Leveraging convolutional neural networks for semantic segmentation of global floods with PlanetScope imagery

Constellations of small satellites are able to achieve the coverage, revisit rate, and spatial resolution needed to generate high resolution global maps of flood waters. The PlanetScope constellation achieves daily coverage at 3-4 m resolution; however, the limited spectral channels available as well as radiometric challenges inherent to large constellations of small satellites create a challenge to accurate global flood mapping. Here, we present a convolutional neural network (CNN) which combines PlanetScope imagery with information about multidecadal water dynamics from the Global Surface Water dataset to create a globally applicable water segmentation model. We also compare this model to locally-trained Random Forest models, which are commonly used for flood mapping. The CNN achieves a mean intersection over union score of 58.1% +/- 9.84% across 7 biomes when evaluated on flooded areas outside of permanent water bodies, and 70.3% +/- 11.6% when evaluated on all water pixels.