Mapping Paddy Rice in Rice–Wetland Coexistence Zone by Integrating Sentinel-1 and Sentinel-2 Data

Abstract

Accurate mapping of vegetation in the coexisting area of paddy fields and wetlands plays a key role in the sustainable development of agriculture and ecology, which is critical for national food security and ecosystem balance. The phenology-based rice mapping algorithm uses unique flooding stages of paddy rice, and it has been widely used for rice mapping. However, wetlands with similar flooding signatures make rice extraction in rice–wetland coexistence challenging. In this study, we analyzed phenology differences between rice and wetlands based on the Sentinel-1/2 data and used the random forest algorithm to map vegetation in the Poyang Lake Basin, which is a typical rice–wetland coexistence zone in the south of China. The rice maps were validated with reference data, and the highest overall accuracy and Kappa coefficient was 0.94 and 0.93, respectively. First, monthly median composited and J-M distance methods were used to analyze radar and spectral data in key phenological periods, and it was found that the combination of the two approaches can effectively improve the confused signal between paddy rice and wetlands. Second, the VV and VH polarization characteristics of Sentinel-1 data enable better identification of wetlands and rice. Third, from 2018 to 2022, paddy rice in the Poyang Lake Basin showed the characteristics of planting structure around the Poyang Lake and its tributaries. The mudflats were mostly found in the middle and northeast of Poyang Lake, and the wetland vegetation was found surrounding the mudflats, forming a nibbling shape from the lake’s periphery to its center. Our study demonstrates the potential of mapping paddy rice in the rice–wetland coexistence zone using the combination of Sentinel-1 and Sentinel-2 imagery, which would be beneficial for balancing the changes between paddy rice and wetlands and improving the vulnerability of the local ecological environment.

1. Introduction

Global climate change and the frequency of extreme weather events brought on by excessive carbon emissions have had a significant negative influence on socioeconomic development. This poses a serious problem for humanity in the modern era [1]. In September 2020, the Chinese government declared “carbon peaking and carbon neutrality” a strategic target. Carbon sequestration by terrestrial plants through photosynthesis is a vital step in the process of reaching carbon neutrality [2]. In the southern regions of China, farmlands and wetland ecosystems are intricately entwined as essential components of terrestrial ecosystems, and they have a major impact on regional climate change and the global carbon cycle [3]. Accurately mapping the spatial distribution of rice paddies and wetlands is essential for remote sensing simulation of terrestrial vegetation in carbon sequestration. This data provides crucial support for decision-making related to national food security and the achievement of the “dual carbon” strategic goal [4].

Traditional methods of obtaining information on vegetation distribution in farmlands rely on manual surveys or statistical analysis, which are time-consuming, labor-intensive, and inaccurate in terms of spatial distribution. Remote sensing technology holds a significant advantage in rice mapping because it uses high-frequency revisits in large-scale coverage. Leveraging the Google Earth Engine (GEE) big data cloud platform provides new opportunities for intelligent interpretation of rice paddy distribution with remote sensing technology. Previous studies have shown cloud computing and storage capability in mapping paddy rice [5,6]. Flood signal characteristics during the irrigation and transplanting stages of rice paddies serve as the primary indicators to distinguish paddy rice from other crops. At this stage, the relationship between the land surface water index (LSWI) and enhanced vegetation index (EVI)/normalized difference vegetation index (NDVI) (LSWI > NDVI/EVI) was proven to be effective in identifying paddy rice [7,8,9]. Due to the inherent defects in requiring clear-weather conditions for optical remote sensing data, weather-independent radar remote sensing has proven to be a helpful means for rice mapping. The VV and VH polarization signals in radar data are sensitive to water and moisture, showing their potential to detect under-canopy water bodies [10,11]. Recent research has found that time series Sentinel-1A VH backscatter signals can enhance the identification of paddy rice during the flooding and transplanting stages [12,13,14].

Meanwhile, rice paddies and wetlands coexist in several regions of southern China, including the plains of Poyang Lake and Dongting Lake. The wetland vegetation area varies with the water surface change of Poyang Lake, and its growing season usually ranges from April to September during the wet season of each year. During the growth season, both paddy rice and wetland vegetation experience flood stages, leading to close signatures of spectral signals in optical imagery and backscatter characteristics in SAR data between them for mapping rice [15,16,17]. Relying solely on optical imagery or radar data were inadequate to meet the requirements of refined mapping in paddy rice. Specifically, cloud contamination is a potential threat to optical imagery in rice mapping in southern China, resulting in few good observations at the transplanting stage of paddy rice. Similarly, subtle vegetation signatures may make it difficult to derive information on paddy rice when using time-series radar data [18,19]. As a result, challenges for rice mapping arise from frequent cloud-contaminated observations during the transplanting stage and minor variations in signal signature between paddy rice and wetland vegetation in the rice–wetland coexisting zones (RWCZ). Consequently, it is essential to combine the fine spectral signature of optical imagery with the weather-independent observation capability of SAR satellites for rice mapping in RWCZ.

Recently, a few studies have demonstrated the potential of the integration of Sentinel-1 radar data and Sentinel-2 optical imagery in rice mapping [20,21]. The high spatial resolution (10–20 m) and high-frequency revisit (5–12 days) can help to obtain dense time series observations during the growing season of paddy rice. However, mapping paddy rice with the combination of Sentinel-1/2 faces challenges in the RWCZ, and its potential should be further investigated.

The Poyang Lake Plain is a typical region where rice paddy and wetland ecosystems coexist due to its fractured terrain and complex agricultural landscape with a broad spread of farmlands, wetlands, and water bodies. We focused on the rice–wetland coexistence zone in the Poyang Lake Plain, integrating field samples, Sentinel-1/2 data, and machine learning algorithms to map paddy rice in the region. Our study explicitly investigated the phenological stages that amplify signal differences between rice and wetlands. A comprehensive solution is provided to solve the difficult problem of land cover classification in the rice–wetland coexistence area of the Poyang Lake Plain. Furthermore, identifying key vegetation indices that delineate significant distinctions between these land cover types is crucial for enhancing the accuracy of mapping results in areas where rice and wetlands intermingle [22,23]. This study asks two questions: (1) which phenological stage can be used to amplify the signal difference between paddy rice and wetland? (2) which vegetation index can find the significant distinctions between the two types of land cover? This method contributes to advancing the development of machine learning approaches for rice–wetland mixed areas and improving classification accuracy. It also provides valuable insights for sustainable land management and achieving carbon neutrality goals in southern China.

2. Materials and Methods

2.1. Study Area

Poyang Lake stands as a vital wetland resource in southern China. The Poyang Lake Plain is a lakeside plain formed by the alluvial action of the Ganhe River, Fuhe River, Xinjiang River, Xiushui River, Raohe River, and other tributaries of the Yangtze River system. Farmlands, rivers, ponds, and wetlands intersect in the region, which has fragmented land parcels, crisscrossing rivers, and dense lakes [24,25,26]. It covers a land area of approximately 32,400 square kilometers (Figure 1) and has a subtropical humid climate with moderate temperatures, abundant rainfall, ample sunlight, and a long frost-free period within one year. With an annual precipitation of about 1400 mm and suitable conditions during the rainy and warm seasons, it is helpful for paddy rice cropping. With such weather conditions, both single rice and double rice were historically cropped, and this region serves as an important grain production base in the middle and lower reaches of the Yangtze River.

2.2. Data Sources

2.2.1. Remote Sensing Data

The Sentinel-1/2 mission comprises two polar-orbiting satellites [27]. Sentinel-1 carries a c-band Synthetic Aperture Radar (SAR) with a revisit period of 6 days at the equator and a spatial resolution of 10 m. It incorporates various polarization modes such as single and dual polarization. On the other hand, Sentinel-2 is equipped with a multi-spectral imager (MSI) comprising 13 spectral bands. Its revisit period at the equator is 5 days with a spatial resolution of up to 10 m. Sentinel 1/2 satellite data can be found at https://code.earthengine.google.com/ (accessed on 1 January 2024).

Sentinel-1 radar data and Sentinel-2 multi-spectral data are both accessible online via the Google Earth Engine (GEE) platform. The data acquisition timeframe spans from April to October, covering the rice-growing season. The Sentinel-1 radar data comprises Ground Range Detected (GRD) images using the Interferometric Wide swath (IW) mode, including VH and VV polarization. Meanwhile, Sentinel-2 imagery represents the surface reflectance product, pre-processed for atmospheric and orthorectification corrections. Cloud masking was performed using the QA band, and a monthly mean composite was generated for spectral indices like vegetation indices and moisture indices derived from optical bands. Masking was conducted based on the study area, constructing a monthly dataset for optical vegetation indices and radar polarization parameters.

Table 1. Band information of images employed in the research.

Sensor	Bands		Wavelength/nm	Resolution/m
Sentinel-2 MSI	B2	Blue	490	10
	B3	Green	560	10
	B4	Red	665	10
	B8	Near-infrared	842	10
	B11	Short-wave infrared 1	1610	20
	B12	Short-wave infrared 2	2190	20
Sentinel-1 SAR GRD	VV	Dual-band cross-polarization, vertical transmission/horizontal receiver	-	10
	VH		-	10

illustrates the information about the various bands in Sentinel-1/2 data.

2.2.2. Validation Data

The training samples are crucial for remote sensing mapping of vegetation with machine learning methods [28,29]. It is essential to select typical and representative pure pixels of different land cover types as sample data for the training of the machine learning algorithm. Based on vegetation phenology and spectral index features (see detailed in Section 2.3.2), we selected samples accurately and effectively through visual interpretation via the GEE platform. After visual screening, a total of 1600 samples were obtained, including single rice, double rice, mudflat, wetland vegetation, water bodies, construction land, forest, and dry land. In summary, we obtained 1981 samples of eight land cover types, and these samples were randomly divided into training and validation samples in a 7:3 ratio for model training and accuracy assessment. The classification system and the sample information are shown in

Table 2. Land use classification system.

Land Cover	Characteristics	Number
Single rice	Also referred to as middle rice. Rice harvested once a year in the same paddy field. Exhibits a smooth and uniform texture within the plot.	330
Double rice	Rice cultivated and harvested twice a year in the same paddy field that includes early rice and late rice. Demonstrates a smooth and uniform texture within the plot.	330
Mudflat	Sandy areas appear bright yellow, while mudflats exhibit a light gray color with clear boundaries, located near water bodies.	134
Wetland vegetation	Plants growing in areas where the soil is permanently or seasonally saturated with water, including emergent, floating, and submerged vegetation.	150
Construction land	Appears silver-white, occurring in contiguous or sporadic distributions with uniform and regular patterns.	320
Forest	Exhibits a deep green color, irregularly distributed in large patchy or cluster formations with clear boundaries.	323
Dry land	Encompasses other crops besides paddy rice, demonstrating relatively smooth and uniform plot patterns.	150
Water	Displays varying shades of blue with distinct boundaries.	244

. In addition, we also collected crop planting area data from the National Bureau of Statistics’ “Rural Statistical Yearbook” data to statistically verify the area extracted for paddy rice classification.

2.2.3. Other Land Cover Datasets Used for Comparison

(1)

Selected Land Cover Classification Data Products. Various typical land cover classification data products were chosen to ensure consistency in the classification results. World Cover Dataset: Developed collaboratively by the European Space Agency (ESA) and several global research institutions, this dataset offers land cover data at 10-m resolution with a producer’s accuracy of 75%. It spans the years 2020 and 2021, defined using the Land Cover Classification System (LCCS) by the Food and Agriculture Organization (FAO) of the United Nations. It encompasses 11 categories such as forests, shrubs, Wetland vegetation, croplands, buildings, deserts, snow and glaciers, water bodies, wetlands, mangroves, and lichens/mosses [30].

(2)

GLC_FCS30 Dataset. Developed by the team of Liangyun Liu at the Aerospace Information Research Institute of the Chinese Academy of Sciences. This dataset provides a detailed global land surface cover classification product at 30 m resolution. It boasts an overall accuracy of 82.5%. It is based on the dynamic monitoring of global land cover from 1985 to 2020 using all available land satellite data, updated every 5 years, and consists of 29 categories [31].

(3)

Single Rice Planting Distribution Dataset at 10-Meter Resolution. Created using a time-weighted dynamic time planning method in conjunction with optical remote sensing data and synthetic aperture radar data for the years 2017–2022. It achieves average user’s accuracy, producer’s accuracy, and overall accuracy of 73.08%, 82.81%, and 85.23%, respectively, across all provincial-level administrative regions [32].

(4)

Double Rice Planting Distribution Dataset from 2016 to 2020. Constructed based on Sentinel-1 remote sensing data and phenology identification methods. The overall accuracy ranges between 88.07% and 95.97% for early rice and 88.25% and 95.68% for late-season rice [33].

All land cover data are available from http://www.nesdc.org.cn/ and https://code.earthengine.google.com/ (accessed on 1 January 2024).

2.3. Machine Learning Classification Method Integrating Phenology and Active-Passive Remote Sensing

We developed a novel approach for extracting both rice and wetland distribution data in their coexistence zone following six steps (Figure 2). (1) Constructing vegetation indices and moisture indices using radar polarization bands and spectral bands. (2) Analyzing the divergent growing stages between rice and wetland based on vegetation phenology. (3) Monthly composite method based on different phenological stages to create a dataset of classification features. (4) Using J-M Distance to select optimal features for mapping. (5) The optimal features were used as input parameters of machine learning to map rice and wetlands. (6) The accuracy was evaluated by calculating the confusion matrix.

2.3.1. Constructing Vegetation Indices and Moisture Indices

Vegetation indices can capture the seasonal variations in chlorophyll content during plant growth [34,35], while water indices can monitor soil moisture and vegetation leaf water content [36,37]; therefore, we selected the two types of spectral indices in rice mapping. Three vegetation indices (NDVI, EVI, and SAVI), three water indices (LSWI, NDWI, and MNDWI), and two polarized band data (VV and VH) were used to identify the growth characteristics of rice paddies and wetlands. The formula of these parameters is shown in

Table 3. Calculation formula of sensible surface parameters.

Spectral Indices	Formulas
Normalized difference vegetation index (NDVI)	NDVI = (NIR − Red)/(NIR + Red)
Enhanced vegetation index (EVI)	EVI = 2.5 × (NIR − Red)/(NIR + 6 × Red − 7.5 × Blue + 1)
Synthetic aperture video index (SAVI)	SAVI = ((NIR − Red)/(NIR + Red + 0.5))(1 + 0.5)
Land Surface Water Index (LSWI)	LSWI = (NIR − SWIR)/NIR + SWIR)
Modified normalized difference water index (MNDWI)	MNDWI = (GREEN − SWIR)/(GREEN + SWIR)
Normalized difference water index (NDWI)	NDWI = (GREEN − NIR)/(GREEN + NIR)

.

2.3.2. Key Phenological Windows for Paddy Rice and Wetland

According to the field survey and literature investigation, paddy rice cropped in the Poyang Lake Plain experiences four crucial phenological stages (

Table 4. Paddy rice and wetland phenology calendar of the study area.

Month	4	5	6	7	8	9	10
Early rice	Transplanting Stage	Growth Stage I	Growth Stage II	Maturation Stage
Late rice				Transplanting Stage	Growth Stage I	Growth Stage II	Maturation Stage
Middle rice			Transplanting Stage	Growth Stage I	Growth Stage II	Maturation Stage
Wetland vegetation	Nutrient Absorption Stage		Dormant Stage			Recovery Stage

). (1) Transplanting stage: after transplanting rice seedlings, using LSWI larger than NDVI or EVI, the signals of water flooding and transplanting can be identified. (2) Growth stage I: when rice began to grow after transplanting, NDVI (EVI) increased monotonously, greenness became the main feature, and the water signal decreased. (3) Growth stage II: the stage in which NDVI and EVI reach their peaks during the rice-growing season. (4) Maturation stage: rice starts to mature and enters the harvesting phase. The chlorophyll content decreases rapidly, resulting in a significant decline in NDVI (EVI) value.

Wetland vegetation possesses three key phenological stages. (1) Nutrient absorption stage: wetland vegetation absorbs light and moisture, initiating growth and development. Chlorophyll content increases, resulting in the first peak of NDVI (EVI). (2) Dormant stage: as Poyang Lake enters the high-water period, wetland vegetation becomes submerged and is unable to undergo photosynthesis, causing NDVI (EVI) to reach a trough. (3) Recovery stage: with the rapid decrease in Poyang Lake’s water level, wetland vegetation reappears and resumes growth and development, leading to the second peak of NDVI (EVI).

2.3.3. Monthly Variations of Spectral Indices and Backscattering Coefficients

The monthly variation curves of eight typical spectral indices and backscattering coefficients of single rice, double rice, wetland vegetation, and mudflats are obviously different (Figure 3). The NDVI and SAVI values of wetland vegetation increased rapidly in April but showed a downward trend in May. The values of NDVI, SAVI, VV, and VH reached their lowest points from June to August, while the values of NDWI and MNDWI reached their highest points from June to August. Double rice has two growth peaks, NDVI and EVI, and SAVI has two peaks in June and September. Single rice NDVI, EVI, and SAVI increased rapidly in July but declined rapidly in September. The NDVI, EVI, and SAVI values of the mudflat are all lower than 0.2 from April to October, and the VV and VH values are also lower than those of other land classes. The time series of different vegetation have distinctive features, especially during critical growth stages, so they can be used to identify vegetation types.

The characteristics of critical phenological stages in vegetation have been demonstrated to be useful for extracting vegetation classification information [38,39,40,41]. In this study, the sentinel-1/2 dataset was constructed using the monthly synthesis method, and the median of all valid observations of sentinel-1/2 was calculated during the monthly period.

2.3.4. Optimal Classification Feature Data Set

To determine the stages and features within the paddy rice growth period that are most distinguishable from other land cover types such as wetlands, we employ the SEaTH algorithm (Separability and Thresholds) [42]. The principle of the SEaTH algorithm is based on the assumption that the selected features follow a normal distribution (if the object features do not adhere to a normal distribution, it indicates poor feature separability) and utilizes the Jeffries-Matusita distance (J-M distance) to assess the separability between land cover types. The calculation formula for J-M distance is as follows:

$$J=2\left(1-e^{-B}\right)$$

(1)

$$B=\frac{1}{8}{\left(m_1-m_2\right)}^2\frac{2}{{\sigma }_1^2+{\sigma }_2^2}+\frac{1}{2}\ln [\frac{{\sigma }_1^2+{\sigma }_2^2}{2{\sigma }_1{\sigma }_2}]$$

(2)

In the formula, B represents Bhattacharyya distance, where m_i and σ_i², i = 1,2, denote the mean and variance of a specific feature for two different categories. The value of J ranges from 0 to 2, where a higher value indicates greater separability between the two categories. The focus lies on the easily confused rice paddies and wetlands; hence, the J-M distance mainly considers these two land cover types.

2.3.5. Machine Learning Algorithms and Design of Experimental Plan

Random forest is a non-parametric machine learning algorithm that employs ensemble learning by integrating multiple decision trees for sample training and prediction. The final classification result is determined through a voting process among decision trees, known as the majority vote. The random forest algorithm is widely used due to its capability to generate high-precision remote sensing classification results while avoiding overfitting. Additionally, this algorithm can effectively handle complex scenarios such as missing values and high-dimensional data [43]. Through a simple majority voting decision, the ultimate classification decision is as follows:

$$H(x)=\mathit{argmax}\underset{Y}{\underbrace{max}}{\sum }_{i=1}^{K}I\left(h_i\right(x)=Y)$$

(3)

In the equation, $H\left(x\right)$ represents the combined classification model, $h_i$ represents a single decision tree classification model, $Y$ is the output variable, and $I(°)$ is the indicator function.

Support Vector Machine (SVM) is a powerful machine learning algorithm used for classification and regression. Its core idea is to find the optimal decision boundary to maximize the margin between different classes, effectively performing classification [44]. It includes a total of 12 parameters, such as decision procedure, SVM type, kernel function, and gamma value, among others. Among these parameters, the choice of kernel function is relatively more critical, including linear kernel, polynomial kernel, radial basis function (RBF)/Gaussian kernel, and Sigmoid kernel. The formulas for the four types of kernel functions are as follows.

$$K(x,z)=(x,z)$$

(4)

$$K(x,z)={(\gamma x·z+r)}^d$$

(5)

$$K(x,z)=exp\left({-\gamma ‖x-z‖}^2\right)$$

(6)

$$K(x,z)=tanh(\gamma x·z+r)$$

(7)

In the equation, $r$ is a constant, γ represents the slope, and $d$ represents the order of the polynomial.

Cart Regression Tree is a decision tree algorithm used to build tree-based models for predicting the values of target variables or categorizing data points into different classes [45]. This decision tree provides a clear and interpretable representation of the relationship between features and targets. The definition of the Gini coefficient is as follows:

$$GI=1-{\sum }_{j=1}^J{p}^2(j/h)$$

(8)

$$p(j/h)=\frac{n_j\left(h\right)}{n\left(h\right)}$$

(9)

$${\sum }_{j=1}^{J}p(j/h)=1$$

(10)

This study devised six schemes, as shown in

Table 5. Scheme information.

Scheme	Feature Group	Classifier
Option 1	Spectral feature (NDVI, EVI, SAVI, LSWI, NDWI, MNDWI)	RF
Option 2	Radar signature (VV, VH)	RF
Option 3	Spectral feature + Radar signature	RF
Option 4	Spectral feature + Radar signature (Key phenology)	RF
Option 5	Spectral feature + Radar signature (Key phenology)	Cart
Option 6	Spectral feature + Radar signature (Key phenology)	SVM

. Scheme 1 uses only Sentinel-2 optical data, scheme 2 solely utilizes Sentinel-1 radar data, while scheme 3 combines both optical and radar data. By employing Random Forest (RF) classification, it aims to explore the impact of incorporating spectral and radar features on land-use classification in the Poyang Lake Basin. Schemes 4 to 6 conduct comparative validation of classifications using RF, Support Vector Machine (SVM), and CART regression tree machine learning algorithms on combinations of spectral features and radar features during crucial phenological periods.

2.3.6. Accuracy Assessment

In this study, sample point data were used to verify the classification results, and the accuracy was evaluated by calculating the confusion matrix. Commonly used indexes include producer’s accuracy (PA), user’s accuracy (UA), overall accuracy (OA), and the Kappa coefficient (

Table 6. Calculation formula of accuracy evaluation index.

Precision Index	Calculation Formula
PA	$PA=\frac{x_{ii}}{x_{+i}}$
UA	$UA=\frac{x_{ii}}{x_{i+}}$
OA	$OA=\frac{{\sum }_{i=1}^r{X}_{ii}}{n}$
Kappa	$Kappa=\frac{n{\sum }_{i=1}^r{X}_{ii}-{\sum }_{i=1}^r(x_{i+}\times x_{+i})}{n^2-{\sum }_{i=1}^r(x_{i+}\ast x_{+i})}$

). UA measures the classification accuracy of various ground objects, and OA and Kappa coefficients are used to evaluate the OA of experimental classification.

The equation defines the variables as follows: n represents the total number of pixels in the study area, and r represents the number of land cover types. $x_{ii}$ represents the number of pixels correctly classified for land cover type i, $x_{+i}$ represents the total number of pixels in the reference data for land cover type i, and $X_{i+}$ represents the total number of pixels for land cover type i in the evaluation data.

3. Results

3.1. Optimization Results of Classification Feature Parameters

Both paddy rice and wetland vegetation have spectral characteristics of vegetation and water, which can easily cause confusion and misclassification. So, this study utilizes six spectral features and two radar features to analyze the separability of paddy rice fields and wetlands at different stages. The results of J-M distances quantitatively analyzed the feature differences between paddy rice fields and wetlands each month. The results showed that in the whole phenology, double rice and single rice were not easily confused with mudflat, while wetland vegetation and paddy rice were relatively poor in separability and easy to confuse. July and August are the key phenological periods for distinguishing wetland vegetation from paddy rice (Figure 4). In these two key phenological stages, single rice and double rice show good separability from wetland vegetation because the selected six spectral features and two radar features have relatively large J-M distance values. Therefore, combining spectral index features and radar features during key phenological periods can effectively address the issue of signal mixing between paddy rice and wetlands. In this study, the feature whose J-M distance is greater than one is selected as the preferred choice.

3.2. Accuracy Assessment of Land Cover Identification in Poyang Lake Basin

The classification accuracy of the proposed six experimental schemes is presented in

Table 7. Accuracy statistics of classification results.

Classification		Option 1	Option 2	Option 3	Option 4	Option 5	Option 6
OA		0.86	0.90	0.91	0.94	0.81	0.74
Kappa		0.84	0.89	0.90	0.93	0.78	0.71
Single rice	PA	0.78	0.83	0.83	0.95	0.73	0.69
	UA	0.79	0.82	0.81	0.89	0.75	0.71
Double rice	PA	0.83	0.86	0.85	0.95	0.74	0.58
	UA	0.86	0.93	0.86	0.98	0.61	0.60
Mudflat	PA	0.90	0.84	0.94	0.97	0.66	0.73
	UA	0.91	0.85	0.96	0.98	0.76	0.68
Wetland vegetation	PA	0.90	0.96	0.94	0.97	0.79	0.64
	UA	0.92	0.93	0.95	0.98	0.87	0.73
Water	PA	0.93	0.95	0.94	0.99	0.86	0.81
	UA	0.96	0.97	0.98	0.99	0.89	0.79
Construction land	PA	0.90	0.85	0.90	0.98	0.91	0.87
	UA	0.90	0.87	0.92	0.99	0.88	0.81
Forest	PA	0.90	0.91	0.94	0.95	0.81	0.85
	UA	0.93	0.90	0.96	0.96	0.85	0.91
Dry land	PA	0.77	0.78	0.79	0.84	0.71	0.64
	UA	0.76	0.86	0.84	0.89	0.73	0.72

Note: Overall accuracy (OA); producer’s accuracy (PA); user’s accuracy (UA).

. Scheme 1, which utilized only Sentinel-2 spectral features for random forest classification, yielded the lowest OA and Kappa coefficient, PA and UA were 0.78 and 0.79 for single rice, and PA and UA were 0.83 and 0.86 for double rice. Scheme 2, using only Sentinel-1 radar features, achieved an OA of 0.90 and a Kappa coefficient of 0.89. The PA and UA for paddy rice fields and wetlands vegetation in scheme 2 were higher than those in scheme 1, indicating that VV and VH polarization features are more sensitive in identifying wetland vegetation and paddy rice fields compared to spectral features. Scheme 3, utilizing both Sentinel-1/2 data, exhibited slightly higher OA and Kappa coefficient compared to schemes 1 and 2, but PA and UA of single and double rice were not improved. This is due to the excessive combination of characteristic variables, leading to information redundancy and increasing the complexity of the model. Therefore, after selecting key phenological feature variables, scheme 4 achieved the best classification performance, with an OA of 0.94 and a Kappa coefficient of 0.93, enhancing the efficiency of data processing.

Schemes 4 to 6 compared the ability of different machine learning algorithms in identifying coexisting paddy rice fields and wetlands. The confusion matrix was used for comparative analysis, and the mapping results of the three machine-learning methods were validated based on sample data. Results indicate that the OA of RF, CART, and SVM were 0.94, 0.81, and 0.74, respectively, with Kappa coefficients of 0.93, 0.78, and 0.71. The PA for single rice was 0.95, 0.73, 0.69, and the UA was 0.89, 0.75, and 0.71. For double rice, the PA was 0.95, 0.74, 0.58, and the UA was 0.98, 0.61, and 0.60. From the perspective of validation accuracy, the RF exhibited a certain advantage in identifying coexisting paddy rice fields and wetlands.

Among all schemes, scheme 4 obtained the best OA, Kappa, PA and UP values of single rice, double rice, and wetland. This is because scheme 4 combines the data of sentinel-1/2 and optimizes the key phenology of rice and wetlands, which can effectively improve the classification accuracy of single rice and double rice. Secondly, compared with SVM and CART, random forest is the most suitable machine learning algorithm for rice extraction using this method.

3.3. Rice–Wetland Spatial Distribution

By utilizing the parameter dataset obtained from feature selection, the random forest classification model was applied for the identification of rice–wetland coexistence areas. The results revealed spatial distribution within the Poyang Lake Basin’s rice–wetland area (Figure 5). Rice cultivation exhibited a relatively wide distribution, characterized by rice fields surrounding the Poyang Lake and its tributaries. The spatial distribution of double rice was relatively concentrated, while that of single rice was more scattered. In the areas of the Fuhe River and Xinjiang River, there was a mix of single and double-season rice cultivation. The Ganjiang River Basin primarily cultivates double rice in the east, while the Xiushui River Basin mainly cultivates single rice, and the Raohe River Basin primarily cultivates double-season rice. The mudflats mainly exist in the central and northeastern parts of Poyang Lake, while the wetland vegetation surrounds the mudflats, forming a kind of “bite” shape towards the lake’s center. The period from 2018 to 2022 exhibited fluctuating characteristics in rice and wetland areas (Figure 6). The area of mudflats and wetland vegetation increased, while the total rice area decreased. Specifically, the rice field areas for the years 2018–2022 were 12,011 km², 10,267 km², 11,475 km², 11,843 km², and 10,767 km², respectively. The average annual rice cultivation area over these years was 11,272.8 km², with double rice covering 7454 km² and single-season rice covering 3818.8 km².

The quantitative accuracy evaluation of the rice–wetland classification results was performed using a confusion matrix (

Table 8. Rice–wetland classification accuracy in Poyang Lake Basin from 2018 to 2022.

Year	Accuracy	Single Rice	Double Rice	Mudflat	Wetland Vegetation
2018	UA	0.9	0.9	0.95	0.97
	PA	0.91	0.98	0.95	0.97
	OA	0.91
	Kappa	0.90
2019	UA	0.89	0.95	0.95	0.97
	PA	0.92	0.89	0.98	0.98
	OA	0.94
	Kappa	0.93
2020	UA	0.88	0.91	0.93	0.97
	PA	0.88	0.91	0.88	0.96
	OA	0.90
	Kappa	0.89
2021	UA	0.83	0.94	0.89	0.96
	PA	0.88	0.92	0.92	0.88
	OA	0.87
	Kappa	0.88
2022	UA	0.86	0.89	0.90	0.98
	PA	0.9	0.95	0.97	0.9
	OA	0.85
	Kappa	0.87

). For the years from 2018 to 2022, the average UA accuracy and PA for single rice, double rice, mudflat, and wetland vegetation ranged from 0.89 and 0.98. The OA ranged from 87% to 94%, and the Kappa coefficient ranged from 0.85 to 0.93.

4. Discussion

4.1. Integration of Times Series Sentinel-1 and Sentinel-2 Imagery

Currently, there are many studies focused on using remote sensing methods to identify the extent of paddy rice cultivation areas [46,47]. However, there is limited research on extracting rice-related information in the rice–wetland coexistence areas of the Poyang Lake Plain and analyzing the growth differences between paddy rice and wetland vegetation using Sentinel SAR and optical data. This study focuses on the Poyang Lake Plain, analyzing the differences in wetland and paddy rice at different stages based on optical and radar index characteristics. After comparative analysis, it suggests that using a combination of key period spectral and radar indices, employing the random forest method for extracting paddy rice and wetland information, is an applicable method for land cover identification in the rice–wetland coexistence areas of the Poyang Lake Plain. Sentinel-2 optical image data [48,49], due to its free source, high temporal and spatial resolution, and ease of processing, can be considered an optimal data source for monitoring paddy rice cultivation areas in the Poyang Lake Plain. SAR is not affected by weather conditions for paddy rice identification and can serve as supplementary data when optical images are unusable due to high cloud cover, thus enhancing the completeness of paddy rice monitoring methods [50,51]. Integration of Sentinel-1 and Sentinel-2 data greatly increases the temporal frequency of good-quality observations, providing an opportunity to generate high-resolution cropping intensity maps based on phenology analysis [52,53].

4.2. Verification of Consistency between Rice and Wetland Monitoring Areas and Verification Data

The distribution of paddy rice in the study area is extensive, with an average annual paddy rice cultivation area of 11,272 km², accounting for approximately 34% of the total study area. The analysis of the collected official statistical yearbook data of 12 counties and cities in 2019 and the rice area extracted by remote sensing in the corresponding counties and cities showed that the average relative error of double-cropping rice area extracted in this study was 10.1%, and the lowest relative error was only 2.48%. Additionally, the average relative error in the extracted single rice area is 9.52%, with the lowest relative error at 1.12% (

Table 9. Comparison of extraction results of rice area in 2019 with statistical yearbook data.

County	Double Rice			Single Rice
	Extraction Area/km2	Yearbook Area/km2	Error	Extraction Area/km2	Yearbook Area/km2	Error
Nanchang	688.51	643.63	6.97%	228.59	197.76	15.59%
Jinxian	316.53	360.46	−12.19%	38.42	37.99	1.12%
Xinjian	528.66	498.49	6.05%	138.64	128.72	7.70%
Yongxiu	347.14	258.56	34.26%	35.57	46.00	−22.68%
De’an	103.50	104.87	−1.31%	3.30	3.70	−10.81%
Duchang	187.20	190.36	−1.66%	83.18	82.58	0.73%
Zhangshu	414.70	393.89	5.28%	24.55	28.52	−13.91%
Fengcheng	697.39	680.54	2.48%	144.24	114.72	25.74%
Linchuan	87.10	78.50	10.96%	343.78	351.40	−2.17%
Dongxiang	245.98	226.20	8.74%	15.34	14.96	2.54%
Chongren	137.82	157.10	−12.27%	36.47	33.20	9.85%
Jinxi	194.36	163.60	18.80%	73.14	72.10	1.44%

).

The classification categories of land cover products are not completely consistent. However, all of them contain paddy rice subclasses. Therefore, the extraction of paddy rice subclasses was conducted for comparative analysis. Initially, statistical analysis was performed on the paddy rice area within each remote sensing product. Although the paddy rice areas from different data sources were roughly similar, specific numerical differences still existed. Subsequently, a comparative overlay analysis was carried out with World Cover data and GLC_FCS30 data (Figure 7). The results showed that because the spatial resolution of Landsat was lower than that of Sentinel-1/2, the number and boundary clarity of rice patches extracted from Sentinel-1/2 data were higher than those from GLC_FCS30 products. At the same time, some wetland vegetation in GLC_FCS30 and World Cover data were incorrectly classified as rice, but the results of this study more accurately excluded the interference of wetlands on rice classification.

The national single rice planting distribution data set has the problem of misclassification from wetland vegetation, while the national double rice 10-m resolution planting distribution data set incorrectly classified a large number of ridges, resulting in fragmented and irregular results (Figure 8). The data were based on large-scale rice classification across the country, and sentinel-1/2 images within one year were used for land classification, and the impact of wetland and other land types on single and double rice was not considered, resulting in fuzzy and confusing classification details of single and double rice in rice wetland coexistence areas. In contrast, the quality of the classification results in this study is better.

4.3. Uncertainty Analysis of Classification Results

The Poyang Lake Basin exhibits complex geographical features with dense lake distribution, undulating terrain along the lake shores, and presenting hills and ridges. This unique landscape contributes to the relatively scattered nature of land parcels within the basin, resulting in substantial spatial variations in the distribution of rice paddies and wetlands. Despite utilizing Sentinel-1/2 imagery with a 10 m resolution in this study, the issue of mixed pixels persists. In remote sensing imagery, similar types of land may exhibit different spectral characteristics (same object, different spectra), while different land types may share similar spectral features (different objects, similar spectra). This phenomenon is the most significant factor affecting the accuracy of rice paddies and wetland classification [54,55,56].

Furthermore, the Poyang Lake Basin, located in the southern region, frequently experiences cloudy and rainy weather conditions. Despite employing a monthly synthesis method that considers phenological characteristics to construct the dataset, the impact of cloudy and rainy weather could still affect the quality of the remote sensing images. Cloud cover obscures the land surface, reducing the image’s usability, while rainfall might cause reflection and scattering, thereby influencing the process of extracting and interpreting spectral characteristics [57].

In conclusion, the combined challenges of mixed pixels and adverse weather conditions significantly contribute to errors in single-rice and double-rice extraction in the Poyang Lake Basin. In order to mitigate the impact of severe weather conditions, time series analysis and image data fusion technology can be used, which is critical to improving the precision and accuracy of rice field classification under complex environmental conditions.

4.4. Research Limitations and Possible Future Improvements

This study used a random forest classifier to solve the problem of feature identification in rice–wetland coexistence areas, but deep learning contains great potential in remote sensing image classification. Under the condition that the quantity and quality of samples are guaranteed, the robustness and portability of deep learning methods are generally better than random forests. However, the sample size of land type interpretation in this study area is small, the diversity is insufficient, the sample image source is single, and the sample Due to problems such as fixed size, it is impossible to construct a sample data set that meets the requirements. In future research, we plan to make full use of the advantages of semi-automatic sampling methods and combine them with deep learning models to solve the problem of a lack of samples. Doing so will expand the application of deep learning in remote sensing image classification and further improve the accuracy and efficiency of feature recognition in rice–wetland coexistence areas.

5. Conclusions

Based on the GEE cloud platform and Sentinel-1/2 data, a classification sample was constructed using visual interpretation sampling. Combined with monthly synthesis and J-M distance methods for parameter optimization, the study aimed to explore the optimal machine learning classification algorithm for fine classification of vegetation in the coexisting rice–wetland areas of the Poyang Lake Basin. The conclusions are as follows: (1) Considering phenology and remote sensing features, the monthly synthesis and J-M distance methods revealed that rice paddies and wetlands exhibited significant separability in indices such as NDVI, EVI, SAVI, NDWI, MNDWI, and VH. The best separability between rice paddies and wetlands was observed in July and August. The combination of spectral index features during critical phenological periods and radar characteristics effectively addressed the signal mixing problem between vegetation and wetlands. (2) Among the six experimental schemes compared, Sentinel-1 data’s VV and VH polarization characteristics were more sensitive in identifying natural wetlands and paddy rice information. These features provided robust support for classification, with the random forest algorithm demonstrating superior performance compared to CART regression tree and support vector machine algorithms. The optimized random forest algorithm for classification feature parameters displayed good application value and potential for identifying land features in the wetland-rice coexisting area, achieving a maximum OA of 94% and a Kappa coefficient of 0.93. (3) From 2018 to 2022, paddy rice in the Poyang Lake Basin showed the characteristics of planting structure around the Poyang Lake and its tributaries. The mudflats are mainly distributed in the middle and northeast of Poyang Lake, and the wetland vegetation is distributed around the mudflats, forming a nibbling shape from the periphery to the center of the lake. This work proposes a new method for identifying single-rice and double-rice planting areas in rice wetland coexistence areas using Sentinel-1/2 data. The results of this study can support the scientific basis for remote sensing simulation of terrestrial vegetation carbon sequestration of the Poyang Lake basin and present a scientifically supported supplemental foundation for decision-making related to national food security and for achieving the region’s “carbon peaking and carbon neutrality” goals.