Object-based image analysis of suburban landscapes using Landsat-8 imagery

Figures & data

Figure 1. The study area located in Beijing, China.

Table 1. Object features used for classification.

Feature Type	Object Features	Description
Spectral	mean value^a, standard deviation^a, ratio^a, min pixel value^a, max pixel value^a, brightness, max difference, NDVI, NDBI, NDWI	Spetral features are used to describe image objects with information derived from their spectral properties.
Geometrical	area, length, length-to-width ratio, number of pixels, width, asymmetry, border index, compactness, density, elliptic fit, radius of largest enclosed ellipse, radius of smallest enclosing ellipse, rectangular fit, roundness, shape index	Geometrical features are built on an object’s shape and calculated from the pixels that form it.
Textural	GLCM homogeneity^a, GLCM contrast^a, GLCM dissimilarity^a, GLCM entropy^a, GLCM ang. 2nd moment^a, GLCM mean^a, GLCM std. dev^a, GLCM correlation^a	Textural features originate from Haralick, Shanmugam, and Dinstein (1973) based on the Gray-Level Co-occurrence Matrix.

^aObject features were calculated for each of the 8 multispectral bands mentioned above.

Figure 2. Importance of the features based on the ReliefF method and Cfs results at the segmentation scales of 30 and 60. (Here, a checkmark means that the feature is included in the optimal subset).

Figure 3. Overall accuracy of the four classifiers with different combinations of shape and compactness at segmentation scales of 30 and 60.

Table 2. Values of shape and compactness corresponding to Max A and Min A with different classifiers.

Classifiers	Max OA				Min OA
	Segmentation Scale30		Segmentation Scale60		Segmentation Scale30		Segmentation Scale60
	S	C	S	C	S	C	S	C
DT	0.8	0.8	0.9	0.9	0	0.6	0.6	0.5
SVM	0.7	0.4	0.2	1	0.2	0.8	0.8	0
RT	0.9	0.3	0\0.5	1	0.1	1	0.9	0.1
NB	0.8	0.2	0.5	1	0.1	1	0.9	0

Note: Max OA: maximum average overall accuracy; Min OA: minimum average overall accuracy; S: shape; and C: compactness.

Figure 4. Number of objects at segmentation scales of 30 and 60 with different combinations of shape and compactness.

Figure 5. Overall accuracy of the four classifiers with increasing segmentation scale size.

Table 3. Z-statistics of the classification results.

Pair comparsion	SVM vs RT	SVM vs NB	SVM vs DT	RT vs NB	RT vs DT	NB vs DT
Z-statistics	0.7681	2.1832^a	3.5664^a	1.4343	2.8483^a	1.4427

^aSignificant difference at the 95% confidence level.

Figure 6. Overall accuracy of a decision tree with increasing maximum depth values at different segmentation scale sizes.

Figure 7. Overall accuracy of a support vector machine with different values of C and gamma at different segmentation scale sizes.

Figure 8. Overall accuracy of a random tree with different values of k and n at different segmentation scale sizes.

Table 4. Optimal parameter values for three classifiers at different segmentation scale sizes.

Scale	DT	SVM	RT
	Maximum depth	(Gamma, C)	(n, k)
10	8	(10^−2,10^2)	(2,50)
20	8/9	(10^−2,10^2)	(1,40)
30	5	(10^−4,10^4)/(10^−5,10^5)	(4,90)
40	5/13	(10^−2,10^2)	(1,60)
50	7	(10^−2,10^2)	(2,70)
60	5/7-10/12-13/15	(10^−210^2)/(10^−2,10^3)	(2,50)
70	5	(10^−210^2)	(2,30)
80	6/8-15	(10^−4,10^6)	(1,40)/(2,80)/(4,80)/(15,50)/(15,100)

Haralick, Robert M., K. Shanmugam, and I. Dinstein. 1973. “Textural Features for Image Classification.” IEEE Transactions on System, Man and Cybernetics SMC-3 (6): 610–621. doi: 10.1109/TSMC.1973.4309314

 Web of Science ®Google Scholar