Fine classification of crops based on an inductive transfer learning method with compact polarimetric SAR images

Figures & data

Figure 1. The color composite images (FP SAR VV (red), VH (green), and HH (blue)) of the backscattering coefficients of FP SAR data (a: July 27, 2012; b: September 16, 2015).

Figure 2. The color composite images of Pauli decomposition of FP SAR data (a, b, c, and d are SAR images of regions a, b, c, and d of study area 2 on March 6, 2017, respectively. e is SAR images of small region e on May 22, 2017. f is the distribution map of the five small regions in study area 2).

Figure 3. The color composite images of Pauli decomposition of FP SAR data (a: May 28, 2014; b: August 18, 2014; c: ground-truth data (blue, brown, red, green, and yellow plots are barley, corn, potato, sugar beet, and wheat, respectively).

Table 1. Full-polarization SAR data parameters of multitemporal RADARSAT-2 C band data for study area 1.

Data acquisition (D/M/Y)	DoY (Day of Year)	Image mode	Pixel spacing (A × R, m)	Incidence angle (°)	Rice phenology stage
27/06/2012	179	FQ20W^1	5.2×7.6	38–41	Seedling
21/07/2012	203	FQ20W	5.2×7.6	38–41	Seedling-Elongation
28/08/2012	241	FQ9W	5.2×7.6	27–30	Booting-Heading
21/09/2012	265	FQ9W	5.2×7.6	27–30	Heading-Flowering
15/10/2012	289	FQ9W	5.2×7.6	27–30	Dough-Mature
08/11/2012	312	FQ9W	5.2×7.6	27–30	Harvest
12/06/2015	163	FQ20W	5.2×7.6	38–41	Seedling
30/07/2015	211	FQ20W	5.2×7.6	38–41	Seedling-Elongation
23/08/2015	235	FQ20W	5.2×7.6	38–41	Booting-Heading
16/09/2015	259	FQ20W	5.2×7.6	38–41	Heading-Flowering
10/10/2015	283	FQ20W	5.2×7.6	38–41	Dough-Mature
03/11/2015	307	FQ20W	5.2×7.6	38–41	Harvest

^aFQW = fine-quad-polarimetry wide, where 20 or 9 is the number of beam positions, which is related to the incidence angle.

Table 2. Full-polarization SAR data parameters of multitemporal GF-3 C band data for study area 2.

Small regions	Data acquisition (D/M/Y)	Image mode	Pixel spacing (A×R, m)	Incidence angle (°)	Wheat phenology stage
Small region a	06/03/2017 06/03/2017 06/03/2017 06/03/2017	QPSI^2	5.0×4.5	48.79	Regreening- elongation
Small region b		QPSI	5.0×4.5	48.79	Regreening- elongation
Small region c		QPSI	5.0×4.5	48.79	Regreening- elongation
Small region d		QPSI	5.0×4.5	48.79	Regreening- elongation
Small region e	22/05/2017	QPSI	2.2×5.2	34.63	Mature

^aQPS = Quad Polarimetry Stripe.

Table 3. Full-polarization SAR data parameters of multitemporal RADARSAT-2 C band data for study area 3.

Data acquisition (D/M/Y)	DoY (Day of Year)	Image mode	Pixel spacing (A×R, m)	Incidence angle (°)	Phenology stage
28/05/2014	148	FQ20W	5.2×7.7	40.2	Seedling stage
21/06/2014	172	FQ20W	5.2×7.7	40.2	Early growth stage
15/07/2014	196	FQ20W	5.2×7.7	40.2	Late growth stage
18/08/2014	229	FQ20W	5.2×7.7	40.2	Mature stage

Figure 4. Methodology flowchart.

Table 4. Transfer experiments in study area 1.

Transfer types	Experiments	Date of DS data	Date of DT data
Cross-year transfer under the same phenological periods	(a)	27/06/2012	12/06/2015
	(b)	21/07/2012	30/07/2015
	(c)	28/08/2012	23/08/2015
	(d)	21/09/2012	16/09/2015
	(e)	15/10/2012	10/10/2015
	(f)	08/11/2012	03/11/2015
Cross-year transfer under the different phenological periods	(g)	27/06/2012	10/10/2015
	(h)	21/09/2012	10/10/2015
	(i)	21/09/2012	03/11/2015

Table 5. Transfer experiments in study area 2.

Transfer types	Experiments	Region of DS data	Region of DT data
Transfer under the same phenological periods of similar scenes	(a)	Small region a	Small region b
	(b)	Small region a	Small region c
	(c)	Small region a	Small region d
Transfer under the different phenological periods of similar scenes	(d)	Small region a	Small region e

Table 6. Transfer experiments in study area 3.

Experiments	Date of DS data	Date of DT data
(a)	28/05/2014	21/06/2014
(b)	28/05/2014	15/07/2014
(c)	28/05/2014	18/08/2014
(d)	18/08/2014	28/05/2014
(e)	18/08/2014	21/06/2014
(f)	18/08/2014	15/07/2014

Table

The proposed TBEL algorithm
1 Select an appropriate weak classifier, preparing DS,DTL, and DTU.
2 The training samplesubset is generated by dividing the whole labeled data (DSUDTL)into Tgroups.
3 Implement the TrBagg algorithm.
4 Obtain F* by filtering F, and F*is used as the judging set H.
5 For n =1,2,3…Ndo
6 Classify DTUby using classifiers in the judging set H.
7 If n =1then
8 Add the data that is completely consistent with the DTUprediction classes to DTL.
9 else
10 Compare the prediction ofH to DTUwith the prediction of ft-1 to DTU, and add the data that is completely consistent with DTL.
11 end if
12 Use DTL to train a weak classifier ft and add the weak classifier to weak classifier set F.
13 If DTL no longer changes then
14 break
15 end if
16 end for
17 Obtain weak classifier set FT.

Figure 5. Fine classification maps of two types of rice based on three transfer learning algorithms (TrBagg, BETL, and TBEL) in six groups of transfer experiments (a, b, c, d, e, and f are the transfer experiments in six groups of cross-year domains under the same phenological periods, and serial 1, 2, and 3 are the fine classifications of rice results based on the TrBagg, BETL, and TBEL algorithms, respectively). gt is the ground-truth area of five kinds of ground objects).

Figure 6. Fine classification maps of two types of rice based on three transfer learning algorithms (TrBagg, BETL, and TBEL) in three transfer experiments (g, h, and i are three groups of cross-year domains under different phenological periods).

Table 7. Transfer learning-based method classification overall accuracy (OA) table with a decision tree as a weak classifier.

Transfer learning algorithms	Transfer experiments (OA/%)
	a	b	c	d	e	f	g	h	i
SA (Fernando et al. 2014)	32.5	26.5	41.8	40.5	41.7	59.8	14.2	57.8	33.9
CORAL (Sun, Feng, and Saenko 2016)	33.2	57.6	54.8	48.2	48.9	52.7	41.7	43.5	42.2
JDA (Long et al. 2013)	41.2	44.5	64.7	67.7	38.4	46.9	39.9	69.0	47.6
BDA (Wang et al. 2017)	33.5	54.1	46.9	26.0	32.6	33.6	57.2	29.3	42.1
TrBagg (Kamishima, Hamasaki, and Akaho 2009)	62.4	75.4	71.5	81.3	75.2	88.2	73.7	77.2	81.1
BETL (X. Liu et al. 2016)	62.8	74.1	75.6	84.8	79.7	87.8	73.9	78.9	89.6
TBEL	67.3	76.3	71.6	86.5	81.0	92.2	79.6	78.2	92.7
Supervised classification	67.0	79.6	75.5	85.8	89.5	94.0	89.5	89.5	94.0

Table 8. Transfer learning-based method classification OA table with K-nearest neighbor classifier as a weak classifier.

Transfer learning algorithms	Transfer experiments (OA/%)
	a	b	c	d	e	f	g	h	i
SA	58.8	66.8	61.6	62.6	28.1	65.7	26.5	56.9	61.6
CORAL	54.2	63.2	54.6	48.6	27.9	52.3	53.7	48.4	49.0
JDA	46.0	46.0	65.6	72.6	36.3	47.5	41.8	70.6	55.0
BDA	45.4	46.3	64.5	75.4	36.6	47.4	40.2	71.2	55.9
TrBagg	62.6	75.7	70.7	82.7	79.6	88.5	70.8	80.8	90.9
BETL	61.0	75.2	74.7	85.2	81.8	90.7	72.3	82.3	89.8
TBEL	63.4	75.9	72.0	86.0	82.6	93.9	81.5	82.4	93.5
Supervised classification	66.6	80.7	76.9	86.5	83.2	94.5	83.2	83.2	94.5

Table 9. Transfer learning-based method classification OA table with random forest as a weak classifier.

Transfer learning algorithms	Transfer experiments (OA/%)
	a	b	c	d	e	f	g	h	i
SA	34.7	33.8	37.1	55.1	49.7	72.8	31.6	57.8	59.2
CORAL	35.0	57.8	52.7	46.4	45.3	48.7	47.0	43.6	42.8
JDA	40.0	48.8	67.5	71.5	37.7	49.6	38.1	72.2	59.5
BDA	30.8	26.6	44.4	10.3	22.6	20.2	51.5	16.1	24.8
TrBagg	64.8	77.1	72.4	83.6	81.3	93.5	73.8	81.8	90.1
BETL	64.3	76.8	72.5	86.0	83.2	93.5	77.6	82.6	91.5
TBEL	65.9	78.0	76.6	86.5	84.5	95.9	82.1	85.2	93.4
Supervised classification	70.2	84.0	80.7	90.8	86.9	94.5	86.9	86.9	94.5

Figure 7. The TBEL transfer learning-based method classification accuracy with different weak classifiers.

Figure 8. Classification maps of wheat based on the TrBagg, BETL, and TBEL algorithms under four groups of transfer experiments (a, b, and c are transfer experiments under the same phenological periods of similar scenes. d is the transfer experiment under different phenological periods of similar scenes, and serial 1, 2, and 3 are the classification results based on the TrBagg, BETL, and TBEL algorithms, respectively).

Table 10. Transfer learning-based method classification overall accuracy (OA) table with a decision tree as a weak classifier.

Transfer learning algorithms	Transfer experiments
	a		b		c		d
	OA (%)	Kappa	OA (%)	Kappa	OA (%)	Kappa	OA (%)	Kappa
SA	49.7	0.242	66.2	0.481	37.6	0.046	49.7	0.255
CORAL	66.3	0.498	67.0	0.500	86.4	0.795	73.1	0.590
JDA	65.7	0.480	80.8	0.707	81.6	0.723	83.5	0.753
BDA	56.1	0.336	77.0	0.648	59.6	0.382	58.4	0.386
TrBagg	88.4	0.807	92.1	0.877	84.0	0.739	86.6	0.799
BETL	84.6	0.745	89.8	0.842	77.2	0.630	88.1	0.821
TBEL	90.8	0.846	92.9	0.881	87.6	0.801	94.4	0.944
Supervised classification	91.9	0.863	94.0	0.905	88.8	0.821	93.5	0.901

Table 11. Transfer learning-based method classification OA table with K-nearest neighbor classifier as a weak classifier.

Transfer learning algorithms	Transfer experiments
	a		b		c		d
	OA (%)	Kappa	OA (%)	Kappa	OA (%)	Kappa	OA (%)	Kappa
SA	77.0	0.656	65.8	0.476	69.5	0.528	80.3	0.701
CORAL	71.4	0.573	60.3	0.391	84.8	0.772	76.6	0.645
JDA	65.5	0.478	80.4	0.700	81.7	0.725	80.4	0.708
BDA	64.7	0.465	80.2	0.698	81.3	0.719	79.5	0.694
TrBagg	89.3	0.821	92.4	0.881	83.8	0.736	89.2	0.838
BETL	85.1	0.754	90.5	0.853	81.7	0.705	89.1	0.836
TBEL	90.3	0.838	93.0	0.882	87.4	0.798	93.6	0.903
Supervised classification	89.5	0.825	94.3	0.910	86.5	0.786	93.8	0.905

Table 12. Transfer learning-based method classification OA table with random forest as a weak classifier.

Transfer learning algorithms	Transfer experiments
	a		b		c		d
	OA (%)	Kappa	OA (%)	Kappa	OA (%)	Kappa	OA (%)	Kappa
SA	76.5	0.649	66.4	0.484	54.4	0.299	62.6	0.447
CORAL	65.5	0.486	66.8	0.496	89.0	0.833	72.9	0.588
JDA	66.0	0.485	68.8	0.503	84.9	0.773	65.5	0.488
BDA	37.6	0.046	68.1	0.508	31.9	0.041	58.4	0.387
TrBagg	88.7	0.813	92.6	0.885	84.5	0.749	87.6	0.814
BETL	85.7	0.762	92.4	0.882	83.9	0.741	90.6	0.858
TBEL	90.9	0.847	92.3	0.872	88.3	0.812	94.4	0.915
Supervised classification	91.9	0.864	94.0	0.906	88.4	0.815	92.2	0.882

Figure 9. The TBEL transfer learning-based method classification accuracy with different weak classifiers.

Figure 10. Fine classification maps of crops based on three transfer algorithms (TrBagg, BETL, and TBEL) using CP SAR of CTLR mode (a, b, c, d, e, and f are six transfer experiments, and a1-f1, a2-f2, and a3-f3 are the transfer classification results based on the TrBagg, BETL, and TBEL algorithms, respectively).

Table 13. Transfer classification accuracy based on three transfer algorithms.

Transfer Experiment	TrBagg		BELT		TBEL
	OA (%)	Kappa	OA (%)	Kappa	OA (%)	Kappa
a	54.7	0.404	59.9	0.471	61.5	0.490
b	42.3	0.253	40.9	0.242	42.5	0.253
c	49.9	0.364	49.4	0.381	51.4	0.399
d	64.2	0.528	62.8	0.512	64.3	0.528
e	60.1	0.474	59.9	0.473	62.1	0.494
f	42.2	0.251	41.5	0.252	42.6	0.257

Figure 11. Transfer classification accuracy of crops based on the TBEL algorithm using CP SAR of CTLR and π/4 modes.

Fernando, B., A. Habrard, M. Sebban, and T. Tuytelaars. 2014. “Salient and Non-Salient Fiducial Detection Using a Probabilistic Graphical Model.” Computer Vision & Pattern Recognition 47 (1). https://doi.org/10.48550/arXiv.1409.5241.

 Google Scholar

Sun, B., J. Feng, and K. Saenko. 2016. “Return of Frustratingly Easy Domain Adaptation.” Computer Vision & Pattern Recognition 30 (1). https://doi.org/10.48550/arXiv.1511.05547.

 Google Scholar

Long, M., J. Wang, G. Ding,Sun, J. and Yu, P.S. 2013. “Transfer Feature Learning with Joint Distribution Adaptation.” Paper presented at the IEEE international conference on computer vision. Sydney, NSW, Australia, December 01-08. https://doi.org/10.1109/ICCV.2013.274.

 Google Scholar

Wang, J., Y. Chen, S. Hao, W. Feng, and Z. Shen, 2017. “Balanced Distribution Adaptation for Transfer Learning.” Paper presented at the IEEE International Conference on Data Mining. New Orleans, LA, USA, November 18-21. https://doi.org/10.1109/ICDM.2017.150.

 Google Scholar

Kamishima, T., M. Hamasaki, and S. Akaho, 2009. “TrBagg: A Simple Transfer Learning Method and Its Application to Personalization in Collaborative Tagging.” Paper presented at the Ninth IEEE International Conference on Data Mining, FL, USA, December 6–9. https://doi.org/10.1109/ICDM.2009.9.

 Google Scholar

Liu, X., G. Wang, Z. Cai, and H. Zhang. 2016. “Bagging Based Ensemble Transfer Learning.” Journal of Ambient Intelligence and Humanized Computing 7 (1): 29–36. https://doi.org/10.1007/s12652-015-0296-5.

 Web of Science ®Google Scholar

Data availability statement

The data that support the findings of this study are available upon reasonable request from the first author [Xianyu Guo. E-mall: http://[email protected]].