Comparison of Methods for Crop Classification and Rice Extraction on Chongming Island Based on Sentinel-1 and Sentinel-2 Data

The work employed the Google Earth Engine Cloud Computing Platform for the extraction of crop spatial distribution in the Chongming Island, Shanghai, China, where a novel approach for integration of Sentinel-2 optical data and Sentinel-1 radar data was developed. The precision and reliability of the derived results were evaluated by a total of 521 ground samples. In the process of image segmentation and classification, a fusion of diverse vegetation indices was utilized for effective categorization. The optical image classification was conducted using the random forest method, ensuring robust and accurate classification results. Concurrently, in the context of temporally dense radar data, stable land patches corresponding to crop growth were identified to extract the time series of radar backscatter coefficients. Relevant rules was established for the deep learning to extract the distribution of rice in spring and autumn periods. The findings demonstrated that the fusion of optical remote sensing imagery and radar data facilitates accurate delineation of agricultural fields in Chongming Island, enabling a comprehensive classification of 13 classes and 8 distinct agricultural categories. The classification outcomes for both spring and autumn seasons exhibited satisfactory results, characterized by Kappa coefficients surpassing 0.81. Notably, the autumn dataset demonstrated superior performance in terms of accuracy and reliability. Through multi-temporal feature analysis of Synthetic Aperture Radar (SAR) data, combined with band features and spatial overlay analysis, a backscattering lookup table was established and applied for extracting the spatial distribution of the predominant crop and rice with a precision rate over 93%.