A Typhoon Center Location Method on Satellite Images Based on Deep Reinforcement Learning

The location of the typhoon center is a key technology for analyzing and forecasting typhoons as well as reducing typhoon disasters. Varied patterns, blurred borders, and uncertain centers of the typhoons in the satellite images make the task challenging. In this work, a method based on deep reinforcement learning is proposed to locate the typhoon center. It turns the problem of typhoon center locating into an issue of searching and views the process of searching for the typhoon center as a series of Markov Decision Processes. Specifically, the agent learns to move and reduce a search box using simple transformation actions to make the search box center close to the typhoon center. Meanwhile, discriminant rules for detecting typhoons are designed based on the relationship of selected action and its value function. Experimental results demonstrate that the proposed method can complete the locating of typhoons of different grades and types in about 16 steps. The recall of the proposed method increases with the typhoon level and reaches 100% at grade 5 and 6, and the average precision reaches 91.85%. The proposed method can effectively locate typhoon centers in different forms with a mean absolute error of 0.265°, which is superior to the comparison algorithms.