Detecting Marine Heat Waves using Deep Learning

Abstract:

Marine heat waves (MHWs) are oceanic extreme events that last from days to months and can have devastating impacts to our environment and society. The frequency of occurrence and intensity of MHWs is expected towill increase in the future under the pressure of anthropogenic climate change. Sea level pressure and wind variability have been suggested as potential triggers for some major MHWs. Therefore, identifying the atmospheric conditions that precede MHWs will substantially improve our predictive skill; deep learning techniques hold promise for such predictions. Here, we present an initial effort for neural network (NN) application towards real-time forecasting of MHWs. Using the community-accepted definition of MHWs, we have generated a database of 57 million MHW events from the NOAA OI SST dataset. These are labeled by duration and intensity and provide the primary sample for training. After generating these labels, a deep NN is trained to identify and predict MHW onset and severity based on similarity with past occurrences. By modeling our input as a rank 3 tensor with sea level pressure, wind direction, and strength as the channels we can approach MHW detection as a computer vision problem and apply standard frame prediction techniques. We utilize a U-Net to produce future segmented states given the current state, and evaluate these predictions using a variety of metrics. The U-Net is trained to forecast the presence or absence of a MHW based on the atmospheric conditions occurring days before.