An automatic rice mapping method based on an integrated time-series gradient boosting tree using GF-6 and sentinel-2 images

Figures & data

Figure 1. The location of the study areas and sample collection points. (a) ~ (c) represent study site 3, study site 2 and study site 1, respectively. The points on the map represent the sample collection points. (d) ~ (f) Photos of crops taken during the field survey, including SR, DR, soybean and corn.

Figure 2. Rice phenology and the time of acquisition of the GF6 and Sentinel–2 images: (a) study site 1; (b) study site 2; (c) study site 3. The middle rice in the figure represents SR, and the DR contains both early rice and late rice.

Table 1. The number of samples for the land cover types in the three study sites.

Land cover type	Xiangyin County	Jingzhou City	Tongjiang City
Single rice	2359	2713	3100
Double rice	2157	0	0
Corn	835	891	1022
Soybean	912	1069	1117
Water	2360	2230	2500
Forest	2040	2110	1100
Grassland	0	670	460
Building	2800	2400	2380
Sum	13463	12083	11679

Figure 3. Comparison between the original data and interval feature data of the NDRE for DR and SR. The interval features include the mean f₁(t₁,t₂), standard deviation f₂(t₁,t₂) and slope f₃(t₁,t₂). (a) is the original data and interval feature data of the NDRE for DR, while (b) is the original data and interval feature data of the NDRE for SR.

Figure 4. Schematic diagram of the GBT algorithm, where X represents the input features, and tree represents the decision tree. In this study, X refers to the interval features and Fourier coefficients of the time-series data.

Figure 5. The flowchart compares the effectiveness of the Auto–ITSGBT method proposed in this paper with competing methods including the TSRF method, DTW method, and SWV method for rice mapping.

Figure 6. Results of rice mapping based on four methods at study site 1: (a–b) the original image and the local area (GF–6 on June 7, 2021); (c–d) Auto-ITSGBT-based rice mapping result; (e–f) TSRF-based rice mapping result; (g–h) DTW-based rice mapping result; (i–j) SWV-based rice mapping result.

Table 2. Accuracy comparison of four rice mapping methods at study site 1.

Paddy rice map	Class	PA%	UA%	OA%	Kappa
Auto-ITSGBT	Double rice	96.19	96.31	96.45	0.95
	Single rice Non-rice	97.12 96.14	95.22 97.55
TSRF	Double rice	92.63	93.97	94.32	0.91
	Single rice Non-rice	95.25 94.98	93.27 95.44
DTW	Double rice	92.39	93.39	94.03	0.91
	Single rice Non-rice	95.13 94.50	93.38 95.05
SWV	Double rice	91.68	92.67	93.17	0.89
	Single rice Non-rice	94.88 93.05	92.34 94.23

Figure 7. Results of rice mapping based on four methods at study site 2: (a–b) the original image and the local area (GF–6 on August 17, 2021); (c) Auto–ITSGBT–based rice mapping result; (d) TSRF–based rice mapping result; (e) DTW–based rice mapping result; (f) SWV-based rice mapping result.

Figure 8. Results of rice mapping based on four methods at study site 3: (a–b) the original image and the local area (GF–6 on August 20, 2019); (c) Auto-ITSGBT-based rice mapping result; (d) TSRF-based rice mapping result; (e) DTW-based rice mapping result; (f) SWV-based rice mapping result.

Table 3. Accuracy comparison of the four rice mapping methods at study sites 2 and 3.

Study site	Paddy rice map	Class	PA%	UA%	OA%	Kappa
Study site 2 (Jingzhou City)	Auto-ITSGBT	Rice	96.99	97.16	97.13	0.94
		Non-rice	97.28	97.12
	TSRF	Rice	93.97	94.31	94.28	0.89
		Non-rice	94.57	94.25
	DTW	Rice	93.79	93.63	93.84	0.88
		Non-rice	93.89	94.05
	SWV	Rice	92.91	93.40	93.32	0.87
		Non-rice	93.72	93.24
Study site 3 (Tongjiang City)	Auto-ITSGBT	Rice	97.34	95.57	96.63	0.93
		Non-rice	95.98	97.59
	TSRF	Rice	94.69	94.27	94.79	0.90
		Non-rice	94.88	95.25
	DTW	Rice	92.92	93.66	93.70	0.87
		Non-rice	94.40	93.74
	SWV	Rice	92.30	92.87	93.04	0.86
		Non-rice	93.70	93.18

Figure 9. Variation in OA with a gradual increase in the number of GF–6 and Sentinel-2 images based on the Auto-ITSGBT method.

Figure 10. The harmonic curves at the frequency indices t = 0 to 4 obtained from GF–6–based NDRE curves of various crops by DFT.

Table 4. Fourier coefficients of the GF-6-based NDRE curves for different crops at spectral coefficients t = 0 ~ 5.

cc	t	Amplitude	Phase	Amplitude		Crop type	t	Amplitude	Phase	Amplitude
				Proportion %	Cumulative proportion %					Proportion %	Cumulative proportion %
DR	0	0.45				SR	0	0.36
	1	0.15	0.47	39.89	39.89		1	0.24	1.20	85.47	85.47
	2	0.10	−0.94	19.19	59.08		2	0.08	−1.47	10.39	95.86
	3	0.11	0.47	20.34	79.42		3	0.04	0.32	1.94	97.80
	4	0.07	0.10	9.22	88.64		4	0.03	0.60	1.65	99.45
	5	0.08	−1.34	11.36	100		5	0.02	−0.15	0.55	100
Corn	0	0.29				Soybean	0	0.32
	1	0.13	−1.02	75.93	75.93		1	0.17	−1.30	86.80	86.80
	2	0.07	0.07	18.97	94.9		2	0.06	−0.64	10.68	97.48
	3	0.03	1.41	3.26	98.16		3	0.01	0.92	0.09	97.57
	4	0.02	0.43	1.45	99.61		4	0.02	−1.21	0.86	98.43
	5	0.01	−0.79	0.39	100		5	0.02	0.23	1.57	100

Data availability statement

The data that support the findings of the study area are available from the first author [Xueqin Jiang, [email protected]] upon reasonable request.