Boundary-Aware Tropical Cyclone Detection from Geostationary Satellites and Climatic Archive Data using Advanced Neural Networks

Abstract:

Tropical Cyclones (TCs) are among the most severe and catastrophic weather phenomena causing human life loss and significant large-scale infrastructure damage. Atmospheric researchers continue to work towards better understanding and improving the prediction of TCs and their impacts. Boundary-aware TC detections provide valuable information regarding the location and region of interest (ROI) of the TC circulation including rainbands. In this study, we first introduce a gridded TC ROI dataset which fuses International Best Track Archive for Climate Stewardship (IBTrACS) and InfraRed (IR) longwave channel of geostationary satellites. The gridded TC ROI product has high spatial (0.04º×0.04º) and temporal (30-minutes) resolutions with global coverage spanning the period from 1983 to the present. The dataset is compared to the IBTrACS-derived maps to assess the data fusion quality. Next, we developed an advanced neural network structure, termed as U-NET, to learn TC patterns and used them to delineate TC in fine details from the fused spatially distributed TC ROIs. The U-NET structure prevents the “loss-of-resolution” problem in the segmentation task and preserves the spatial information while using the pooling operation. The introduced model uses multiple information such as GOES IR images, latitude and longitude information, and land/ocean mask, as inputs. The model is trained in two main steps: (1) positive calibration in which the model is trained on samples with TCs, and (2) negative calibration in which the model is trained on samples with no TCs in them. The proposed model outputs are evaluated and compared to the fused dataset and IBTrACS in terms of Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI) and accuracy. The model results demonstrate high-quality performance in fine spatial features detection compared to the above-mentioned datasets. The results are in good agreement with the fused dataset when assessed with qualitative visual comparison.