Satellite-based crisis information is frequently requested in the context of flood disasters to gain rapidly situational awareness and to prioritize response actions under often limited resources during emergency response. To assure that information products have the highest possible spatial, temporal and thematic resolutions, it is critical to be able to simultaneously analyse data from a large variety of satellite sensors. In this contribution, we present a solution to rapidly extract water bodies from Landsat TM, ETM+, OLI and Sentinel-2 for up-to-date situational awareness during emergency response. A convolutional neural network is used to segment water extent in these images, while clouds, cloud shadows and snow / ice are specifically handled by the network to remove potential bias from any downstream analysis. Atmospheric correction, post-processing and ancillary data are not required. To distinguish flood from permanent water we present a reference water mask that is derived by means of time-series analysis of archive imagery. Compared to widely-used mono-temporal reference water masks, it can be adapted to any area and time of interest. This study builds up on previous work of the authors and presents new results from recent flood disasters in Germany, Peru, China, India and Mozambique, as well as a flood monitoring application centred on the Indian state of Kerala. The processing chain produces very high overall accuracy and Kappa coefficient (0.87) and shows consistent performance throughout a monitoring period of 12 months that covers 143 Landsat OLI and Sentinel-2 images.
展开▼
