Ratoon rice mapping based on Sentinel-1 and Sentinel-2 imagery

Rice ratooning has gained increasing interest in Asia as a way to boost rice production by allowing two rice harvests from a single growing season. Accurate mapping of this practice can improve rice production estimates. However, current efforts have mainly relied on optical sensors, which are limited by cloud cover, especially during the wet season when ratooning is common. This study systematically assessed the use of optical Sentinel-2, Synthetic Aperture Radar (SAR) Sentinel-1 data and their combination to map ratoon rice crops. Field data were collected in four provinces of the Philippines in 2018-19. Backscatter intensity from Sentinel-1, spectral information, and six commonly used vegetation indices (VIs) from Sentinel-2 were analysed using the Mann-Whitney U significance test to examine differences between the main and ratoon rice crops. Next, we compared the classification performance of decision tree (DT), support vector machine (SVM), and random forest (RF) classifiers. Results show that ratoon and main rice crop significantly differed in VV and VH polarisations, red edge and near-infrared bands, and all VIs. The highest accuracy was achieved with selected features in an RF classifier (overall accuracy of 92%), compared to SVM (87%) and DT (81%). Classification using features from both Sentinel-1 and 2 consistently yielded higher accuracy than using features from one sensor alone. The total planting of ratoon rice was estimated at approximately 223 km 2 (± 4% of the wet season rice area). This study demonstrates the value of combining SAR Sentinel-1 and optical Sentinel-2 for ratoon rice mapping. • Ratoon rice had significantly higher backscatter and lower reflectance in comparison to main rice crops • VV, VH, Red-edge bands, NIR bands, and vegetation indices allowed to distinguish between ratoon and main rice crops • Random forest classifier resulted in a more accurate ratoon rice crop detection than decision tree and support vector machine • Classification accuracy improved using features from both Sentinel-1 and -2 data