Abstract

Synthetic aperture radar (SAR) remote sensing offers a flexible approach and brings the opportunity to collect crop information that is not limited by weather conditions. The applicability of Sentinel-1 SAR data with dual-polarization enables the identification of individual rice fields, and with sufficient repeatability to monitor the growth status of different crops. In recent years, with the continuous development of machine learning algorithms, deep learning in the world, especially convolutional neural networks (CNN), has obtained good results in detecting and extracting information on remote sensing images. In this study, we propose a classification model based on deep convolutional neural network (DCNN) to extract rice fields from  sentinel-1 SAR data. Physical indices were calculated from (VH and VV) polarization and a mobile team examined the growth
morphology of rice plants. The results were checked using visual field data with the overall accuracy, and cross-validation values of the rice parameters extracted were higher than 0.85. The accuracy of rice biomass estimation reached (R2=0.79, RMSE=0.12 kilograms) for the Winter-Spring crop and (R2=0.77, RMSE=0.15 kilograms) for the Summer-Autumn crop. The results showed that Sentinel1 data could map the spatial distribution of retrieved rice biomass in various weather conditions. The integrated methodology framework developed in this study can be applied to rice fields across Vietnam and similarly rice fields in the world.