Exploring the use of random forest classifier with Sentinel-2 imagery in flooded area mapping

Optical satellite sensors represent a reference for Earth imaging applications, including land monitoring and flood management, directly allowing the visual interpretation of acquired scenes or the exploitation of surfaces' spectral signatures. Simultaneously, the advancements in machine learning (ML) methods led to a proliferation of supervised and unsupervised algorithms applied to classification problems in the field of flood hazard and risk mapping. In particular, random forest (RF) is a powerful and widespread ML algorithm applied in remote sensing for image classification. In the present study, the ability of two RF models for flood detection using Sentinel-2 imagery was explored. The models were developed considering spectral bands at 20m resolution in combination with one or more multispectral indices, namely the ratio between the two SWIR bands, the Modified Normalized Difference Water Index (MNDWI), the Normalized Difference Moisture Index (NDMI), and the Red and Short-Wave Infra-Red (RSWIR) Index. The investigation was carried out to map the extent of a flood event that occurred along the Sesia River (Vercelli, Italy) in October 2020. The performance of the two RF models was assessed using the maps delivered by the Rapid Mapping service of the Copernicus Emergency Management Service (CEMS) as a reference. Results revealed some very interesting findings regarding the performances of the examined methods, underlying the added value of Sentinel-2 and the RF proficiency in accurately identifying flood-affected regions.