Single-Season Rice Area Mapping by Combining Multi-Temporal Polarization Decomposition Components and the Two-Stage Segmentation Method

Abstract

Recently, Synthetic Aperture Radar (SAR) data, especially Sentinel-1 data, have been increasingly used in rice mapping research. However, current studies usually use long time series data as the data source to represent the differences between rice and other ground objects, especially other crops, which results in complex models and large computational complexity during classification. To address this problem, a novel method for single season rice mapping is proposed, based on the principle that the scattering mechanism of rice paddies in the early flooding period is strongly influenced by water bodies, causing the volume scattering to be lower than that of other crops. Thus, a feature combination that can effectively and stably extract rice planting areas was constructed by combining multi-temporal volume scattering in the early flooding period of rice using dual-polarization SAR data, so that a simple semantic segmentation model could realize high-precision rice mapping tasks. A two-stage segmentation structure was introduced to further improve the mapping result with the Omni-dimensional Dynamic Convolution Residual Segmentation model (ODCRS model) as the bone model. In the experiment, Suihua City, Heilongjiang Province was selected as the study site, and the VH/VV polarized data of Sentinel-1 satellite in 2022 was used as the data source. The mapping accuracy of the ODCRS model was 88.70%, and the user accuracy was 84.19% on the field survey data. Furthermore, experiments with different years and regions also proved the effectiveness and stability of the proposed method.

1. Introduction

As the world’s fourth largest cereal crop [1], rice meets the staple food needs of more than half of the world’s population and is also a key crop for global food security [2]. Monitoring rice planting areas is of great importance in reducing hunger, promoting food security, improving nutrition, and promoting sustainable agriculture, further achieving Goal 2 of the 17 Sustainable Development Goals (SDGs) proposed by the United Nations in 2015: zero hunger [3]. China is the world’s largest rice producer. Except for a few provinces in the southwest region, most of the rice-growing provinces are dominated by single-season rice [4]. Therefore, it is of great importance to study the mapping method of single-season rice.

In recent years, SAR data has played an important role in the monitoring of cropland because of its ability to operate all day and in all weather conditions [5]. It has also been increasingly used in rice mapping research, where all bands of SAR data have been. For example, L-band (1–2 GHz) ALOS PALSAR data were introduced in rice field detection by [6]. Multi-temporal C-band (4–8 GHz) RADASAT-2 data were used by [7] to achieve high-precision fine classification of rice fields. The time series backscatter coefficient data of TerraSAR in the X-band (8–12 GHz) from May to October were used by Nelson et al. to monitor rice in South and Southeast Asia [8].

Current rice mapping methods generally describe the differences between rice growth curves and other ground objects (mainly vegetation) using long time series data. Sun et al. [9] used a feature combination sequence calculated from VH- and VV-polarized SAR data as inputs for the dual-branch Bidirectional Long Short-Term Memory (BiLSTM) model to achieve rice mapping in subtropical mountainous areas. Zhu et al. [10] combined multiple temporal remote sensing index data obtained from optical images and Long Short-Term Memory (LSTM) models to achieve rice field distribution mapping. Yang et al. [11] used temporal optical data together with dual polarized SAR data as data sources to achieve rice mapping with Simple Non-Iterative Clustering (SNIC) segmentation. Whether it is individual SAR data, individual optical data, or a fusion of both, when applied to rice mapping research, many studies directly use long time series data as classification input to achieve better mapping results. However, the use of long-term data can lead to a significant increase in computational complexity, as well as a significant increase in memory and time consumption during the mapping process.

To avoid the use of long-term data, this study depicts the differences between rice and other vegetation/crops from a different perspective. At present, most of the research on rice mapping using SAR data only uses the backscatter coefficients of time series, and the application of polarization information has not been sufficiently explored. Polarization decomposition methods can reveal the deep scattering mechanism of ground objects, while utilizing the intensity and phase information of ground object scattering. There are already many applications of polarization decomposition methods in other remote sensing fields, such as cropland extraction [12], forest biomass estimation [13], crop growth descriptor estimation [14], etc. The polarization decomposition method has some applications in rice monitoring, such as rice phenology monitoring [15,16], lodging monitoring [17], etc. However, its application in rice mapping is just beginning. Yonezawa et al. [18] explored the applicability of multi-temporal full polarimetric airborne L-band SAR scattering in rice field surveying by applying the polarization decomposition method to fully polarized SAR data. Ma et al. [19] used the covariance matrix and the results of the H/α decomposition of five phases of dual-polarized SAR data as the input of 35 channels to achieve the identification of rice planting areas. However, current research lacks a comparison of different polarization decomposition methods, and the physical mechanisms of the corresponding mapping methods are not clear. In conclusion, the application of polarization decomposition methods in rice mapping still requires in-depth research.

Deep learning methods have been widely used in the field of remote sensing, especially in remote sensing image classification, due to their excellent performance [20].In rice mapping research, deep learning methods have also been widely applied, such as the LSTM model [10] and the dual-branch BiLSTM model [9] mentioned above. However, these models require huge input data such as the SAR data of a whole year of optical data of multiband, increasing the difficulty to train. If the polarization decomposition method is used to obtain effective features to distinguish between rice and other ground objects, the use of long time series data as input for segmentation in the subsequent segmentation step can be avoided, further avoiding the use of complex models such as LSTM. Thus normal semantic segmentation small models can be applied, such as UNet [21], which has stable performance and excellent segmentation results in the field of remote sensing image classification, with a series of improved models based on it. Ge et al. [22] used the UNet model in the crop classification experiment and achieved an average optimal overall accuracy of more than 87% in all target regions. The RAUNet (Residual Attention UNet) proposed by [23] achieved an average accuracy of 90% in fine-scale crop mapping of very high resolution (VHR) images. Xu et al. [24] proposed the HRUnet (High Precision UNet) for the task of cropland extraction, which achieved an accuracy of 92.81% in the experiment. In order to achieve the goal of reducing the computational complexity and complexity of rice mapping, the ODCRS model was selected for experiments on extracting rice planting areas, showing excellent performance in extracting cropland in rice planting areas due to the good balance between the loss of detail features and information redundancy [12]. To further reduce the difficulty in sample making for training deep learning models, a two-stage segmentation structure was used in the study, which also helped to improve the mapping accuracy.

Dual-polarized SAR data, especially the data of the Sentinel-1 satellites launched by the European Space Agency (ESA), are widely used in agricultural-related applications due to its large coverage range, relatively high resolution, and relatively complete time series [25]. Therefore, the VH/VV polarization data of Sentinel-1 and the corresponding polarization decomposition methods are considered for use in the rice mapping research in this article. For non-coherent agricultural targets, polarization decomposition methods are generally divided into model-based decomposition methods and eigenvalue–eigenvector-based decomposition methods [26]. For dual polarized SAR data, two types of widely used methods are the m/χ decomposition [27] and H/α decomposition [28] methods, which are applied in the rice mapping experiments in this study.

Considering the above, a rice mapping method based on polarization decomposition methods of dual-polarized SAR data and deep learning is proposed in this study, with Suihua City, Heilongjiang province, China as the study site and VH/VV-polarized data of Sentinel-1 as the main data source. The specific contributions include:

1. The sensitivity of m/χ decomposition and H/α decomposition to rice in different growth stages are compared. An effective feature is proposed to extract the rice planting areas: the combination of volume-scattering components during the early flooding period of rice.

2. A two-stage segmentation structure is introduced to reduce the effort and difficulty in sample making, which also improves the accuracy of rice mapping.

3. The application experiments of the proposed rice mapping method are conducted in different years (Suihua City in 2019 and 2022) and regions (Suihua City in 2022 and the central region of Jilin Province), verifying the stability of the proposed features and providing a new approach for early rice detection.

2. Materials and Methods

2.1. Study Site

The main study site for this research was Suihua City, Heilongjiang province, China, located at 45°10′ N–48°05′ N, 124°53′ E–128°35′ E. The site lies in the Hulan River Basin of the Songnen Plain, with the northeastern part being the western slope of the Xiaoxing’an Mountains, the central part being a hilly area, and the southwestern part being a plain. Overall, it forms a high eastern and gradually sloping southern strip. This region has a continental monsoon climate in the cold temperate zone, with significant seasonal changes and significant differences in climate during the four seasons. The specific location of Suihua City is shown in Figure 1.

The rice growing season in this region is from April to September. Seeding is completed in early May, and transplanting is completed by the end of May. After about a week of transplanting, the tillering stage lasts for approximately 30 days, followed by a booting stage of approximately 20 days, a heading stage of approximately 15 days, and a grain-filling stage of approximately 40 days. The rice matures at the end of September, and harvesting begins in October. Irrigation in the paddy field stops after the heading stage, so the preceding periods are the flooding period. Figure 2 shows photos of field surveys in Suihua in 2023, taken on 6 June, 2 August, and 7 September, respectively. The corresponding rice growing stage are the end of the transplanting stage, the heading stage, and the grain-filling stage in the region.

2.2. Dataset

Sentinel-1 satellite data were selected as the main data source, with the optical data from the Google Earth platform as auxiliary data for sample making. In order to apply the polarization decomposition method, the Sentinel-1 data used were the Single Looking Complex (SLC) data of VH/VV polarization. The year 2022 was selected as the research period in the experiment, and four frames of data were required to fully cover the entire area of Suihua, which are orbit-frame 32–433, 32–438, 105–435, and 105–441. However, due to the malfunction of the Sentinel 1-B satellite, there were no data for the entire year 2022 for the two frames 32–433 and 32–438. Therefore, only the area covered by the 105–435 and 105–441 frames in Suihua City was selected as the actual study site, as shown in Figure 1. Specifically, when investigating the sensitivity of the polarization decomposition method for rice extraction, all data for the rice growing season (April to September) in 2022 were obtained from the 105–435 frame. When applying and validating the proposed rice extraction method, three phases (the decision of the specific phases is presented in Section 3) of 105–435 and three phases of 105–441 in 2022 and 2019, respectively, were used. In addition, when validating the method of this study in other regions, three phases of 105–448 in 2022 were used. The specific phases used are shown in

Table 1. Information of Sentinel-1 data used.

Orbit-Frame	105–435	105–441	105–448
Year		Date (MM/DD)
2022	04/08
	04/20
	05/02
	05/14	05/14	0503
	05/26
	06/07
	06/19	06/19	0608
	07/01	07/01	0620
	07/13
	07/25
	08/06
	08/30
	09/11
	09/23
2019	05/13	05/13
	06/18	06/18
	06/30	06/30

.

2.3. Method

In response to the problem of using long time series data and neglecting the scattering phase information of ground objects in current rice mapping research, a novel single-season rice mapping method based on the decomposition components of dual-polarized SAR data and a high-precision small deep learning model is proposed in this study. Firstly, the SLC data of Sentinel VH/VV polarization on frame 105–435 from April to September 2022 are pre-processed by SNAP, including calibration, covariance matrix calculation, multi-looking, and terrain correction; secondly, m/χ decomposition and H/α decomposition are performed and compared, obtaining the most sensitive components to extract rice planting areas; thirdly, the same preprocessing and decomposition are performed on the whole study site in the selected phases, constructing the final feature map with corresponding decomposition components. After the feature construction is completed, the classification step is performed on the constructed feature map with a two-stage segmentation structure. In the first stage, rice region labels are manually annotated based on optical images, which are used to obtain coarse extraction results using the ODCRS model. In the second stage, a larger sample set is created using the coarse extraction results as new labels. Then the final rice mapping results are obtained after training and prediction again with the ODCRS model and validated with the field survey results. Meanwhile, the proposed method is applied to different years in the same region and different regions in the same year for effectiveness verification. The specific flow chart is shown in Figure 3.

2.3.1. Feature Construction

The feature construction experiment is carried out on frame 105–435. According to the field survey, the rice planting areas in this region are distributed along the Hulan River, Tongken River, Kayin River, Nomin River, and Ougen River basins. The distribution of its geographical location in frame 105–435 is also shown in Figure 1. The green box represents the distribution area of the main rice production area. The gray image is the volume scattering image in 14 May 2022. Within the green box, rice is distributed in the darker areas along rivers.

First, the corresponding decomposition components of H/α decomposition and m/χ decomposition are obtained from the pre-processed covariance matrix for the rice growth season (April to September) in 2022.

H/α Decomposition

The covariance matrix $C$ is first decomposed into eigenvalues and eigenvectors when computing the $H$/$\mathsf{\alpha }$ decomposition,

$$C=U\left[\begin{array}{cc}{\lambda }_1& \\ & {\lambda }_2\end{array}\right]U^T,{\lambda }_1\ge {\lambda }_2$$

(1)

In the equation is the matrix of eigenvectors $\overrightarrow{u_1}$ and $\overrightarrow{u_2}$,

$$U=\left[\overrightarrow{u_1}\overrightarrow{u_2}\right],\overrightarrow{u_i}=e^{i{\varphi }_i}\left[\begin{array}{c}cos{\alpha }_i\\ e^{i{\delta }_i}sin{\alpha }_i\end{array}\right]$$

(2)

Therefore, entropy $H$, anisotropy $A$, and average scattering angle α are

$$H={\displaystyle \sum }_{i=1}^2-P_i{\log }_2{P}_i,0\le H\le 1$$

(3)

$$A=\frac{{\lambda }_1-{\lambda }_2}{{\lambda }_1-{\lambda }_2},0\le A\le 1$$

(4)

$$\mathsf{\alpha }={\displaystyle \sum }_{i=1}^2{P}_i{\alpha }_i={\displaystyle \sum }_{i=1}^2{P}_i{\cos }^{-1}\left(\left|u_{1i}\right|\right),0°\le \alpha \le 90°$$

(5)

Of which,

$$P_i=\frac{{\lambda }_i}{{{\displaystyle \sum }}_{j=1}^2{\lambda }_j},i=1,2$$

(6)

Figure 4 shows the decomposition results of frame 105–435 of all phases from April to September 2022. It can be seen that the overall entropy $H$ in the region shows an increasing trend during this time period. The entropy of the rice distribution areas increases faster in May and June (transplanting and tillering stages), which shows a significant difference from the surrounding areas but is similar to the changes in some non-watershed areas on the left, causing difficulty in distinguishing rice from all ground objects. The overall change in anisotropy $A$ in this region is relatively small, with the rice region slightly lower than other regions in late May and early June. The average scattering angle $\mathsf{\alpha }$ in the region is relatively small and has little variation, with the rice region also being difficult to distinguish from other regions.

$m$/$\mathsf{\chi }$ Decomposition

The $m$/χ decomposition is calculated from the Stokes vector. The formula for calculating Stokes vectors from the covariance matrix is

$$g=\left[\begin{array}{c}{g}_0\\ g_1\\ g_2\\ g_3\end{array}\right]=\left[\begin{array}{c}{C}_{11}+C_{22}\\ C_{11}-C_{22}\\ 2R\left(C_{12}\right)\\ 2I\left(C_{12}\right)\end{array}\right]$$

(7)

Then, the degree of polarization $m$ and the sign of rotation of the polarization ellipse and its ellipticity $\mathsf{\chi }$ are calculated:

$$m=\frac{\sqrt{g_1^2+g_2^2+g_3^2}}{g_0}$$

(8)

$$\sin 2\mathsf{\chi }=-\frac{{\mathrm{g}}_3}{m{g}_0}$$

(9)

From this, double bounce scattering, volume scattering, and odd bounce scattering components are obtained:

$$\left[\begin{array}{c}{V}_{DoubleBounce}\\ V_{Volume}\\ V_{SingleBounce}\end{array}\right]=\left[\begin{array}{c}\sqrt{g_{0}m\frac{1+sin2\chi }{2}}\\ \sqrt{g_0\left(1-m\right)}\\ \sqrt{g_{0}m\frac{1-sin2\chi }{2}}\end{array}\right]$$

(10)

Figure 5 shows the decomposition results of frame 105–435 of all phases from April to September 2022. It can be seen that during this time period, the variation trend of the scattering component intensities in the region is similar and gradually increases with time. From mid-May to the end of July, the scattering intensity in the rice planting area is lower than that in the surrounding area, and the difference in volume scattering components between the rice planting area and the surrounding area is greater than that of the double-bounce scattering and odd-bounce scattering. The largest difference appears in mid-May and mid- to late June, which are the early stage of transplanting and the middle to late stage of tillering, respectively. This period belongs to the early stage of rice flooding, when the rice field is adequately irrigated and the plants are short. Therefore, the scattering of rice planting areas is dominated by water bodies, which can be distinguished from other vegetation. Due to the close scattering intensity between the rice area and the non-watershed areas on the left in a single phase, the idea of combining multiple temporal volume scattering components is adopted to enhance the difference between the rice and non-rice areas.

Feature Map

Based on the above analysis, the combination of the volume scattering components of m/χ decomposition in the early stage of transplanting and the middle to late stage of tillering is selected as the final classification feature, with the specific dates of 14 May, 19 June, and 1 July 2022. The feature map of the whole study site is shown in Figure 6 (for the best visual effect, the pseudo-color composition is presented in the order of R: 1 July, G: 19 June, B: 14 May, but the order of composition does not affect the classification performance and results. The subsequent feature maps are also arranged in this order).

It can be seen that in the pseudo-color composite image, most of the rice planting areas appear dark red, while a few areas appear bluish purple due to differences in seeding time, as shown in Figure 7.

2.3.2. Two-Stage Segmentation

To further improve the accuracy of rice mapping and reduce the difficulty in sample making, a two-stage segmentation structure is introduced in the segmentation part.

In the first stage, a small number of rice samples are first annotated manually with reference to the optical data and the feature map, as there is a significant difference between rice and other ground objects in the feature map. Specifically, the rice labels are polygons drawn in ENVI, presented in Section 3.1. The coarse extraction result is then obtained by segmenting the feature map with the ODCRS model. In the second stage, referring to the optical data and feature map, the areas with high classification accuracy in the coarse extraction result are selected to create a new sample set, thus achieving the expansion of the sample set, effectively improving the classification accuracy, and reducing the workload. Then, ODCRS is applied for the second segmentation to obtain the final mapping result.

Since rice is clearly distinguished from other ground objects in the pseudo-color composite image, the difficulty of classification from the data source is greatly reduced. Therefore, in the segmentation step, a small model with excellent classification performance and low training difficulty could be used for classification. The ODCRS model has shown excellent performance in extracting cropland in the Mekong Delta region, where the main cropland in the region is rice fields with abundant water resources, similar to the study site in this study. Therefore, the ODCRS model is used in the two-stage segmentation structure.

By introducing the residual network structure and omni-dimensional dynamic convolution, the ODCRS model has excellent fitting ability. Furthermore, the introduction of complementary attention can effectively suppress activation in irrelevant regions, reduce redundant information in skip connections, simplify the training of the decoder, and improve the model training speed and classification accuracy. The application of the dropout module can also effectively improve the generalization ability of the model. The specific structure of the ODCRS model is shown in Figure 8. The RODConv module is a full dimensional dynamic convolutional residual network.

3. Results

In this section, the experimental results are presented and analyzed based on four aspects: the effectiveness of the two-stage structure, the comparison between the model performance and the rice mapping results, the accuracy of the test on field survey data, and the feature validity.

During the experiment, the models were built on the Pytorch framework (1.13.1 version) in Python 3.9.16. The input data were three-channel Tiff image and the output data were one-channel Tiff image. During the training process, the batch size, learning rate and the momentum were 8, 0.0001, and 0.9, respectively.

3.1. Effectiveness of the Two-Stage Structure

Figure 9 shows the manually drawn sample label area used in the first stage of segmentation and the sample labels cut out from the coarse extraction results used in the second stage of segmentation. The black areas are non-rice areas, and the green areas are rice areas. The slice size of the sample set is 256 × 256. The number of slices in the first stage sample set is 1244, and the number of slices in the second stage sample set is 5206, more than 4 times the number of slices in the first stage.

In the segmentation experiment, the loss function for model training is the cross-entropy function, and the optimizer is the Adam optimizer. During the training process, the sample set is divided into a training set and a validation set at a ratio of 0.7/0.3, and the training epoch of all models are 40.

Table 2. Segmentation indexes on the validation set.

	Rice			All Classes (Rice and Non-Rice)
	IoU	Recall (PA)	Precision	mIoU	mPA	Accuracy
One-Stage ODCRS	72.23%	87.54%	80.50%	84.82%	92.95%	97.57%
Two-Stage ODCRS	78.67%	89.83%	86.36%	88.39%	94.36%	98.24%

shows the segmentation indices of the one-stage segmentation and the two-stage segmentation on the validation set, and Figure 10 shows the loss variations. It can be seen that for the rice class, the introduction of the two-stage structure improves the IoU (Intersection of Union) by 6.44%, the recall (equivalent to pixel accuracy) by 2.29%, and the precision by 5.86% compared to those of the one-stage segmentation. And for all classes (rice and non-rice), the two-stage structure improves the mIoU (mean intersection of union) by 3.57%, the mPA (mean pixel accuracy) by 1.41%, and accuracy by 0.67%. In addition, the loss and its convergence speed are also improved, indicating the effectiveness of the two-stage segmentation structure and the expansion of the sample set.

3.2. Comparison of the Model Performance and the Rice Mapping Results

The model performance of UNet [21] and two-stage ODCRS is compared in

Table 3. Model performance comparison.

	Rice			All Classes (Rice and Non-Rice)
	IoU	Recall (PA)	Precision	mIoU	mPA	Accuracy
UNet	66.04%	85.77%	74.17%	81.32%	91.72%	96.82%
Two-Stage ODCRS	78.67%	89.83%	86.36%	88.39%	94.36%	98.24%

. For the rice class, the introduction of two-stage structure improves the IoU by 12.63%, the recall by 4.06%, and the precision by 12.19% over those of UNet. And for all classes (rice and non-rice), the two-stage structure improves the mIoU by 7.07%, the mPA by 2.64%, and the accuracy by 1.42%.

Figure 11 shows the mapping results of the two-stage ODCRS model and UNet. It can be seen from the results of the UNet model that there are large areas of misclassification in the non-watershed area (marked with red boxes). In turn, for the two-stage ODCRS model, the distribution of rice is concentrated in river basins, which is consistent with prior knowledge, indicating that the two-stage ODCRS model greatly improves the mapping results as well as the segmentation indices. More detailed results are presented in Figure 12.

In terms of mapping details, the mapping result of UNet has relatively more smooth and continuous edges, but there are many missing rice regions (such as the areas marked with red boxes in Figure 12). Although the ODCRS model result has a slight decrease in edge smoothness and connectivity, due to its excellent fitting and generalization ability, the omission of rice regions are greatly reduced. And for the non-rice regions in the fourth column, the result of UNet has large areas of misclassification, while the two-stage result of the ODCRS model has almost no misclassification in this area. This also proves the effectiveness of the two-stage segmentation structure and the proposed feature.

3.3. Testing Accuracy on Field Survey Data

This section tests the accuracy of the classification results based on the field survey data from 6 June 2023 to 8 June 2023 in Suihua City. In this field survey, a total of 340 rice fields with a total of 186,929 pixels and 222 non-rice fields with a total of 104,551 pixels were recorded. As shown in Figure 13, the green area represents rice and the orange area represents non-rice. Since the UNet result has large visible misclassifications outside the field survey area, it is not meaningful to test it on the field survey data, so it is not tested in this section.

Table 4. Mapping accuracy of two-stage ODCRS model.

	Field Investigation Data/Pixels				User Accuracy
Mapping Results /Pixels		Rice	Non-rice	Sum
	Rice	165,814	31,130	196,944	84.19%
	Non-rice	21,115	73,215	94,330	77.62%
	Sum	186,929	104,555	291,484
Mapping Accuracy		88.70%	70.03%

shows the mapping accuracy of the two-stage ODCRS model on field survey data.

When evaluating the segmentation model performance on field survey data, mapping accuracy and user accuracy are:

$$\mathrm{Mapping}\mathrm{Accuracy}=\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(11)

$$\mathrm{User}\mathrm{Accuracy}=\mathrm{Precesion}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(12)

It can be seen that for the rice region, the mapping accuracy of the ODCRS model is 88.70%, and the user accuracy is 84.19%, proving the effectiveness of the proposed features for rice mapping and the two-stage segmentation structure.

3.4. Feature Validity

This section demonstrates the validation of rice extraction effectiveness based on the proposed features and the ODCRS model in different years and regions.

3.4.1. Validity in Time

Figure 14 shows the feature maps and rice extraction results of Suihua City in 2022 and 2019. It can be seen that in the same region in different years, the rice regions in the proposed feature combination have the same characteristics. The characteristic that the volume scattering component of rice is lower than that of other ground objects in the early stage of transplanting and the middle to late stage of tillering is similar in different years. At the same time, this result also proves that the model trained based on data from 2022 also has temporal scalability.

3.4.2. Validity in Space

In this part of the experiment, the area of frame 105–448 in 2022 was selected as the validation experimental area for rice extraction. The feature map of the region and the rice extraction results are shown in Figure 15a. An example of the zoomed feature map and corresponding extraction results are shown in Figure 15b.

This area is located in the central part of Jilin Province, bordering Liaoning Province in the lower left corner, and includes some areas of cities such as Siping City, Liaoyuan City, Tonghua City, Changchun City, and Jilin City. The latitude of this area is lower than that of Suihua City, so the phases of the source data 11 days earlier than that of Suihua City are selected to construct the feature map. In the feature map of the region, the rice region is bluish purple, and the rice extraction results are also distributed along water bodies, which is consistent with the prior knowledge, proving that the effectiveness of the proposed feature combination for single-season rice does not change with spatial variation. At the same time, the model trained using data from the Suihua region also has some scalability in spatial dimensions

4. Discussion

This study first analyzed the scattering characteristics of VH/VV polarization data of single season rice during the growing season and compared the difference between rice and other ground features on the components of H/α decomposition and m/χ decomposition at different time periods. It was found that the volume scattering of rice paddy is lower due to the influence of water bodies during the early flooding period, especially during the early transplanting period and the middle to late tillering period. The combination of the volume scattering components of m/χ decomposition can effectively extract rice planting areas.

From Section 3, it can be seen that the two-stage segmentation structure introduced in this study can effectively improve the accuracy of rice extraction and reduce the difficulty in sample making. The temporal and spatial applicability of the proposed feature combination are also verified with experiments conducted in Section 3.4. Compared to mapping methods using long-term time series of SAR data, optical data or a combination of both as data sources, the proposed method only requires three phases of VH/VV-polarized SAR data, and the model input is a mere three-channel image, achieving the goal of reducing computational complexity and the difficulty in model training. And the physical mechanism of the proposed rice mapping method is clear: the volume scattering of rice paddy is notably lower than other vegetation during the early flooding period. With the two-stage segmentation structure, the sample set for the deep learning method could be effectively extended, thus reducing the difficulty in sample making and improving the segmentation result no matter how the samples are labeled in the first stage.

Although the proposed rice mapping method performs well, some improvement can still be made. For example, non-rice regions were regarded as background instead of another category in the segmentation experiments. Further research is needed to determine whether this is the reason why the mapping accuracy and user accuracy of non-rice regions are lower than those of rice regions when testing the segmentation accuracy with field survey data.

5. Conclusions

In response to the problem of large computational complexity and complex classification models caused by the use of long time series data in current rice mapping research, and based on the principle that the scattering mechanism in the early transplanting period and the middle to late tillering period of rice during the early flooding period is greatly affected by water bodies and clearly distinguished from other ground objects, this study proposed a set of feature combinations that can effectively extract rice planting areas with VH/VV-polarized SAR data and the m/χ decomposition method, specifically the combination of volume scattering components in the early transplanting period and the middle to late tillering period of rice. Using a two-stage ODCRS model, high-precision rice mapping was achieved in Suihua City. Experiments showed that the introduction of the two-stage segmentation structure reduces the difficulty of deep learning sample production and greatly improves segmentation accuracy. Meanwhile, validation experiments in different years and regions in the study demonstrated the temporal and spatial applicability of the proposed features, providing a new approach for early detection of large-scale single-season rice.

In future work, we will consider extending the method of this study to rice mapping research under more complex terrains and planting patterns.