Evaluation of neural network models for landslide susceptibility assessment

Figures & data

Figure 1. Study area: (a) and (b) geographical location of the study area; (c) shaded relief map of the study area showing the geomorphic settings of this region; and (d), (e), and (f) Google Earth images of three typical landslides.

Figure 2. Fifteen landslide conditioning factors are utilized in LSM.

Table 1. Data utilized in the present study.

Source	Format	Time	Scale	Factor
Landslides	Shapefile	2019	–	–
ASTER DEM (https://www.usgs.gov)	TIFF	2011	30 m	Slope
				Aspect
				Elevation
				Relief amplitude
				Plan curvature
				Profile curvature
				TWI
				SPI
Geological map	Shapefile	–	1:500,000	Lithology
				Distance to faults
Road map (http://www.webmap.cn)	Shapefile	2017	1:250,000	Distance to roads
River map (http://www.webmap.cn)	Shapefile	2017	1:250,000	Distance to rivers
LULC map (http://doi.org/10.5281/zenodo.4280923)	TIFF	2020	30 m	LULC
Landsat 8	TIFF	May, 2020	30 m	NDVI
Meteorological observation station data	Shapefile	1989–2019	–	Annual Precipitation

Table 2. Description of the lithology map and LULC map.

Legend number	Lithology	LULC
a	Granite	Rainfed cropland
b	Metavolcanics	Herbaceous cover
c	Slate, limestone and sandstone	Irrigated cropland
d	Schist and tuff	Open evergreen broadleaved forest
e	Diorite	Closed evergreen broadleaved forest
f	Dolomite, limestone and marble	Open deciduous broadleaved forest (0.15 < fc < 0.4)
g	Limestone, shale and siltstone	Closed deciduous broadleaved forest (fc > 0.4)
h		Open evergreen needle-leaved forest (0.15 < fc < 0.4)
i	Slate intercalated with tuff, metasandstone and pebbly sandstone	Closed evergreen needle-leaved forest (fc >0.4)
j	Phyllite, slate, metasandstone, marl and marble	Deciduous shrubland
k	Pyroclastic rock	Grassland
l	Siliceous rock, slate and phyllite	Sparse vegetation (fc < 0.15)
m	Dolomite, sandstone and shale	Wetlands
n	Sandstone and siltstone	Impervious surfaces
o	Shale, limestone and marl	Bare areas
p	Bioclastic limestone	Consolidated bare areas
q	carbonate rock and clastic	Water body
r	Mudstone and siltstone	Permanent ice and snow

Figure 3. Workflow of the present study.

Figure 4. Architecture of a typical three-layer ANN. @15 represents the neurons in this layer are 15.

Figure 5. Architecture of 1D CNN.

Figure 6. Architecture of a typical RNN.

Figure 7. (a) Multicollinearity and (b) relative importance analysis of landslide conditioning factors.

Figure 8. Selection of optimal model structure parameters for ANN, 1D CNN and RNN using tenfold cross-validation with the stratified sampling method.

Table 3. Training hyper-parameter setting in this study.

Hyper-parameters	ANN	1D CNN	RNN	Optional parameters
Optimizer	SGD	SGD	SGD	{SGD, Adam, Adagrad, Adadelta}
Activation	ReLU	Tanh	Tanh	{Tanh, ReLU, Sigmoid}
Batch size	120	40	80	{40, 80, 120, 160, 200, 240, 280}
Initial learning rate	0.001	0.001	0.001	{0.001–1 × 10^−9}
Maximum number of epochs	500	500	500	{0–1000}

Figure 9. Landslide susceptibility maps generated by three neural network models.

Table 4. Statistical analysis of different landslide susceptibility maps.

Models	Susceptibility levels	Area (km^2)	Percentage of area	Number of landslides	Percentage of number	Density
ANN	Very low	911.55	28.30%	2	0.66%	0.06
	Low	597.46	18.55%	22	7.24%	1.08
	Moderate	536.94	16.67%	50	16.45%	2.74
	High	549.05	17.05%	66	21.71%	3.54
	Very high	625.92	19.43%	164	53.95%	7.71
1D CNN	Very low	1163.78	36.13%	5	1.64%	0.13
	Low	562.87	17.48%	14	4.61%	0.73
	Moderate	500.81	15.55%	38	12.50%	2.23
	High	493.57	15.32%	67	22.04%	3.99
	Very high	499.87	15.52%	180	59.21%	10.59
RNN	Very low	1742.20	54.09%	12	3.95%	0.20
	Low	290.94	9.03%	13	4.28%	1.31
	Moderate	204.19	6.34%	17	5.59%	2.45
	High	222.82	6.92%	43	14.14%	5.68
	Very high	760.76	23.62%	219	72.04%	8.47

Figure 10. Comparison of the performance of neural network models: (a) ROC curves and (b) histogram of all accuracy evaluation measures.

Table 5. Wilcoxon rank-sum statistical test (two-tailed) between different neural network models.

Pairwise	Statistic	P value	Significance
ANN ∼ 1D CNN	502.2134	0.0000	Yes
ANN ∼ RNN	693.4044	0.0000	Yes
1D CNN ∼RNN	463.0361	0.0000	Yes

Figure 11. Predictive performance of different neural network models with different training hyper-parameter settings.

Table 6. Comparison of model complexity.

Model	Parameters	Training time (ms/epoch)	Inference time (s)
ANN	244	5	7.038
1D CNN	60342	7	17.184
RNN	6722	10	14.058

Data availability statement

The data or code used in this study are available from the corresponding author on reasonable request.