Object-wise joint-classification change detection for remote sensing images based on entropy query-by fuzzy ARTMAP

Figures & data

Figure 1. The architecture of the FAM.

Figure 2. Object-based entropy query-by fuzzy ARTMAP joint classification change detection diagram.

Figure 3. Study area in Liaohe estuarine wetland.

Table 1. Detailed information of remote sensing images used in the experiments.

Year	Data	Satellite	Sensor
1987	21 September	Landsat 5	TM
1989	22 June	Landsat 5	TM
1995	11 September	Landsat 5	TM
2000	22 July	Landsat 5	TM
2004	19 August	Landsat 5	TM
2009	3 October	Landsat 5	TM
2013	23 May	Landsat 8	OLI
2014	26 May	Landsat 8	OLI

Figure 4. True color remote sensing images of the study area captured in 1987, 1989, 1995, 2000, 2004, 2009, 2013, and 2014.

Figure 5. Comparison of classification accuracy in 1989 remote sensing images with different number of initial training samples, (a) the initial number of training samples is 60, (b) the initial number of training samples is 70, (c) the initial number of training samples is 80, (d) the initial number of training samples is 90.

Figure 6. The process of pixel-based and object-based joint classification comparison in the sub-study area.

Table 2. The parameters of the FAM neural network.

					k	pct
0.001	1.000	0.010	0.900	0.990	4	0.6

Figure 7. Comparison of classification result for different methods between 1989 and 2013 remote sensing images. (a) The original false color images, (b) FAM-PCC with average accuracy of 86.93%, (c) K-means-JCC with average accuracy of 88.89%, (d) FAM-JCC with average accuracy of 90.52%, (e) EQFAM-JCC with average accuracy of 91.98%, (f) EQFAM-OBJCC with average accuracy of 94.12%.

Figure 8. Comparison of change detection results for different methods between 1989 and 2013 remote sensing images. (a) FAM-PCC with total error of 39.38%, (b) K-means-JCC with total error of 37.78%, (c) FAM-JCC with total error of 32.04%, (d) EQAM-JCC with total error of 30.94%, “(e) EQFAM-JCC (f) EQFAM-OBJCC with total error of 27.03%.

Table 3. Comparison of change detection results of five different methods.

	Misdetection rate (%)	False alarm rate (%)	Total error (%)	Average accuracy (%)	Kappa coefficient
FAM-PCC	11.10	27.27	38.38	86.93	0.7043
K-means-JCC	18.89	18.89	37.78	88.89	0.7324
FAM-JCC	15.82	16.48	32.04	90.52	0.7725
EQAM-JCC	18.89	12.05	30.94	91.18	0.7826
EQFAM-OBJCC	15.75	11.28	27.03	94.12	0.8233

Figure 9. Land use/cover classification results of Liaohe estuarine wetland in 1987, 1989, 1995, 2000, 2004, 2009, 2013, and 2014.

Table 4. Statistical area of reclamation activity (km²) 1987–2014.

	1987	1989	1995	2000	2004	2009	2013	2014
Land area	872.465	936.515	1017.634	1035.392	1041.265	1091.336	1157.110	1194.937
Reclamation area	41.858	105.908	187.026	204.785	210.658	260.728	326.503	364.330

Figure 10. Reclamation monitoring results during 1987–1989, 1989–1995, 1995–2000, 2000–2004, 2004–2009, 2009–2013, 2013–2014, and 1987–2014.

Figure 11. Area change of reclamation from 1987 to 2014 (km²).