Due to the abilities of all-time and all-weather data acquisition, the microwave remote sensing holds great potentials to identify rice in regions dominated by cloudy and rainy weather. The aim of this study was to analyze the consistency of classification results for single and double cropping rice by using optical and SAR remote sensing data, and then to explore the optimal SAR imagery features for rice classification.

In this study, using the object-based random forest classifier on the Google Earth Engine platform, Sentinel-1 and Sentinel-2 images were adopted to extract the single and double cropping rice from four typical rice growing areas in the Dongting Lake Plain. To analyze the optimal SAR features for the single and double cropping rice identification and the consistency of classification results based on Sentinel-1 and Sentinel-2 images, nine scenarios were established by the combination of different sensors and features and compared the performances of different scenarios. Furthermore, the R² and DTW distance between the NDVI time series and the SAR backscatter coefficient time series (VH, VH/VV) were calculated, respectively.

The overall accuracy of single and double rice cropping identification by using VH, VV and VH/VV time series was 90.42%, 82.08% and 88.33%, respectively. Moreover, the combination of VH and VH/VV time series could achieve a better performance (91.67%) for mapping single and double cropping rice. The derived R² and DTW distance between VH (VH/VV, VV) time series and NDVI time series were 0.870 (0.915, 0.986) and 4.715 (1.896, 5.506) for single cropping rice, as well as 0.597 (0.783, 0.673) and 2.396 (1.839, 3.441) for double cropping rice, respectively. Higher R² and lower DTW distance suggested that VH/VV time series, like NDVI, was more sensitive to the rice growth cycle. Furthermore, the flooding signals in rice transplanting phase could be well captured by VH time series. Additionally, the overall accuracy of single and double cropping rice classification based on optical and SAR features (S-2: NDVI, EVI, LSWI; S-1: VH, VH/VV) in six time windows was 91.25% and 90.00%, respectively, and their consistency was high, with the area correlation of 95.70%.

There was high consistency of classification results for single and double cropping rice based on optical and SAR imagery. Thus, Sentinel-1 imagery held great potentials to identify rice area in cloudy and rainy regions. Specifically, VH and VH/VV backscatter coefficient were optimal features for mapping rice. This study provided vital technical supports for feature optimization by using SAR imagery in cloudy and rainy regions to identify single and double cropping rice accurately.

The decametric-resolution (≤30 m) image is an important data source to identify crop types in South China dominated by fragmented croplands and complex cropping patterns. Due to the relative long revisit frequency of decametric-resolution sensors and persistent rainy/cloud weather in South China, it is critical to integrate multi-source decametric-resolution images using the temporal composition method for the generation of spatiotemporal continuous crop type map. Due to the different temporal resolutions of different satellites, and the significant differences in phenological quaternal rhythms of various crop types, selecting the optimal temporal composition window for integrating multi-source images is vital to map crop type distribution accurately.

This study aims to explore the impact of image temporal composition windows on crop type identification, and to provide significant references for large-scale crop type mapping in regions with complex terrain.

In this study, Landat-8 and Sentinel-2 data were integrated to extract the crop type distribution in the Jianghan Plain, Hubei Province, characterized by the various crop types and cloudy and rainy weather. Then, seven scenarios (15, 20, 25, 30, 40, 50, and 60 d) were established to analyze the effect of different temporal composition windows on crop type identification. Specifically, three aspects, including image coverage rate, spectral-temporal feature curves for different crops and classification accuracies, were combined to understand the performances of different scenarios comprehensively.

The crop type mapping using 20-day composition window performed the best, with the overall accuracy (OA) of 93.13%. In contrast, the scenarios that used narrower temporal composition window derived lower accuracy of crop type identification (e.g., OA=90.91% for the 15-day composition window), which can be primarily attributed to the low coverage rate of good observations in the study area. Meanwhile, since time series images composited in the wide window blurred the key phenological information for different crops, the classification accuracy of crop type mapping scenarios using wide temporal interval was also lower (e.g., OA=86.06% for the 60-day composition window). Additionally, the effect of temporal interval on different crops classification was ranked as following: other crops>rapeseed>rice>wheat>rice-crayfish. In detail, the reason why the classification performance of other crops was the most sensitive to the temporal composition window can be due to the high intra-class phenological variance of this type. Flowering period is the key phenology window to identify rapeseed, therefore, the classification accuracy of rapeseed decreased while the temporal composition window exceeds 30-day, and rapeseed was easily confused with wheat. Furthermore, because the key phenology window to distinguish rice-crayfish from single-cropping rice (i.e., the flooding stage of rice-crayfish fields) lasted a long period (e.g., from October 2020 to June 2021), the classification accuracy of rice-crayfish was less sensitive to the temporal composition window.

In general, the 20-day impact of the temporal composition window can take into account the high-quality image coverage and capture of key phenological characteristics of crop identification, but the optimal temporal composition window of different crops identification is affected by the key phenological period of crops. This study provides theoretical reference and method support for selecting the optimal temporal composition window to generate multi-source image time series, which is promising to improve the efficiency and accuracy of large-scale crop type mapping.