Weather forecasting helps the drone industry take flight

Drones operate nearly everywhere, including regions where there are significant weather data reporting gaps, especially below 1,000 ft (often referred to as the planetary boundary layer (PBL)). These gaps are often in rural areas and have terrain that impacts wind speed and direction, as well as turbulence, and draining batteries and shortening travel distance. Accurate, low altitude weather forecasts are not widely available today. A new company, Climavision, is setting out to change that by bring a unique new approach to the weather services industry.

Numerical Weather Prediction (NWP) has become the most important tool for the more accurate weather forecasting the drone industry needs. NWP focuses on taking current observations of the atmosphere and processing these data with computer models to forecast the future state of the atmosphere. Yet, the company believes that knowing the current state of the weather is just as important (if not more so) than the numerical computer models processing the data and producing the forecast.

Current weather observations serve as input to the numerical computer models through a process known as data assimilation. The importance of high density and high-quality weather observations in NWP cannot be overstated.

Accurate atmospheric measurements of the PBL are crucial for improving the ability to capture and forecast the micro-climates that directly impact drone operations. To complement the existing global weather observing network and fill in the spatial and temporal gaps in PBL coverage, Climavision is deploying a high-density X-band radar network across the US. The X-band radar data will not only provide unparalleled real-time weather surveillance capabilities, but it will also fuel highly accurate short-term weather forecasts of those regions of the airspace most challenging to drone operators.

In addition to the gap filling radar network, the company is partnering with commercial satellite observation provider, GeoOptics, to fill in existing weather observation data voids over large areas of the Atlantic and Pacific Oceans. GeoOptics captures and produces high density GNSS-RO (Global Navigation Satellite System-Radio Occultation) observations that serve as unbiased anchor weather measurements in their partner’s global data assimilation engine.

Combining the proprietary radar and GNSS-RO data with the existing surface and upper-air network, their high-performance data scientists are building a unique weather observation data tank that will serve as the foundation for all their NWP initiatives.

Equally important to the capture and collection of existing and novel observing platforms, the company is working with several research groups towards next-generation observational quality control systems.  These quality control systems filter out poor or suspect quality observations that can have a negative impact on the analysis and subsequent low-altitude forecast.

Care must be taken, however, to ensure that these systems are not filtering out good data as well. It is well understood within the atmospheric science community that the current generation of quality control systems are overly aggressive at times in this data screening process.

Climavision scientists are developing state-of-the-art machine learning-based techniques to better screen and filter the global observing datasets described above. This greatly improves their ability to better leverage not only the existing global weather observing networks, but also the growing availability of new, proprietary commercial observing networks.

After the global weather observations are screened and quality controlled, they are then suitable to be used within their data assimilation system. Data assimilation algorithms are being developed to better exploit these newly available proprietary weather observations. The data assimilation framework includes existing operational data assimilation algorithms and facilitates exploration of new data assimilation science across various domains and commercial applications.

This is a rather different approach than what traditional commercial weather companies have taken with their focus on the weather forecast model and visualisation, and much less on the observations and data assimilation systems that are the main drivers of numerical forecasting model accuracy. This is analogous to focusing all new vehicle efforts on the frame– and not focusing on the engine, drivetrain and transmission that ultimately propel the vehicle forward.  

The result of combining high quality weather observations with advanced quality control and data assimilation processes are to produce an atmospheric analysis. This analysis serves as the starting point for their numerical forecasting models…the process that directly produces the proprietary low-altitude weather forecasts. Unlike traditional commercial weather providers, Climavision incorporates machine learning algorithms into all components of the NWP process.

Incorporating machine learning into the dynamic integration of numerical weather prediction models is a new and emerging field – one that shows much promise. Numerical prediction models are being cycled at higher and higher horizontal and vertical resolutions. This is an important development as they work to better capture and forecast the micro weather events that threaten safe and productive drone operations.

However, this increasing resolution creates challenges in both the computational expense (how many computers are required to produce a forecast that is timely enough for forecasters to integrate into the operational forecasting process) and in the complexity of the physics parameterisations required in the forecasting models themselves.

The staff is continuously developing and deploying numerical weather prediction models at increasingly high resolutions as they work to advance the science of numerical weather prediction.  Increasing the model resolution, however, is only one of the challenges. The resolution increase expands the computational requirements needed to produce these sophisticated forecasts within the time constraints required by the company’s clients and decision makers.

While high performance computing technology continues to advance, the computer processors themselves are not increasing in performance as quickly. To account for this, they have begun to integrate machine learning techniques into what is the most computationally expensive component of the numerical model itself, the physical parameterisations.

Given its central importance at Climavision, it is worth briefly exploring its High-Performance Computing (HPC) approach. The supercomputers on which data assimilation and numerical forecasting systems are run are becoming increasingly complex, incorporating always more and heterogeneous processing elements. Using these efficiently is a growing challenge, with much effort exploring the ways to improve the scalability of their forecasting models and data assimilation systems.

The Climavision approach works on the assumption of scalability to facilitate adaptation to new and varied HPC architectures. This allows them to quickly adapt and improve model efficiencies while lowering costs as new technologies emerge. This approach allows the company to take advantage of the latest advancements in HPC technology – as well as improving the operational efficiency and skill of the numerical forecast models themselves.

The ability to have very granular understanding of micro-climates in real-time will enhance safe and profitable drone operations around the world. Climavision’s combination of proprietary radar observations, GNSS-RO satellite observations, flexible and forward-looking numerical weather prediction technology, and advancements in cloud computing are aligning to provide a roadmap for the company to become the industry leader in low-altitude weather prediction solutions and technology.