Monitoring flood extent in the lower Amazon River floodplain using ALOS/PALSAR ScanSAR images



Arnesen AS, Silva TSF, Hess LL, Novo EMLdM, Rudorff CM, Chapman BD & McDonald KC (2013) Monitoring flood extent in the lower Amazon River floodplain using ALOS/PALSAR ScanSAR images. Remote Sensing of Environment, 130, pp. 51-61.

The Amazon River floodplain is subject to large seasonal variations in water level and flood extent, due to the large size and low relief of the basin, and the large amount of precipitation in the region. Synthetic Aperture Radar (SAR) data can be used to map flooded area in these wetlands, given its ability to provide continuous information without being heavily affected by cloud cover. As part of JAXA's Kyoto & Carbon Initiative, extensive wide-swath, multi-temporal SAR coverage of the Amazon basin has been obtained using the ScanSAR mode of ALOS PALSAR. This study presents a method for monitoring flood extent variation using ALOS ScanSAR images, tested at the Curuai Lake floodplain, in the lower Amazon River, Brazil. Twelve ScanSAR scenes were acquired between 2006 and 2010, including seven during the 2007 hydrological year. Water level records, field photographs, optical images (Landsat-5/TM and MODIS/Terra and Aqua) and topographic data were used as auxiliary information. A data mining algorithm allowed the implementation of a hierarchical, object-based classification algorithm, able to map land cover types and flooding status in the study area for all available dates. Land cover based on the entire time series (classification levels 1 and 2) had overall accuracies of 90% and 83%, respectively. Level 3 classifications (one map per image date) were validated only for the lowest and highest water stages, with overall accuracies of 76% and 78%, respectively. Total flood extent (Level 4) was mapped with 84% and 94% accuracies, for the low and high water stages, respectively. Regression models were fitted between mapped flooded area and water levels at the Curuai gauge to predict flood extent. A polynomial model had R2 = 0.95 (p < 0.05) and an overall root mean square error (RMSE) of 241 km2, while a logistic model had R2 = 0.98 (p < 0.05) and RMSE = 127 km2.

Object-based image analysis; Multi-temporal analysis; Incidence angle; Wetlands; Synthetic aperture radar; Kyoto & Carbon Initiative

Remote Sensing of Environment: Volume 130

FundersBrazilian National Research Council
Publication date15/03/2013
Publication date online17/12/2012
Date accepted by journal27/10/2012

People (1)


Dr Thiago Silva

Dr Thiago Silva

Senior Lecturer, Biological and Environmental Sciences