The Xiamen Symposium on Marine Environmental Sciences (XMAS 2027)

Tropical cyclones (TCs) occurrence can reorganize the surrounding mean sea level pressure (MSLP) field, but whether this process is encoded in climate network remains unclear. Here, we construct 10-day evolving undirected climate networks from 6-hourly MSLP anomalies over the Western North Pacific (WNP) based on temporal coherence and magnitude similarity of pressure series. Four network metrics characterize the spatial and temporal responses of TC-induced pressure field change, revealing track-aligned low-value bands and U-shaped lifecycle variations. A convolutional neural network (CNN) trained on these metrics distinguishes TC and non-TC windows, while saliency maps identify a Philippine Key Region (PKR) where network changes are most informative for classification. Composite analysis further shows that the PKR presents different underlying ocean-atmospheric conditions between TCs and non-TCs, including variations in the Western Pacific Subtropical High (WPSH) and convective activity. These environmental contrasts shape local pressure field organization and translate TC-induced disturbances into systematic network changes. This framework provides an interpretable network-based perspective for TCs occurrence detection.