1. The "Regional Nowcasts on GIS Web Portal" provides the following map layer products:
(a) Deep learning nowcasts of satellite-derived reflectivity based on (i) ResConvLSTM-GAN (Residual Convolutional Long-Short Term Memory with Generative Adversarial Network) and (ii) Earthformer for the next 4 hours and 8 hours respectively.
(b) AI-STORMVIS (AI-driven Satellite-based Tropical cyclone Object Recognition, Motion Visualization and Intensity estimation System) on automatic identification of centre position and intensity estimation of TCs.
(c) Aviation nowcast products on polygons enclosing areas of potential ice crystal icing and significant convection for the next 6 hours.
2. Deep learning nowcast of satellite-derived reflectivity
(a) The satellite-based reflectivity image is derived based on various imagery channels of JMA Himawari-8/9 satellite using artificial neural network. The satellite-derived reflectivity images are then input to two deep learning nowcast models, namely, ResConvLSTM-GAN and Earthformer to predict reflectivity for the next 4 hours (240 minutes) and 8 hours (480 minutes) respectively. More background and references on introducing deep learning techniques in HKO nowcasting system can refer to this web page.
(b) User can use the time slider near the top of the page to choose different forecast time levels. When user clicks on the map layer, a pop-up dialog will be displayed to show the predicted reflectivity at that location from the two deep learning nowcast models (figure below).
(c) Brief specifications of deep learning nowcast products are provided in the following table:
(d) Users should note that the reflectivity is derived from satellite observations of cloud top in different channels that the location and intensity can differ from the reflectivity detected by ground-based weather radars, although the radar reflectivity has been employed to train the algorithm. In this connection, the movement and spatial coverage of satellite-derived reflectivity that forecast by the AI/ML models may not be identical to the actual precipitation systems – especially those generated by low-level convection or underneath of deep thunderstorms with extensive cloud tops.
3. AI-STORMVIS
(a) AI-STORMVIS (AI-driven Satellite-based Tropical cyclone Object Recognition, Motion Visualization and Intensity estimation System) provides automatic identification of tropical cyclone (TC) on its centre location and estimation of TC intensity. More information on background and verification of AI-STORMVIS can refer to this link.
(b) The map layer of AI-STORMVIS has coverage of satellite imagery over 15°S – 45°N ; 85°E – 160°E and updated every 10 minutes. The available time of the latest AI-STORMVIS layer including the detected TC information will normally be around 35 minutes after the satellite observation time.
(c) A brief illustration on using AI-STORMVIS map layer is given in the following figure:
(d) Users should note that the identification of TC system and location, as well as intensity estimation are automatically based on the AI/ML algorithms in AI-STORMVIS. For official assessments, and taking precautionary actions against TC and associated inclement weather, users are reminded to refer to the latest analysis, forecast and warning information from the Hong Kong Observatory, NMHSs and RSMCs.
4. Aviation Nowcast Products
(a) Map layers on polygons enclosing areas of (i) ice crystal icing and (ii) significant convection and their 6-hour nowcasts (in step of 60 minutes) are provided. More details and their plain image / animation versions are accessible from this page.
(b) The time of availability of the products will be around 20 minutes (30 minutes) for significant convection (ice crystal icing) after the satellite observation time.
