Combining Sentinel-1 and Sentinel-2 data for improved land use and land cover mapping of monsoon regions

• Published in: International journal of applied earth observation and geoinformation Vol. 73; pp. 595 - 604
• Main Authors: Steinhausen, Max J., Wagner, Paul D., Narasimhan, Balaji, Waske, Björn
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2018
• Subjects: Cloud-prone; India; Land use and land cover classification; Multi-sensory; Random forest; SAR; Sentinel-1; Sentinel-2; SAR; Sentinel-1; Sentinel-2; Multi-sensory; Cloud-prone; Random forest; Land use and land cover classification; India
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2018.08.011

•Combined Sentinel-1 and 2 data can improve land cover mapping in cloud-prone regions.•One additional radar scene already improves the accuracy significantly.•The radar data time series should represent the entire cropping season.•The classification of paddy fields is greatly improved by additional radar data. Land use and land cover maps can support our understanding of coupled human-environment systems and provide important information for environmental modeling and water resource management. Satellite data are a valuable source for land use and land cover mapping. However, cloud-free or weather independent data are necessary to map cloud-prone regions. This particularly applies to monsoon regions such as the Chennai basin, located in the north of Tamil Nadu and the south of Andhra Pradesh, India, which is influenced by the South Asian Monsoon and has abundant cloud cover, throughout the monsoon season. The Basin is characterized by small-scale agriculture with multiple cropping seasons and the rapidly developing metropolitan area of Chennai. This study aims to generate a land use and land cover map of the Chennai Basin for the cropping season of Rabi 2015/16 and to assess the influence of combining the new ESA Copernicus satellites Sentinel-1 and -2 on classification accuracies. A Random Forest based wrapper approach was applied to select the most relevant radar (Sentinel-1) images for the combination with the optical (Sentinel-2) data. Area proportion weighted accuracy with 95% confidence interval were estimated for the Random Forest models, which differentiated 13 land cover classes. The highest overall accuracy of 91.53% ± 0.89 pp was achieved with a combination of 1 Sentinel-2 and 8 Sentinel-1 scenes. This is an improvement of 5.68 pp over a classification with Sentinel-2 data only. An addition of further Sentinel-1 scenes showed no improvement in overall accuracy. The strongest improvement in class-specific accuracy was achieved for paddy fields. This study shows for the first time how land use and land cover classifications in cloud-prone monsoon regions with small-scale agriculture and multiple cropping patterns can be improved by combining Sentinel-1 and Sentinel-2 data.