Towards operational multi-resolution monitoring of water bodies from optical satellite images

Abstract

Given current population growth rates and the increasingly visible effects of climate change on the environment and human activities, continuous monitoring of the spatio-temporal distribution of surface water bodies is becoming ever more important. This becomes particularly prominent in emergency response applications for which near-real time information about flood water extent and duration are crucial components to target often limited resources and prioritize response actions. To assure that the delivered information products have the highest possible spatial, temporal and thematic resolutions, it is critical to simultaneously harvest data from a large variety of satellite sensors. In this contribution, we present an automated processing chain that covers all modules required for operational large-scale surface water monitoring at different spatial and temporal resolutions. This includes data ingestion and preparation, cloud / cloud shadow masking, extraction of water bodies and the preparation of thematic information products. All modules are specifically designed to be compatible with a large range of available optical satellite sensors. Cloud / shadow masking is performed with a globally trained convolutional neural network. Water bodies are mapped with a semi-supervised stochastic gradient descent classifier using training seeds which are automatically identified for each image scene by sampling a Normalized Difference Water Index (NDWI) image. Compared to previous work in this direction, our method is purely data-driven and parameters are dynamically learned from the image to adapt to the sensor specific feature domain. Cloud / shadow masking and water mapping modules aim at being robust to radiometric, atmospheric and geometric variations across scenes and sensors. Hence, we evaluate the performance of the water mapping method against a globally distributed test dataset derived from Landsat TM, ETM+ and OLI, Sentinel-2 and RapidEye images, and compare it to a widely used water index thresholding method. Furthermore, we present the application of our processing chain to surface water monitoring in the North-Eastern Indian state of Bihar, which is seasonally affected by flooding due to monsoon rain. In this context, we introduce a novel reference water layer that is dynamically derived for any given area and time period of interest. Combining the reference water layer with water masks allows distinguishing permanent water bodies from temporarily flooded areas. The results of our processing chain feed into a larger automatic flood monitoring system that jointly uses Synthetic Aperture Radar (SAR) and optical products to provide valuable information for emergency response and for an index-based flood insurance that aims at strengthening the resilience of vulnerable people in the region.