Deep Learning Semantic Segmentation for Climate Change Precipitation Analysis

Deep learning has achieved many successes and advances over the years for climate data analytics, from simple binary classification tasks to object detection to semantic segmentation. Semantic segmentation is the identification of a class (amongst multiple classes) for every pixel of an image. This can be used, for example, for automated identification of weather and climate events such as atmospheric rivers, tropical cyclones, and weather fronts. We use this automated semantic segmentation to identify different types of weather events and then analyze the precipitation conditioned on these weather events. Further, we analyze the relationships between precipitation and other atmospheric variables, such as temperature and precipitable water, conditioned on these weather events. Following this, we apply these analyses to climate simulations under different climate scenarios in order to better understand the anthropogenic influence on the statistics and dynamics of extreme weather events (the event attribution problem). We compare tropical cyclone precipitation of historical simulations with those from a 2-degree Celsius warmed world from 25-km CAM5.1 model output. Such analyses give us insight into how different climate scenarios impact TC precipitation statistics at any location across the globe. Examining precipitation PDFs conditioned on TCs, we find that TCs tend to have higher levels of extreme precipitation at most locations across the globe. We perform similar analyses for atmospheric rivers. Examining precipitation PDFs conditioned on ARs, we find that ARs tend to have lower levels of precipitation. In conclusion, pixel-wise semantic segmentation allows us to now perform highly precise analytics conditioned on specific types of weather and climate events to answer various questions of great importance in the warming world.