figure 8. Feature importance using the final model and the corresponding train dataset: (a) RF1, (b) RF2.

Table 4. Performance metrics of the trained and tested RF models.

Model	LS	NLS	TP	TN	FP	FN	Accuracy	Precision	Recall	F1-score
RF1 - train	147	213	145	213	0	2	0.99	0.99	1.00	1.00
RF1 - test	35	55	33	53	2	2	0.96	0.96	0.96	0.96
RF2 - train	135	208	135	204	4	0	0.99	0.98	1.00	0.99
RF2 - test	49	73	39	66	7	10	0.86	0.90	0.87	0.89

Table 5. Segment classes in classification.

Class	LTR Limits	Description	Binary interpretation	Color code
1	0-0.33	Non-landslide	Non-landslide	green
2	0.33-0.50	Uncertain non-landslide	Non-landslide	yellow
3	0.50-0.67	Uncertain landslide	Landslide	orange
4	0.67-1.00	Landslide	Landslide	red

Figure 9. Landslide detection for L0: (a) post-landslide optical image, (b) segmentation and final classification with RF1 model, (c) segmentation and final classification with RF2 model post-landslide image.

Figure 10. Landslide detection for L1: (a) post-landslide optical image, (b) segmentation and final classification with RF1 model, (c) segmentation and final classification with RF2 model post-landslide image.

Figure 11. Landslide detection for L2: (a) post-landslide optical image, (b) segmentation and final classification with RF1 model, (c) segmentation and final classification with RF2 model post-landslide image.

Figure 12. Landslide detection for L15 (center of the image): (a) post-landslide optical image, (b) segmentation and final classification with RF1 model, (c) segmentation and final classification with RF2 model post-landslide image.

Herrera Herrera M. 2019. Landslide Detection using Random Forest Classifier [MSc thesis]. Delft University of Technology.

 Google Scholar