Improving wetland cover classification using artificial neural networks with ensemble techniques

Figures & data

Figure 1. (a) Location of the Zoige wetland in China, (b) the boundary of the study area, (c) the true color composite image of the Landsat 8 OLI data of the whole study area

Table 1. List of input features used for wetland cover classification. NDVI = normalized difference vegetation index, NDWI = normalized difference water index, MEAN = mean, HOMO = homogeneity, ASM = angular second moment, DIS = dissimilarity, ENTR = entropy, VAR = variance

Feature types	Features	Equation
Spectral features	Coastal, blue, green, red, NIR, SWIR1, SWIR2	-
Normalized difference ratios	NDVI	$\frac{NIR-red}{NIR+red}$
	NDWI	$\frac{green-NIR}{green+NIR}$
Textural features	MEAN	$\sum _{i=0}^{N-1}\sum _{j=0}^{N-1}i\times P\left(i,j\right)$
	HOMO	$\sum _{i=0}^{N-1}\sum _{j=0}^{N-1}\frac{P\left(i,j\right)}{1+{\left(i-j\right)}^2}$
	ASM	$\sum _{i=0}^{N-1}\sum _{j=0}^{N-1}P{\left(i,j\right)}^2$
	DIS	$\sum _{i=0}^{N-1}\sum _{j=0}^{N-1}P\left(i,j\right)\times \left|i-j\right|$
	ENTR	$-\sum _{i=0}^{N-1}\sum _{j=0}^{N-1}P\left(i,j\right)\times \mathrm{l}\mathrm{n}\left(P\left(i,j\right)\right)$
	VAR	$\sum _{i=0}^{N-1}\sum _{j=0}^{N-1}P\left(i,j\right)\times {\left(i-\mathrm{M}\mathrm{E}\mathrm{A}\mathrm{N}\right)}^2$
Topographical factors	Elevation	-

$N$ is the number of the gray levels; $P\left(i,j\right)$ is the normalized gray level value of the cell $\left(i,j\right)$ of the GLCM.

Figure 2. Locations of reference samples in the study area and typical examples of wetland covers in the Landsat 8 OLI images. AM = alpine meadow, MM = marsh meadow, M = marsh, RL = river and lake, FP = floodplain, S = sediments, RA = residential area, BR = bedrock

Table 2. Descriptions of wetland cover categories used in this study

Class name	Class code	Class description
Alpine meadow	AM	A vegetation community dominated by mesophytic Kobresia species
Marsh meadow	MM	Mixed areas of alpine meadow and marsh; characterized by rough texture
Marsh	M	Wet areas mainly consisting of emergent and submerged graminoid vegetation; displayed in deep violet under true color space
River and lake	RL	Permanent and seasonal water areas with the slim or elliptical shape; displayed in pink under true color space
Floodplain	FP	Periodically inundated areas; characterized by rough texture
Sediments	S	Sand land or deposits with high brightness
Residential area	RA	Regular shape with high brightness; usually distributed in low-elevation valleys
Bedrock	BR	Irregular shape with high brightness; usually exposed in ridges

Figure 3. The flowchart of wetland cover classification. ANN = artificial neural network, MBANN = MultiBoost artificial neural network, RANN = rotation artificial neural network, VGG = visual geometry group, RF = random forests

Figure 4. Illustration of the MBANN method. ANN = artificial neural network, MBANN = MultiBoost artificial neural network, AB = AdaBoost

Figure 5. Illustration of the RANN method. ANN = artificial neural network, RANN = rotation artificial neural network, PCA = principal component analysis

Table 3. Classification accuracy evaluation and comparison of various classification methods. ANN = artificial neural network, MBANN = MultiBoost artificial neural network, RANN = rotation artificial neural network, VGG11 = visual geometry group, RF = random forests. AM = alpine meadow, MM = marsh meadow, M = marsh, RL = river and lake, FP = floodplain, S = sediments, RA = residential area, BR = bedrock

Classes	ANN			MBANN			RANN			VGG11			RF
	UA	PA	F	UA	PA	F	UA	PA	F	UA	PA	F	UA	PA	F
AM	0.930	0.930	0.930	0.964	0.925	0.944	0.969	0.950	0.960	0.939	0.967	0.953	0.935	0.935	0.935
MM	0.811	0.888	0.848	0.826	0.919	0.870	0.872	0.938	0.904	0.944	0.958	0.951	0.870	0.919	0.894
M	0.975	0.975	0.975	0.976	1.000	0.988	0.988	1.000	0.994	0.990	1.000	0.995	0.949	0.938	0.943
RL	0.955	0.960	0.958	0.995	0.990	0.992	1.000	1.000	1.000	0.923	0.904	0.914	0.961	0.975	0.968
FP	0.930	0.887	0.908	0.957	0.887	0.920	0.957	0.900	0.928	0.934	0.867	0.899	0.937	0.887	0.911
S	0.905	1.000	0.950	0.927	1.000	0.962	0.987	1.000	0.993	0.946	0.957	0.951	0.923	0.947	0.935
RA	0.971	0.708	0.819	0.951	0.813	0.876	1.000	0.958	0.979	0.823	0.879	0.850	0.915	0.896	0.905
BR	0.902	0.822	0.860	0.978	1.000	0.989	0.978	1.000	0.989	0.930	0.964	0.946	1.000	0.911	0.953
Overall evaluation	WF	OA	K	WF	OA	K	WF	OA	K	WF	OA	K	WF	OA	K
	0.915	0.916	0.900	0.942	0.942	0.931	0.962	0.961	0.954	0.934	0.934	0.922	0.931	0.931	0.918

Table 4. Z-score test with associated probability value (P) for model pairs. ANN = artificial neural network, MBANN = MultiBoost artificial neural network, RANN = rotation artificial neural network, VGG11 = visual geometry group, RF = random forests

Comparative pairs	z score	P value	Significance level
ANN vs. MBANN	3.273	0.001	Yes
ANN vs. RANN	4.041	0.000	Yes
ANN vs. RF	0.962	0.336	No
ANN vs. VGG11	0.283	0.777	No
MBANN vs. RANN	2.121	0.034	Yes
MBANN vs. RF	2.500	0.012	Yes
MBANN vs. VGG11	2.082	0.037	Yes
RANN vs. VGG11	3.413	0.001	Yes
RANN vs. RF	3.771	0.000	Yes

Figure 6. Wetland cover maps obtained by the five classifiers: (a) artificial neural network, (b) MultiBoost artificial neural network, (c) rotation artificial neural network, (d) visual geometry group, and (e) random forests

Figure 7. Area percentages of wetland cover classes predicted by different classification methods. ANN = artificial neural network, MBANN = MultiBoost artificial neural network, RANN = rotation artificial neural network, VGG11 = visual geometry group, RF = random forests

Figure 8. Robustness evaluation concerning the impact of the data size on the classification accuracy of classifiers. RANN = rotation artificial neural network, MBANN = MultiBoost artificial neural network, ANN = artificial neural network, RF = random forests, VGG11 = visual geometry group

Figure 9. Robustness evaluation concerning the impact of the data noise on the classification accuracy of classifiers. RANN = rotation artificial neural network, MBANN = MultiBoost artificial neural network, ANN = artificial neural network, RF = random forests, VGG11 = visual geometry group

Figure 10. Robustness evaluation concerning the impact of the feature variation on the classification accuracy of classifiers. RANN = rotation artificial neural network, MBANN = MultiBoost artificial neural network, ANN = artificial neural network, RF = random forests