Leveraging machine learning to improve satellite flood detection: seeing through (thin) clouds and identifying features

Abstract:

Machine learning algorithms are increasingly used to identify clouds, shadows, and flood water from satellite imagery. Pixel-based machine learning algorithms (e.g. random forests) are commonly employed to identify flood water for both radar and optical satellites (e.g Shen et al 2019). Deep learning algorithms (e.g. fully convolutional neural networks) that include spatial context in models can identify flooded features and yield improved accuracy. New public datasets (e.g. Sen1Floods11, Bonafilia et al, 2020) are available to train radar-optical deep learning fusion models. Here we show how Cloud to Street is using machine learning algorithms to improve flood maps for near real time monitoring. We train a random forest classifier sampled globally across biomes, urban development density, and cloud conditions to detect water from Sentinel-2 and Landsat. We test classifier accuracy to identify permanent water versus flood water for each sensor. Preliminary results on Landsat 8 show random forest models outperform the Automated Water Extraction Index (AWEI) in clear pixels, and skillfully identify flood pixels under thin clouds and in cloud shadows (>80% accuracy) where most thresholding algorithms fail. Feature importance analysis reveals NIR and SWIR signals penetrate thin clouds and vary within water under shadows to enable classification. The ability to map floods under clouds increases “mappable” area by up to 32% during major flood events in Sri Lanka, for example. We will compare variance of flood detection accuracy with machine learning validated against over 4,000 points of critical assets over rainy seasons in Sri Lanka and the Eastern Nile Basin. Finally, we compare flood maps generated with random forest to deep learning algorithms such as fully convolutional neural networks for Sentinel-1 and Sentinel-2. We conclude with a research agenda to leverage advances in computer vision for improved operational flood mapping.