Ground-based gamma-ray telescopes use the atmosphere itself as their detector. A gamma ray or a cosmic-ray proton hits the upper atmosphere and produces a cascade of particles, called a shower. The shower emits a faint flash of Cherenkov light, a blue glow given off by particles moving faster than light travels in air. Cameras on the ground photograph that flash, and a classifier decides from the image whether it came from a gamma ray or from the proton background, which is $10^3$ to $10^4$ times more common. These classifiers are trained on simulations that assume one fixed, clean atmosphere. Meanwhile, the same observatories run Raman lidars, laser rangefinders that measure how much light each altitude slice of the air absorbs. The lidar readings are used only afterwards, to correct the energy scale. This proposal asks a narrow question: does giving the classifier the measured height-by-height extinction profile improve its separation of gamma rays from protons, and in which atmospheric conditions does that matter?
Our central bet is that the answer depends on height. A dust or haze layer below the altitudes where the shower emits its light just dims the whole image evenly. That dimming can be corrected with a single number, and nothing more elaborate is needed. A layer inside the emission altitudes is different. It dims one part of the shower more than another and changes the image's shape, which is the very thing the classifier relies on. Gamma and proton showers reach their brightest point at different and differently scattered heights, so such a layer may distort the two classes unequally. If this is right, the vertical structure of the profile carries information that no total-transmission correction can recover. If it is wrong, the cheap correction is enough, and that is worth knowing too.
Core Claim
We propose a gamma/hadron classifier conditioned on the measured extinction profile. "Conditioned" means the profile enters the network as an extra input that changes how the image features are processed. We test it against three cheaper ways of handling the atmosphere. The claim is specific: when an aerosol layer sits within the shower's emission heights, conditioning on the full height-resolved profile lowers the proton false positive rate at fixed gamma efficiency. It also reduces the shift in gamma efficiency that occurs when cuts tuned on the standard atmosphere are applied to the real one. Both benefits should go beyond what training on randomized atmospheres achieves, and beyond rescaling the image by the transmission at the reconstructed emission height. Below the emission region, we predict no meaningful gain over the rescaling baseline.
The work connects the two research directions in a lopsided way. Discrimination is the task. Multi-level vertical profiling supplies the conditioning signal and shapes the evaluation, because the main experimental variable is the height of the layer.
What Prior Work Does and Leaves Open
Two papers motivate the gap. Calibration of the Cherenkov Telescope Array describes the atmosphere as an active calorimeter. It specifies Raman lidars, radiosondes and cloud monitors, but uses them to correct energy scale and flux, not to classify individual events. Raman LIDARs and atmospheric calibration for the Cherenkov Telescope Array designs the lidars themselves: Raman channels reach 18 km with under 5% systematic uncertainty on the aerosol and molecular components. It never connects these profiles to shower data. We take their stated precision as the noise level that our conditioning inputs must tolerate.
Three papers supply the classifiers and show how they are limited. Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC replaced hand-tuned cuts on image-shape parameters with a random forest that outputs a continuous "hadronness" score. Its authors name the static atmosphere in their simulations as the main source of mismatch with real data. Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes combines per-camera convolutional networks with a graph network linking the telescopes; graph networks pass messages between connected nodes, here the telescopes. Its "Strong" variant adds the reconstructed shower-core position to the graph. That paper trains only at a 20° zenith angle under one uniform atmosphere, and its architecture has no input for environmental data. We reuse that architecture as our backbone and add the missing input.
$S_b$ for gamma/hadron separation in SWGO is a ground-particle array analogue that also assumes a static atmosphere. We note it as a later extension, not a target here.
Graph Neural Networks for IceCube Signal Classification gives a cautionary result. Its network quietly learned to pick out events that start inside the detector volume, rather than the physics the authors intended. That warns us that a conditioned network may exploit artifacts of how the simulation is set up, so our evaluation holds out entire families of atmospheres.
Research Program
We would build a library of simulated air showers matching the H.E.S.S.-style setup of the CNN-GNN study: diffuse protons and gamma rays at 20° zenith. Each shower is rendered under many atmospheres. The key efficiency is that aerosol extinction acts on Cherenkov photons on their way to the ground. The same simulated cascade can therefore be re-imaged under different extinction profiles by applying altitude-dependent transmission to its photons, without re-simulating the particle cascade. This yields paired data: the same shower seen under a clean sky and under a layered one. Pairing sharpens every comparison, because shower-to-shower fluctuations cancel. Changes to the molecular density profile do require re-simulation, so we treat molecular profiles more coarsely.
The atmosphere library covers several cases:
- a clean reference atmosphere;
- low boundary-layer haze of varying optical depth;
- elevated aerosol layers placed at a sweep of heights from just above the telescope site to beyond the typical emission altitude;
- thin high cloud.
We take the low-altitude case from the profiling literature. Inversion-capped haze of the kind resolved by a 325 m tower in Rapid formation of intense haze episodes via aerosol–boundary layer feedback in Beijing is the physical template for that bottom layer.
Each profile is given to the network as extinction coefficients on a fixed altitude grid. Before use, it is corrupted with noise and bias at the under-5% systematic level quoted for the CTA lidars. The conditioning works by having a small network map the profile to a per-channel scale and shift, which are applied to the image features and the graph features. We also test a physically structured variant. There, the profile is first re-expressed relative to each event's reconstructed height of brightest emission, so the network sees "how much extinction lies above versus below this shower's light-emitting region." Comparing the two variants tells us whether the network needs that physics handed to it.
The analysis has two parts. The first is a curve of benefit against layer height: the gap in false positive rate between conditioned and unconditioned models, plotted against where the layer sits relative to each class's distribution of emission heights. The second is a check of whether any gain comes from proton and gamma images being distorted differently, which the paired data lets us measure directly.
Baselines
The naive reference is the Strong CNN-GNN trained only on the clean atmosphere, with its decision threshold tuned on clean simulations, then applied unchanged to layered test atmospheres. This is the current practice inherited from Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes. It tells us how much the atmosphere actually hurts. The only number that paper reports for its strongest model is relative: an "additional 9% reduction in the False Positive Rate" for Strong over Fast CNN-GNN at a threshold of 0.5, with higher ROC AUC than a Hillas-parameter GNN. No absolute AUC appears in the file, and we found no reported result on Lacuna for gamma/hadron performance under non-standard or layered atmospheres. We will therefore first reproduce the Strong-over-Fast ordering on our clean-atmosphere data as a sanity check. We also run the random-forest hadronness pipeline of Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC, trained on Hillas parameters (the classic ellipse-fit width, length and orientation of the image), as the legacy reference.
The strongest competitor is the same Strong CNN-GNN trained on the full randomized atmosphere library but given no profile input. If the atmosphere's effect can be read off the image itself, this model absorbs it, and conditioning adds nothing.
The second strongest competitor is the physics-based correction. Pixel intensities are divided by the transmission from the reconstructed emission height to the ground, computed from the lidar profile, and the result is fed to a randomized-trained Strong CNN-GNN. This uses the lidar, but only through a single number per event.
Two ablations isolate the proposal's key ingredient. The first conditions on a single scalar, the total aerosol optical depth, instead of the height-resolved profile; this tests whether the vertical structure matters at all. The second is an oracle trained only on the exact test atmosphere, which bounds what any conditioning could achieve.
All learned models share:
- the same backbone and the same number of training showers;
- the same atmosphere-library draws for training;
- an equal hyperparameter-search budget, with the same number of trials on a validation set built from atmospheres disjoint from the test set;
- five random seeds.
The proposal counts as beating a baseline only if three conditions hold on the held-out elevated-layer atmospheres. Its proton false positive rate at 50% and 80% gamma efficiency, measured per reconstructed-energy bin, must be lower than the randomized and rescaling baselines by more than the seed-to-seed spread. Its quality factor, gamma efficiency divided by the square root of background efficiency, must improve correspondingly. And its gamma-efficiency bias at fixed cut must be smaller.
Why not just…?
Why not just train on many atmospheres and let the network figure it out? This is the most serious threat. It is exactly what the randomized baseline tests. A single image cannot always reveal its atmosphere. A dim image from a hazy night looks much like a lower-energy shower on a clear night. We expect randomized training to blur its decision boundary across atmospheres rather than adapt to each one. If the randomized model matches the conditioned one across all layer heights, the conclusion is that the lidar is unnecessary for classification, and we would report that plainly.
Why not just rescale image brightness by the measured transmission? For layers below the emission region, we predict this works as well as full conditioning. The experiment is designed so that this outcome is visible, not hidden. The open question is layers inside the emission heights, where a single per-event scale factor cannot represent the uneven dimming along the shower. The benefit-against-height curve settles this. If rescaling closes the gap at every height, the practical recommendation becomes "rescale, don't condition," which is still a result that CTA-style pipelines can adopt immediately.
Why not just retrain on each night's simulated atmosphere, or on real background data? Per-night simulation and retraining is the oracle in our design. It is the target, but it is costly to keep up for every run. The question is how close a single conditioned model comes to it. Training the background class on real off-source data, which the MAGIC random forest tolerates with up to 1% signal contamination, captures the real atmosphere for protons only. The gamma class still comes from simulation under an assumed sky, so the mismatch moves rather than disappears.
Why not just throw away bad-weather data? That is current practice. The CTA calibration paper notes that 20–30% of observation time is historically lost to non-ideal weather. Recovering part of that time with controlled background leakage is the practical payoff. Our gamma-efficiency-bias metric measures whether recovered nights can be trusted.
Proposed Experiments
The first decisive check is the benefit-against-height curve on held-out layer heights. It is computed on paired showers, comparing the conditioned model with the randomized and rescaling baselines. Our hypothesis predicts a gap near zero for low layers that rises when the layer overlaps the emission altitudes. A flat curve near zero would refute the claim. Gains spread evenly across all heights would suggest the network is exploiting a shortcut rather than physics; we would then check by shuffling profiles between events, which should erase any real benefit.
The second check is robustness. We degrade the profile input from noiseless to the under-5% lidar systematic level, then to twice that level, and report where the advantage disappears. This tells us whether existing lidar precision is good enough.
The third check is the scalar-versus-profile ablation at matched training budget. It directly tests whether multi-level profiling, as opposed to a single optical depth, carries information about discrimination.
As a final, opportunistic check, if real paired lidar and telescope data from a steady source become available through a collaboration, we would measure the night-to-night stability of the excess rate as a function of lidar-measured layer height. We would report how much the conditioned model narrows that spread relative to the clean-atmosphere model. The file supplies no such public dataset, so the core results stand on simulation alone.
References
- Calibration of the Cherenkov Telescope Array
- Raman LIDARs and atmospheric calibration for the Cherenkov Telescope Array
- Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC
- Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes
- $S_b$ for gamma/hadron separation in SWGO
- Graph Neural Networks for IceCube Signal Classification
- Rapid formation of intense haze episodes via aerosol–boundary layer feedback in Beijing