Wednesday, 15 January 2025: 2:00 PM
354 (New Orleans Ernest N. Morial Convention Center)
ABSTRACT: Small Unmanned Aerial System (sUAS) operations are quickly becoming a significant fraction of air traffic in the United States. Due to their small size, flight operations and planning are particularly sensitive to weather. Precipitation poses a unique forecasting challenge, as any precipitation significantly hinders flight. A free, cloud-based forecasting application was developed to provide short-term weather guidance to sUAS pilots (www.uasforecast.com). Built upon the High-Resolution Rapid Refresh (HRRR) model, this application provides 0-18-hour precipitation, winds, ceilings, and visibility forecasts. It also provides a real-time display of Multi-Radar/Multi-Sensor (MRMS) precipitation data, watches and warnings, and MPING reports. Weather risk and tolerances differ greatly between sUAS operations and the general public, and this application provides a streamlined way for drone operators to manage risk without having to sift through meteorological data on their own, often without a meteorology background.
This project aims to improve precipitation forecasts in the application by blending MRMS and HRRR guidance to develop probabilistic precipitation forecasts. The project is broken into three phases: nowcast neighborhoods, radar projection analysis, and model analysis. The project's domain is the Contiguous United States, broken into nine climate regions. The analysis for each climate region was done for each meteorological season from spring 2023 to winter 2024 (March 2023 - February 2024). The thirty cities in each climate region were chosen subjectively to balance both population (where we expect many user queries) and a wide distribution across each region. The analysis is performed in intervals of six hours, where the prior six hours of data are used to forecast for the next six hours.
Nowcasts were performed for thirty cities in nine climate regions in the Contiguous United States. The data used is the precipitation flag product, which separates precipitation into convective and stratiform types. The nowcast aims to determine an optimal radius for detecting convection in the radar projection and model analysis phase. The radius that is tested for each city ranges from 2.5 km to 250 km. If convection is detected in the neighborhood, the analysis checks if precipitation occurs in the city pixel. The Heidke Skill Score (HSS) is used to quantify the neighborhood's skill as it is a better indicator of accuracy where events are uncommon. Over the entire climate region for each season, the HSS is averaged. The HSS increases with smaller radii, with the optimal radii being where the HSS begins to level off asymptotically to one. The result is an optimal convection detection radius for each climate region in each season. General precipitation detection is also done as a control to compare to the separated convection detection.
Next, MRMS data is projected out in six-hour intervals using the Lucas-Kanade optical flow algorithm. The data used is the MRMS quality-controlled composite reflectivity. A vector field is created using the prior few hours as training data on which the live radar data will be advected. Future radar data is generated for each hour to a six-hour lead time. A watershed algorithm is used to classify the precipitation data and identify convective cores. 35dbz and above returns surrounding the convective cores are classified as convective. Bayes theorem is then used to return a precipitation probability dependent on convection detection, season, and region. The farther we project the radar data, the worse the projection skill becomes.
Once the HRRR becomes more skilled, the HSS determines the appropriate cut-off time for radar projection. A similar process to determine precipitation probabilities is then applied to the HRRR.
Preliminary results show that separating precipitation events by convection detection, region, and season improves the skill of the forecasts. In particular, separating convective events can increase skill significantly in some cases compared to using general precipitation. Early results also show that simple projection is accurate in the short term, as the model spin-up limits early lead time accuracy. Once the model spins up and the projections' skill becomes poor, the results are more skillful with the model.
This project aims to improve precipitation forecasts in the application by blending MRMS and HRRR guidance to develop probabilistic precipitation forecasts. The project is broken into three phases: nowcast neighborhoods, radar projection analysis, and model analysis. The project's domain is the Contiguous United States, broken into nine climate regions. The analysis for each climate region was done for each meteorological season from spring 2023 to winter 2024 (March 2023 - February 2024). The thirty cities in each climate region were chosen subjectively to balance both population (where we expect many user queries) and a wide distribution across each region. The analysis is performed in intervals of six hours, where the prior six hours of data are used to forecast for the next six hours.
Nowcasts were performed for thirty cities in nine climate regions in the Contiguous United States. The data used is the precipitation flag product, which separates precipitation into convective and stratiform types. The nowcast aims to determine an optimal radius for detecting convection in the radar projection and model analysis phase. The radius that is tested for each city ranges from 2.5 km to 250 km. If convection is detected in the neighborhood, the analysis checks if precipitation occurs in the city pixel. The Heidke Skill Score (HSS) is used to quantify the neighborhood's skill as it is a better indicator of accuracy where events are uncommon. Over the entire climate region for each season, the HSS is averaged. The HSS increases with smaller radii, with the optimal radii being where the HSS begins to level off asymptotically to one. The result is an optimal convection detection radius for each climate region in each season. General precipitation detection is also done as a control to compare to the separated convection detection.
Next, MRMS data is projected out in six-hour intervals using the Lucas-Kanade optical flow algorithm. The data used is the MRMS quality-controlled composite reflectivity. A vector field is created using the prior few hours as training data on which the live radar data will be advected. Future radar data is generated for each hour to a six-hour lead time. A watershed algorithm is used to classify the precipitation data and identify convective cores. 35dbz and above returns surrounding the convective cores are classified as convective. Bayes theorem is then used to return a precipitation probability dependent on convection detection, season, and region. The farther we project the radar data, the worse the projection skill becomes.
Once the HRRR becomes more skilled, the HSS determines the appropriate cut-off time for radar projection. A similar process to determine precipitation probabilities is then applied to the HRRR.
Preliminary results show that separating precipitation events by convection detection, region, and season improves the skill of the forecasts. In particular, separating convective events can increase skill significantly in some cases compared to using general precipitation. Early results also show that simple projection is accurate in the short term, as the model spin-up limits early lead time accuracy. Once the model spins up and the projections' skill becomes poor, the results are more skillful with the model.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner