What’s brown and sticky and can be fired out of a (GPU enabled) Cannon? ^[1]

Many of the technologies we use in modern astronomy have their origins in other fields of human endeavour. In an effort to ensure our instrumentation program at ANU is sustainable into the future, we are of course always looking for ways to ensure this technology transfer flows both ways. As part of the wider ranging iLAuNCH Trailblazer program (a major Australian Government sponsored industry engagement initiative managed by the University of Southern Queensland and coordinated through InSpace here at ANU^[2]), our team here at Mount Stromlo has partnered with home-grown and international Industry groups to develop detector control technologies for Earth observation remote sensing.  We call the detector control system Rosella. Funding for the project ($3.8M over three years) officially began in July this year.

The Rosella program will deliver a quasi-commercial detector control solution configurable for use with most of the detectors used in space-based astronomy and remote sensing detectors. The target market is small satellite systems (or better still constellations) that have demanding performance specifications, but need the small form-factor and mission specific operational demands.  Our own OzFuel mission (in collaboration with Professor Marta Yebra’s Bushfire Research Centre of Excellence in the Fenner School of Environment and Society on campus^[3]), is a great example of an early customer. 

Figure 1.

Image

However, one of the problems for Earth observation applications is the required frame rate and the associated data deluge that this generates. This data volume can’t readily be transmitted to the ground for detailed analysis in real time. In fact, it typically can’t even be stored short-term for more than an orbit or two before it must be overwritten with fresh observations.  Some sort of value-added data compression process is needed. One that can turn millions of independent pixels from hundreds of thousands of independent spectra, into a small subset of population defining parameters that represent the “science” traced by the raw data. This sounds very familiar…

As astronomical surveys moved to generating increasingly larger volumes of raw and processed datasets, stellar spectroscopists have had to develop data-driven approaches to efficiently extract fundamental properties of objects from this morass of data. Realising that spectra are in fact “just spectra”, regardless of where they come from, and inspired by Marta's recent colloquium at Mount Stromlo, RSAA postdoctoral researcher Dr Sven Buder suggested a simple exchange of “stellar” with “leaf” spectroscopy. He sought to enhance is spectroscopy class by feeding a one of the astronomical classification pipelines with the spectra Marta and Dr Nicolás Younes (Fenner School of the Environment and Society) have been collecting in the field as part of the OzFuel bushfire spectral library. This pipeline - The Cannon, developed by Dr Melissa Ness (RSAA) et al. (2015)^[4] for use with the Apogee and later applied to the GALAH program by Sven (Buder et al. 2018^[5]) – has turned up some interesting results.

After a crash-course in the physical and chemical properties of leaf spectra with Nicolás, he and Sven have been able to show The Cannon appears to identify discrete spectral regions that flag specific information about the vegetation that makes up the critical bushfire fuel load which the OzFuel team seek to measure on a national scale to predict and inform bushfire hazard management strategies.

The software model for The Cannon is not overly complex or computationally onerous in the context of astronomical survey data (that typically generate 100s - 1000s of spectra an hour). However, in the context of remote sensing (which is looking to generate 100s - 1000s of spectra a second) this becomes a pixel throughput problem. And how do we solve a pixel throughput rate problems in astronomy? We give it to a bunch of PhD students and promise them a conference trip if they sort it out, of course! However, this violates WHS policies on the safe work environments when in low Earth orbit. The slightly more expensive solution is, according to ChatPGPT "GPU Acceleration: Utilise GPU computing (e.g., CUDA, OpenCL) for data-intensive operations, especially in image processing tasks.” This is extraordinarily convenient, as GALAH’s latest data release (Buder et al., 2024^[6]) has implemented GPU-accelerated neural networks for their spectrum analysis; and, one of our iLAuNCH partners, Spiral Blue (a Sydney-based SME) just happens to make a space-qualified, GPU-enabled, “Space Edge Computer” specifically designed for AI data analysis at high data throughput.  You’d almost think we knew what we were doing!

Figure 2.

Image

It’s still early days, undertaking the cross-analysis to check that we properly understand what The Cannon is telling us about bushfire fuel loads.  The code was initially designed for stellar surface gravities, temperatures and metallicities, etc., rather than cellulous and lignin levels and eucalyptus oil load. Age estimates at the ±1 Myr level for example are not quite what we’re looking for. Interestingly it turns out dust content is still an issue, as is sand, gravel, and soil content (and if we go out to 5 microns it will all turn into PAH soot again, I’m sure). However, the team is excited to be exploring technical and scientific crossovers between very different types of remote sensing, and is a step closer to making Australia a little safer in bushfire season. 

Rob Sharp, Sven Buder, Marc White, Joice Mathew (The Rosella technical team RSAA/AITC),  Marta Yebra (Fenner School of the Environment and Society, School of Engineering) and Nicolás Younes (Fenner School of the Environment and Society)

==========================================================================================