Abstract

Dual-polarized (VV and VH) Sentinel-1 Synthetic-Aperture Radar (SAR) Ground Range Detected (GRD) data are available in 9-m spatial resolution and 12-day repeat orbit. A constellation of two satellites, Sentinel 1A and Sentinel 1B, capture these data with ascending and descending orbits, thus increasing the revisit time at the equator to every six days. Those specifications allow creating dense cross-orbit time-series data with a relatively high spatial resolution, beneficial for identifying land-covers and land-uses with unique temporal dynamics, such as paddies. This study was intended to assess the accuracy of time-series dual-polarized cross-orbit Sentinel 1A and 1B GRD data for mapping paddy extents. The monthly median value of these data was processed in Google Earth Engine and used as inputs in the paddy identification in Magelang District using bagging random forests (RF) and extreme gradient boosting (XGB) algorithms. Variables were ranked based on importance and selected using recursive feature elimination (RFE) and RF model to reduce the data dimensionality and understand the variable importance corresponding to a different month of the year. The resulting variable importance demonstrates better contributions of VV polarization and ascending orbit to the mapping model, and the producer’s and user’s accuracies achieved by RF classifier were 75% and 93.9%. For these reasons, an ascending (ASC) dataset provides better accuracy than its descending (DSC) counterpart and the combination of both (ASC+DSC). The user’s accuracy of the paddy identified using the RF model with ascending Sentinel 1-data is 4% and 6% higher than the XGB models built using ASC and cross-orbit (ASC+DSC) datasets, respectively.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call