Abstract:
In the era of global climate change and rising temperatures, it is more important than ever to develop accurate methodologies for forecasting cyclonic storms. In this art...Show MoreMetadata
Abstract:
In the era of global climate change and rising temperatures, it is more important than ever to develop accurate methodologies for forecasting cyclonic storms. In this article, we design a novel attention mechanism for cyclone intensity estimation based on the perceptual Mach band effect that facilitates boosting and suppression of edges in the input image. To best of the authors’ knowledge, this is the first work to establish a useful connection between the perceptual Mach band effect and the adaptive feature weighing mechanism intrinsic to the attention model deployed within a deep convolutional neural network (CNN). The Mach band attention model (MBAM) aims to amplify or suppress the prominence of feature locations within the convolutional feature space by leveraging attention weights, which are estimated based on the proportion of response normalization-induced Mach band overshoot or undershoot. The MBAM is devoid of any extra model parameters, and the generated attention maps clearly demonstrate that it effectively redirects the attention of the deep attention network to specific discriminative regions such as the shape of the cyclone eye, the eyewall, and associated cloud structures, which are the key salient characteristics indicative of cyclone intensity. We have created a new in-house infrared cyclone image dataset consisting of 15 637 images captured from the INSAT-3DR satellite, particularly focused over the Indian subcontinent region (Bay of Bengal and Arabian Sea). Extensive experimentation of the proposed MBAM-integrated deep attention network demonstrates its superior performance (in terms of mean absolute error (MAE), root mean squared error (RMSE), accuracy, precision, recall, and F1 -score) over the existing methods for very accurate and fast predictions of real-world cyclone severity.
Published in: IEEE Transactions on Instrumentation and Measurement ( Volume: 73)