Paddy rice is one of the world’s most important staple food crops, feeding half of the population (Kuenzer and Knauer, 2013). Globally, the planting area of rice reached 161 × 10⁶ ha in 2018, accounting for 15% of the planting area of main crops, and 89.8% of paddy rice was planted in Asia (FAO Statistical Databases, 2018). As the world’s population continues to increase, the importance of rice as a food supply increases (Qin et al., 2015; Karthikeyan et al., 2020). Due to its special physiological structure, rice needs to be grown in flooded soils; therefore, rice cultivation has a crucial impact on the world’s freshwater resources (Bouman and Tuong, 2001; Bouman et al., 2007; Dong and Xiao, 2016). In addition, rice fields are one of the main sources of atmospheric methane, which contribute 12%–26% of global anthropogenic CH₄ emissions and play an important role in global climate change (Sass et al., 1999; Dong and Xiao, 2016; Canadell and Monteiro, 2021). Therefore, timely and accurate information on rice planting areas is crucial for maintaining food security and freshwater resource stability and studying climate change. In addition, China has a vast rice planting area, and there are significant differences in temperature during the rice growing season in different regions. The impact of climate change on China’s rice yield will also have significant spatiotemporal differences (Saud et al., 2022). Therefore, understanding the exact growth location of rice is crucial for assessing the impact of climate change on rice yield.

Satellite-based methods are increasingly becoming the primary means for identifying and monitoring crop distributions, as they can provide accurate and timely information on crop phenology and growth (Atzberger, 2013; Zhang X. et al., 2018; Karthikeyan et al., 2020; Zhan et al., 2021). Most existing studies uses two types of remote sensing data: optical data and SAR data (Joshi et al., 2016; Zhao et al., 2021). Optical remote sensing data has been widely used to study the spatial distribution of rice at regional and local scales. For example, Jin et al. (2016) used normalized difference vegetation index (NDVI), land surface water index (LSWI), and enhanced vegetation index (EVI) data from multi-temporal Landsat to produce rice maps of the Sanjiang Plain in northeast China from 2010 to 2012, with a resolution of 30 m and user’s and producer’s accuracy of 90% and 94% respectively. Liu et al. (2018) established a paddy rice map of northeast China with high accuracy basing on a sub-pixel method using LSWI time series of MODIS products.

Although it is possible to use optical data for mapping rice in northeast China, where the longer growing seasons and fewer clouds allow for sufficient high-quality data availability (Dong et al., 2016), the frequent clouds and rain of the southeast make it much more challenging (Zhan et al., 2021). In this case, the use of SAR data is more effective. SAR signals are able to operate day and night independently of weather conditions (Rogan and Chen, 2004; Zhu et al., 2012). Early studies have found that SAR can accurately capture the flooding of rice, as the radar backscatter signal of rice fields is low during the transplanting stage (Kurosu et al., 1995; Shao et al., 2001). Kurosu et al. (1995) first tried to use ERS-1 C-band SAR data to monitor rice growth and found that SAR data could suitably capture the signal of the rice growth period. Later, common-polarization backscatter (e.g., HH) and common-polarization ratio (e.g., HH/VV) images became more common in rice mapping due to their high correlation with rice growth cycles (Bouvet et al., 2009). A recent study used VH backscatter data provided by Sentinel-1 to map rice in the Mekong Delta in 2015 and found that the VH backscatter is more sensitive than the VV backscatter to capture changes in the growth period of rice, with an overall accuracy of 87.2% (Nguyen, et al., 2016). Sun et al. (2023) produced a 20 m resolution rice region map of five countries in Southeast Asia in 2019, with an overall accuracy of 92.2%. Carrasco et al. (2022) created seven rice field maps of Japan from 1985 to 89, 1990 to 94, 1995 to 99, 2000 to 04, 2005 to 09, 2010 to 14, and 2015 to 19, with an overall accuracy of 83%–95%. Other studies have combined optical and SAR data for rice identification (Onojeghuo et al., 2018; Chen et al., 2020; He et al., 2021). For example, a phenology-based rice mapping study in Peninsular Malaysia using VH (capturing low backscatter signals in the rice transplanting stage) and NDVI (reflecting the differential growth conditions from rice growth to maturity) time series data of Sentinel-1/2 satellites has achieved good results with an overall accuracy of 95.95% (Fatchurrachman et al., 2022).

In the past few decades, a plethora of remote sensing-based crop identification methods, such as threshold-based methods (Bazzi et al., 2019; Li et al., 2020; Wei et al., 2022), phenology-based methods (Xiao et al., 2002; 2006; Xiao et al., 2021; Kobayashi and Ide, 2022), and machine learning-based methods have been developed (Ndikumana et al., 2018; Zhong et al., 2019; Chen et al., 2020). For instance, in a rice mapping study based on SAR data, researchers collected 50,000 ground sample sites and used a feature-based decision method to map the distribution of rice in Yunlin and Changhua counties in central Taiwan in 2017, with an overall accuracy of 91.9% (Chang et al., 2020). Fiorillo et al. (2020) collected about 400 sample sites per year from 2017 to 2019 and mapped the spatial distribution of rice in Senegal using Sentinel-1/2 imagery based on the random forest method, with a maximum accuracy of 87% and a kappa coefficient of 0.8. Despite the existence of different methods for rice mapping, most of them require massive samples and rely heavily on the selection of training samples. Obtaining large amounts of real-time samples is time-consuming and labor-intensive, which limits the application of these methods to update rice maps in real time over large areas (Mosleh et al., 2015; Dong et al., 2020).

As the world’s largest rice producer, China had 30 × 10⁶ ha of rice planted area in 2018, accounting for 18% of the world’s total rice sown area (FAO Statistical Databases, 2018). Particularly the sown area of double-season rice in southern China was 10 × 10⁶ ha, accounting for one-third of China’s total sown area, making it an important agricultural production area in China (National Bureau of Statistics of China, 2019). In hilly areas of southern China, the mixed patterns of cultivation and high fragmentation of cultivated land, resulting from abundant rainfall and heat, pose challenges for identifying rice planting areas (Qiu et al., 2003; Pan et al., 2021; Wei et al., 2022; Zheng et al., 2022). The backscattering coefficient of rice growth showed a V-shaped growth curve and the number of flood signals varied across rice cultivation systems (Pan et al., 2021). The VH time series of single-season rice theoretically shows a V-shaped valley, while double-season rice shows two. The accurate identification of flood signals is critical for mapping rice. However, due to the short growth period of double season rice and the weak flood signal intensity of late rice, it is challenging to fully capture the signals of two V-shaped valleys, making it difficult to identify double season rice using traditional methods (Xu et al., 2023).

In order to solve the problem of traditional machine learning relying on samples for identifying rice, we have developed a rice index that can be used to identify double season rice in a large area without the need for samples. In this study, we used a phenology-based sample-free identification method to map rice planting areas using time series SAR data provided by Sentinel-1. Using the proposed method, we mapped double-season rice in nine provinces in southern China (accounting for 99% of the total area) in 2018 with a spatial resolution of 10 m. To validate the accuracy of rice identification, we used county level statistics data obtained from various municipal statistical bureaus for regional scale verification. In the main producing provinces of double-season rice, field survey data were used to verify the accuracy at the pixel scale. Our results indicate that the proposed method is feasible for rice identification on a large scale without using samples, which is very helpful for large-scale crop recognition updates and backtracking identification.

We produced a spatial distribution map of double-season rice at 10 m spatial resolution in nine provinces of China in 2018 using the SPRI method (Figure 3). To verify the stability and reliability of our method, we first conducted a verification in the four major producing provinces (85% of China’s total double-season rice area) using both GPS survey data and UAV farmland data. We randomly selected sample points in the UAV survey area and combined all GPS survey samples to assess pixel-level accuracy. Based on double-season rice and non-double-season rice survey samples, the overall identification accuracy across the four provinces ranged from 81.02% to 90.05% for early rice (Table 1) and between 79.7% and 88.89% for late rice (Table 2). Hunan is one of the largest double-season rice planting provinces in China, accounting for more than 28.38% of China’s double-season rice planting areas. The user’s, producer’s, and overall accuracy of early (late) rice in Hunan were 90.33% (76.91%), 88.28% (86.27%), and 90.05% (85.01%), respectively. Jiangxi is also one of the main double-season rice planting areas in China, with a planting area accounting for 25.28% of double-season rice in China. The user’s, producer’s, and overall accuracy of early (late) rice in Jiangxi were 73.84% (73.85%), 80.19% (86.56%), and 82.33% (79.70%), respectively.

Subsequently, the agricultural statistical data we gathered was utilized to assess the accuracy at the county level across all examined provinces (Figure 4). Our method exhibited satisfactory performance in accurately identifying both early and late rice within regions where county-level data was available. The statistical metrics, namely, R ², RMSE, and RMAE, were calculated to quantify the accuracy. For early rice, the R ², RMSE, and RMAE values were determined as 0.77, 4.52 (10³ ha), and 0.35, respectively. Notably, the performance for late rice was even better, with R ², RMSE, and RMAE values of 0.82, 4.29 (10³ ha), and 0.30, respectively.

For each province, the R ² between the identified and agricultural statistical areas of early rice ranged from 0.63 to 0.82. The RMSE ranged from 1.21 to 5 .50 (10³ ha), and RMAE ranged from 0.27 to 0.58 (Figure 5). For late rice in each province, the method performed better in some areas, with R ² ranging from 0.59 to 0.88, RMSE ranging from 1.72 to 6.12 (10³ ha), and RMAE ranging from 0.20 to 0.55 (Figure 6). The accuracy of late rice identification in Hunan, Guangdong, Guangxi, Anhui, Hainan, and Hubei provinces was higher than that of early rice. For example, in Guangxi, the R ², RMSE, and RMAE of late rice were 0.78, 4.37 (10³ ha), and 0.30, respectively while the R ², RMSE, and RMAE of early rice were 0.63, 5.14 (10³ ha), and 0.43, respectively. For the two largest double-season rice cultivation areas (Hunan and Jiangxi) in China, the method showed good performance between identified and agricultural statistical areas for both early and late rice, where the R ², RMSE, and RMAE of early rice were 0.76, 5.50 (10³ ha) and 0.28 in Hunan, and 0.82, 4.94 (10³ ha), and 0.36 in Jiangxi, respectively. For late rice, the R ², RMSE, and RMAE were 0.88, 3.96 (10³ ha), and 0.20 in Hunan, and 0.80, 6.12 (10³ ha), and 0.37 in Jiangxi, respectively. Although there was a high R ² (0.64) between identified and agricultural statistical areas in Zhejiang, the identified area noticeably underestimated agricultural statistical areas. Zhejiang had the worst verification performance at the county level where the RMAE was 0.57 for early rice and 0.54 for late rice, with severe omission errors in areas where statistics were available.

In this study, we produced double-season rice maps with a spatial resolution of 10 m in southern China in 2018 based on the SPRI method. Validation using field sample data and agricultural statistics showed that the produced double-season rice maps have high accuracy at both the pixel and regional scales.

High spatial resolution rice maps can capture the distribution of rice fields more precisely and reduce the impact of mixed pixels on recognition accuracy (Xiao, 2003; Zhang et al., 2009). Compared to the 500 m resolution maps of Xiao et al. (2005) and Sun et al. (2009), the 10 m spatial resolution rice map produced in this study more accurately reflects the distribution of rice fields in southern China, especially in hilly areas with high fragmentation of planting. In addition, our method can automatically identify rice planting areas without sampling points, and has achieved good accuracy in major producing provinces. Compared with traditional machine learning algorithms (Zhang M. et al., 2018; Bazzi et al., 2019; Cai, et al., 2019; Zhang et al., 2020), our method also saves a lot of time and labor (Dong et al., 2020). Therefore, our method is more flexible in both large-scale applications and long-term retrospective mapping research. Recently, Pan et al. (2021) used the time-weighted dynamic time warping (TWDTW) method to produce the first large-scale maps of double-season rice in China, with a spatial resolution of 10 m. The overall accuracy of the map for four major producing provinces (HuN, JX, GD, GX) were 90.5%, 91.88%, 88.07%, and 88.17% for early rice, and 90.36%, 89.18%, 88.25%, and 88.40% for late rice, respectively. While our accuracy is slightly lower than that reported by Pan et al. (2021), we achieved satisfactory results without the need for any samples. Other studies have also used machine learning to classify double-season rice in smaller areas of Hunan Province and Jiangxi Province. He et al. (2021) achieved an overall accuracy of 85% and 96% for early and late rice, respectively, while Tian et al. (2018) achieved user’s accuracy of 97.9% and 96.3% for early and late rice, respectively. However, these studies were limited to regional-scale applications, and their results were uncertain when applied on a large scale due to the limitations of sample size.

Although our method can effectively and accurately identify large-scale rice areas, there are still some uncertainties in the identification process. Data quality affects the accuracy of classification. First, while SAR data is not affected by cloud or lighting conditions, its inherent noise makes classification results uncertain, and SAR backscattered signals are damaged by topographic effects, even when they are radiometric terrain corrected (Steele-Dunne et al., 2017; Fatchurrachman et al., 2022). Pan et al. (2021) found that rice paddies in mountainous areas are small and scattered, especially in Fujian province, where 64% of rice fields are in mountainous areas. Affected by topography, our classification results in Fujian province were relatively unsatisfactory, in which the RMAE of early and late rice was 0.45 and 0.46 respectively. Secondly, we used two optical indices, NDVI and NDWI, to determine the area of vegetation coverage and temporary water bodies. The W-line value is the value at the bottom 10% of the minimum backscattering value in all temporary water bodies, while the V-line value is the value in the top 10% of the maximum backscattering value in the vegetation-covered area. In southern China, due to the influence of the cloudy and rainy weather, the number of effective optical images is insufficient, which affects the determination of the final value of W and V lines, and ultimately the recognition accuracy (Wu et al., 2011). Finally, we calculated the quantity and proportion of data used for county-level validation (Figure 7). Our results performed poorly in Zhejiang Province, possibly due to the limited data available for county-level validation and the low rice planting area in the validation area, which cannot represent the identification of the main production area.

In this study, a map of double season rice in southern China for the year 2018 was generated using a sample-free automatic method at a large scale. The study further validated the feasibility and adaptability of this method for large-scale applications, establishing its reliability and effectiveness. Our proposed method mainly distinguished rice from other crops by capturing and amplifying characteristics during growth. In future work, our method can be applied to the identification of other crops.

China is the world’s largest rice producer, accounting for 18% of the world’s rice-sown area in 2018. In this study, we used SAR data from the Sentinel-1 satellite to map double-season rice in southern China in 2018 based on SPRI, a sample-free identification method. In the main producing provinces, we use randomly generated samples based on UAV survey images and all samples from handheld GPS field expeditions for verification. The highest recognition accuracy of early rice was recorded in Hunan Province, where the producer’s accuracy, user’s accuracy, and overall accuracy of early rice were 88.28%, 90.33%, and 90.05%, respectively. By comparison, Guangdong Province had the highest recognition accuracy of late rice, in which the producer’s accuracy, user’s accuracy, and overall accuracy were 88.49%, 90.63%, and 88.89%, respectively. In addition, the sown area of early and late rice determined in this study showed a high correlation with the county-level agricultural statistics from the statistical bureaus of each city. The early and late rice areas of all surveyed provinces were compared with the existing county-level agricultural statistics, where the R ² of early and late rice were 0.77 and 0.82 respectively. More importantly, using the SPRI method for rice identification does not require ground samples, and larger-scale rice mapping can be performed by adjusting the W and V lines of rice in different regions. The results of this study provide support for subsequent research on rice yield estimation, and the method used in this study enables years of rice retrospective mapping.