A Modular Processing Chain for Automated Flood Monitoring from Multi-Spectral Satellite Data

Emergency responders frequently request satellite-based crisis information for flood monitoring to target the often-limited resources and to prioritize response actions throughout a disaster situation. We present a generic processing chain that covers all modules required for operational flood monitoring from multi-spectral satellite data. This includes data search, ingestion and preparation, water segmentation and mapping of flooded areas. Segmentation of the water extent is done by a convolutional neural network that has been trained on a global dataset of Landsat TM, ETM+, OLI and Sentinel-2 images. Clouds, cloud shadows and snow/ice are specifically handled by the network to remove potential biases from downstream analysis. Compared to previous work in this direction, the method does not require atmospheric correction or post-processing and does not rely on ancillary data. Our method achieves an Overall Accuracy (OA) of 0.93, Kappa of 0.87 and Dice coefficient of 0.90. It outperforms a widely used Random Forest classifier and a Normalized Difference Water Index (NDWI) threshold method. We introduce an adaptable reference water mask that is derived by time-series analysis of archive imagery to distinguish flood from permanent water. When tested against manually produced rapid mapping products for three flood disasters (Germany 2013, China 2016 and Peru 2017), the method achieves >= 0.92 OA, >= 0.86 Kappa and >= 0.90 Dice coefficient. Furthermore, we present a flood monitoring application centred on Bihar, India. The processing chain produces very high OA (0.94), Kappa (0.92) and Dice coefficient (0.97) and shows consistent performance throughout a monitoring period of one year that involves 19 Landsat OLI (<mml:semantics>mu Kappa=0.92</mml:semantics> and <mml:semantics>sigma Kappa=0.07</mml:semantics>) and 61 Sentinel-2 images (<mml:semantics>mu Kappa=0.92</mml:semantics>, <mml:semantics>sigma Kappa=0.05</mml:semantics>). Moreover, we show that the mean effective revisit period (considering cloud cover) can be improved significantly by multi-sensor combination (three days with Sentinel-1, Sentinel-2, and Landsat OLI).