Boost Cyclone Intensity Estimation via Processing Motion Information and Embedding Rotation Invariance

Tropical cyclones (TCs) in tropical and subtropical regions often evolve into severe natural disasters, causing significant economic and human losses. Accurate and timely TC intensity information can better initialize TC numerical prediction models, thereby improving the accuracy of TC intensity and path forecasts. To this end, we propose a TC intensity estimation model based on convolutional neural networks. To enhance the accuracy of estimating TC intensity from satellite remote sensing channels, such as far-infrared (IR) and mid-IR, the model attempts to learn and process cloud and water vapor motion information and incorporate rotation invariance. In the motion information processing module, we employ a dual-branch design. One branch directly processes motion information from the pixel space, while the other branch first computes motion matrices between different time steps to map the features from the pixel space to the motion space, then cascades multiple downsampling modules for encoding and compressing the information. To ensure that the feature extraction module of the network possesses rotation invariance, each basic convolution operation is replaced with a rotation invariant convolution operation. In terms of mean absolute error and root-mean-square error, our model shows an improvement of over 10% compared to previous “state-of-the-art” models. Notably, in terms of MedAE, our model achieves an improvement of up to 17% .