The team, led by Martin Bonte with Lesley De Cruz, Fabian Debal and Stéphane Vannitsem, evaluated the two models over Belgium using the RADCLIM radar product, a quantitative precipitation estimate that merges radar measurements with rain gauge observations through Kriging with External Drift. Crucially, neither model was fine-tuned for the Belgian domain. LDCast was used with its original pre-trained weights, and STEPS was run through the open-source pysteps library exactly as a national weather service would deploy it. This choice matters, because in practice most forecast offices will not have the time, data or expertise to retrain a state-of-the-art AI model before using it. The evaluation therefore reflects a realistic operational scenario rather than a best-case laboratory demonstration.

The two models represent fundamentally different philosophies of nowcasting. STEPS, short for Short-Term Ensemble Prediction System, is built on Lagrangian persistence: it advects the observed rainfall field along a motion field estimated by optical flow, decomposes the field into a cascade of spatial scales, and lets each scale evolve through an auto-regressive process. Randomness enters through stochastic perturbations of both the motion field and the intensities. LDCast, by contrast, is a latent diffusion model. A variational autoencoder compresses sequences of rainfall fields into a compact latent space, a forecaster network predicts the latent representation of future fields, and a denoiser stack then generates different ensemble members conditioned on that output. The generative approach naturally produces diverse, nondeterministic forecasts, which makes ensemble construction almost effortless.

To compare the two fairly, the researchers selected ten convective and ten stratiform rainfall events, classified using ERA5 reanalysis data on convective and large-scale precipitation rates. For each event they generated 50-member ensembles and examined the spread-error relationship scale by scale, using Fourier power spectra. In a perfectly calibrated ensemble, the spread among members should match the error of the ensemble mean against observations. The spectral analysis showed that both models come remarkably close to this ideal for scales larger than about five kilometres. LDCast was slightly underdispersive overall, though somewhat overdispersive at large scales during convective events, while STEPS was underdispersive at short lead times for stratiform rain and became well calibrated as lead time grew.

One subtlety the authors had to untangle involved the smallest scales, below roughly five kilometres, where the power spectrum of radar observations flattens out. This flattening is not a physical feature of rainfall but an artefact of white noise contamination in the radar data and of the smoothing introduced by merging with rain gauges. The apparent underdispersion of both models at these scales therefore cannot be blamed on the ensembles themselves. Interestingly, the error at these smallest scales saturates almost immediately, within the first five minutes of the forecast, at scales between five and ten kilometres depending on the model and event type, meaning the issue is largely irrelevant for practical forecasting.

Beyond the raw size of uncertainty, the team probed the internal geometry of the ensembles using the covariance matrix of the members and its eigenvectors, essentially a principal component analysis of the forecast uncertainty. The results revealed a striking architectural difference. STEPS ensembles, especially in convective cases, are dominated by a handful of directions in phase space: convective ensembles concentrate their variability in roughly ten dominant modes, and stratiform ensembles in about five. The residual vectors of individual members align along these few leading eigenvectors, meaning the members are quite similar to one another. LDCast ensembles, in contrast, display much more homogeneous eigenvalue spectra, spreading their variability more evenly across many modes. Both models, however, adapt their ensembles to the meteorological situation, developing larger-scale perturbations for stratiform events and smaller-scale ones for convection, a sign that the models are responding to genuine dynamical differences between the two regimes.

The eigenvalue growth itself followed power laws in lead time, a behaviour consistent with the theoretical predictability limits of atmospheric flows whose energy spectra follow the classical minus five-thirds power law at nowcasting scales. Errors seeded at small scales cascade toward larger ones, and the perturbation modes gradually develop power at progressively larger scales as the forecast advances. This hierarchy, in which the largest-amplitude perturbations affect the largest scales, mirrors the way predictability is lost in the real atmosphere, and it is encouraging that both a physics-inspired statistical scheme and a purely data-driven generative model reproduce it.

The critical test came when the researchers asked whether the ensembles could localise the error in space, not just quantify it. They used two complementary metrics: the cosine of the angle between the actual forecast error and its projection onto the ensemble perturbations, a geometric measure related to the PECA score, and the probabilistic Fraction Skill Score, which assesses spatial accuracy while tolerating displacement errors. The clever twist was the construction of surrogate ensembles. MAAFT surrogates, inspired by the Iterated Amplitude-Adjusted Fourier Transform, preserve the statistical fingerprints of the original ensembles, namely the distribution of pixel values and the power spectra of the perturbations, but scramble the complex Fourier phases that encode spatial structure. If the real ensembles contain genuine spatial information, they should outscore these statistical ghosts.

They did not. The scores of the MAAFT surrogates were essentially indistinguishable from those of the actual STEPS and LDCast ensembles for both metrics, at all lead times and for both event types. In Fourier language, the complex phases of the ensemble perturbations, which carry all the spatial localisation information, are effectively random. The ensembles know how big the error should be at each scale, and they even reproduce the right distribution of rainfall intensities, but they have no dynamical insight into where the error will actually materialise. The skill they display is statistical, not dynamical. A similar picture emerged from the Fraction Skill Score analysis, where both models’ scores saturated with lead time toward values achievable by purely random predictions constrained to the same value distributions.

The authors are careful about what this means for operations. The finding does not invalidate generative nowcasting; LDCast’s spread still provides a useful, nearly unbiased estimate of error magnitude across most scales, even without regional retraining, and both models demonstrably adapt their uncertainty to the weather regime. But forecasters hoping that an ensemble of AI rainfields will point toward the specific locations where the forecast is going wrong will be disappointed. The study also notes a practical asymmetry: when STEPS runs out of radar data at the edge of its domain, it produces explicit missing values that honestly flag where information is lacking, whereas a generative model like LDCast simply invents plausible-looking values with no such transparency. Future work, the authors suggest, should test a version of LDCast fine-tuned for Belgium, evaluate other generative architectures such as DGMR, which produces less diverse members, and explore blending nowcasting models with numerical weather prediction, an approach already available in pysteps. For now, the message is clear: AI ensembles have learned the statistics of rainfall uncertainty, but the dynamics of where that uncertainty strikes remain, for the moment, beyond their grasp.

Subject of Research: Evaluation of ensemble spread-error relationships and spatial error representation in precipitation nowcasting using STEPS and the generative AI model LDCast over Belgium

Article Title: Spread/error relationship and spatial error representation in precipitation nowcasting: comparison of STEPS and generative AI

Article References: Bonte, M., De Cruz, L., Debal, F., & Vannitsem, S. (2026). Spread/error relationship and spatial error representation in precipitation nowcasting: comparison of STEPS and generative AI. Weather and Climate Dynamics, 7(3), 1837-1852. https://doi.org/10.5194/wcd-7-1837-2026

Image Credits: AI Generated

DOI: 10.5194/wcd-7-1837-2026

Keywords: precipitation nowcasting, LDCast, STEPS, generative AI, ensemble forecasting, spread-error relationship, latent diffusion model, radar, uncertainty quantification, Fraction Skill Score, convective precipitation, Weather and Climate Dynamics

The team, comprising Ajay Waghumbare and Upasna Singh from the Department of Computer Science and Engineering and A. Arockia Bazil Raj from the Department of Electronics Engineering, calls its approach TriPath-WGAN-GP. The name encodes the three structural innovations that distinguish it from a standard generative adversarial network. At its core sits the Wasserstein Generative Adversarial Network with Gradient Penalty, or WGAN-GP, a variant of the GAN family introduced by Arjovsky and colleagues and refined by Gulrajani and colleagues, which replaces the notoriously unstable original GAN objective with a Wasserstein distance measure that provides smoother, more informative gradients during training. That mathematical change matters enormously when the thing being generated is not a natural photograph but a micro-Doppler spectrogram, a specialized two-dimensional representation of radar returns in which the horizontal axis is time, the vertical axis is Doppler frequency, and pixel intensity encodes the energy scattered by moving body parts.

Micro-Doppler signatures are the reason radar can tell a walking person from a crawling one. When radar waves reflect off a human body, the bulk motion of the torso produces a strong, slowly varying Doppler shift, while the swinging of arms and legs, the rotation of limbs, and even the micro-vibrations of the chest impose fine, rapidly oscillating modulations on top of that shift. First systematically modeled by Chen and colleagues in 2006, these modulations form distinctive time-frequency patterns that act almost like kinematic fingerprints. A classifier that reads them well can distinguish walking, running, sitting, falling, and suspicious movements, which is precisely what makes radar attractive for surveillance, fall detection in elderly care, inattentive driving monitoring, and noninvasive person authentication. But those fingerprints are subtle, and a neural network trained on a few hundred examples tends to memorize them rather than learn the underlying physics, collapsing when it meets people or conditions it has never seen.

The first of the three paths in TriPath-WGAN-GP addresses the generator, the network responsible for synthesizing fake spectrograms from random noise. Rather than a plain feed-forward stack, the researchers built progressive residual refinement into the generator, an architecture in which residual connections allow information about fine spectrogram structure to survive the journey through many layers. Residual learning, popularized in image recognition, lets each layer learn only the correction it needs to apply to its input, which stabilizes deep networks and preserves high-frequency detail. In the radar context, that high-frequency detail corresponds exactly to the limb-induced micro-Doppler modulations that a classifier must see to do its job. A generator that blurs them away produces spectrograms that look plausible at a glance but are useless for training discriminative models.

The second path concerns the critic, the adversarial counterpart of the generator whose job is to tell real spectrograms from synthetic ones. TriPath-WGAN-GP equips the critic with multi-scale discriminative learning, meaning it evaluates generated samples at several spatial scales simultaneously. This is a deliberate response to the nature of micro-Doppler data: the coarse envelope of the spectrogram captures gross body motion, while the fine texture captures limb dynamics, and a critic operating at a single scale can be fooled by samples that match one level of structure while violating the other. By judging both, the critic applies pressure that forces the generator to reproduce the full hierarchy of physical detail, from torso Doppler down to the fine striations of a swinging arm.

The third path is stabilization. Training adversarial networks is famously delicate, and the researchers layered two forms of regularization on top of the standard gradient penalty: spectral regularization alongside the gradient constraint. Spectral normalization, analyzed by Lin and colleagues, limits the Lipschitz constant of the critic’s layers by constraining the singular values of its weight matrices, preventing any single layer from amplifying signals catastrophically. Combined with the gradient penalty that keeps the critic’s gradients well behaved along the path between real and generated distributions, this dual regularization produces the stable optimization that the framework’s results depend on. The authors also evaluated their outputs with both pixel-level metrics and distribution-aware measures, an important distinction because two spectrograms can be pixel-similar while differing in the statistical structure that classifiers actually exploit.

The quantitative results are the study’s headline. The synthetic dataset produced by the framework, named DIAT-μRadHAR-α, achieved a Fréchet Inception Distance of 140.78 and an Inception Score of 1.96, outperforming conventional GAN-based augmentation methods on the same task. The FID, introduced by Heusel and colleagues, measures how similar the distribution of generated images is to the distribution of real ones by comparing activations of a pretrained Inception network; lower is better, and the score indicates that the synthetic spectrograms occupy a statistically recognizable neighborhood of the real data manifold. The Inception Score, from Salimans and colleagues, rewards generated samples that are individually confident in their class membership yet collectively diverse, capturing the balance between fidelity and variety that any augmentation scheme must strike.

More consequential than the generative metrics are the downstream classification experiments, because the ultimate test of synthetic training data is whether it actually teaches networks to recognize real activities. The researchers trained three deep convolutional architectures from scratch on the generated data and compared them against pretrained counterparts. The from-scratch models won consistently and decisively: MobileNetV2 reached 98.15 percent accuracy, EfficientNetB0 achieved 99.15 percent, and InceptionV3 attained 96.57 percent. That last comparison carries a lesson that ripples well beyond radar. EfficientNetB0, designed by Tan and Le, and MobileNetV2, from Sandler and colleagues, are compact architectures whose inductive biases suit spectrogram-like images, whereas InceptionV3, a heavyweight of natural-image classification, underperformed despite its pedigree. Generic pretraining on photographs, the results suggest, imports priors that actively conflict with the statistics of radar data.

This finding challenges a widespread habit in the field. Transfer learning from ImageNet-pretrained networks has become the default recipe for small-data problems, and several prior studies, including work on transfer learning for micro-Doppler classification by Seyfioglu, Erol, and Gurbuz, built on exactly that assumption. The DIAT results indicate that when the domain gap is wide enough, the features transferred from natural images are not merely unhelpful but harmful relative to what a network can learn from a sufficiently rich synthetic corpus. In other words, a well-designed generative augmentation pipeline can substitute for pretraining, supplying domain-specific statistical structure that generic weights cannot. The study builds on a lineage of GAN-based radar augmentation, from the DCGAN spectrogram schemes of Mi and colleagues in 2018 through the kinematically sifted ACGAN signatures of Erol, Gurbuz, and Amin, and the WRGAN-GP approach of Qu and colleagues, each iteration refining how faithfully the physics of human motion is reproduced in synthetic form.

The practical implications extend to AI-driven surveillance and real-time sensing systems, where the authors see strong deployment potential. Because the dataset analyzed in the study is publicly available and the implementation code is available from the authors on reasonable request, the framework is positioned to be adopted and stress-tested by other groups working on radar perception, from smart-home fall detection to through-wall sensing. The broader significance is methodological: it demonstrates that the data scarcity problem in specialized sensing domains can be attacked not by collecting more data but by generating it, provided the generator is architecturally constrained to respect the physics that downstream classifiers depend on. For a field where every labeled radar recording represents a person performing a scripted activity in a lab, that is a meaningful shift. If synthetic micro-Doppler spectrograms can push recognition accuracy above 99 percent, the argument for camera-free, privacy-preserving human activity recognition becomes considerably harder to ignore.

Subject of Research: Generative data augmentation with WGAN-GP for radar-based human activity recognition using micro-Doppler spectrograms

Article Title: TriPath-WGAN-GP for radar-based human activity recognition via micro-Doppler spectrogram augmentation

Article References: Waghumbare, A., Singh, U., & Bazil Raj, A. A. (2026). TriPath-WGAN-GP for radar-based human activity recognition via micro-Doppler spectrogram augmentation. Multimedia Tools and Applications, 85(10), Article 777. https://doi.org/10.1007/s11042-026-21934-y

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21934-y

Keywords: radar, human activity recognition, micro-Doppler spectrogram, generative adversarial networks, WGAN-GP, data augmentation, deep convolutional neural networks, EfficientNetB0, MobileNetV2, surveillance, signal processing, synthetic data

To appreciate why this matters, it helps to understand what an optoelectronic oscillator actually does. A continuous-wave laser feeds an intensity modulator, the modulated light travels through a long fiber, a photodetector converts it back into an electrical signal, and that signal is amplified and fed back to drive the modulator. If the loop gain exceeds unity at the right frequency, the system self-oscillates, producing a microwave tone whose stability is set by the optical delay in the loop. Because optical fibers store light with extraordinarily low loss, the delay can be made enormous—far larger than any practical microwave cavity could provide—which is why OEOs routinely beat electronic oscillators in spectral purity.

The catch is the mode structure. The oscillation frequencies of a loop are spaced by the inverse of the loop delay, so a kilometer-scale fiber produces modes separated by only a few hundred kilohertz. A physical filter in the loop must select exactly one of these closely spaced modes while rejecting all the others, and the narrower the filter, the harder it becomes to build and to keep stable against temperature drift and vibration. Worse, in a conventional single long loop, the mode-selection filter must simultaneously deliver very high rejection of neighboring modes and very low insertion loss, and these demands pull against each other. The result is the classic phase noise–mode purity trade-off: shorten the loop and the spectrum cleans up but close-in noise worsens; lengthen the loop and the noise floor drops while spurious modes multiply.

The new work attacks this dilemma with an architecture built from cascaded zero-dispersion recirculating loops. Instead of relying solely on an electronic or optical bandpass filter to pick a single mode, the researchers exploit the physics of chromatic dispersion inside the fiber itself. When a recirculating loop operates near its zero-dispersion wavelength, the phase of light circulating in the loop becomes exquisitely sensitive to wavelength, and the interference between successive circulations reshapes the effective transmission spectrum of the loop. By carefully biasing the loop to the zero-dispersion point, the team created a comb-like spectral response in which only modes satisfying a strict phase condition can build up, while others are suppressed by destructive interference accumulated over many round trips.

Cascading is the second key ingredient. A single zero-dispersion loop, however elegant, still leaves a periodic transmission function with a free spectral range tied to its own length. By connecting two or more such loops of deliberately different lengths in series, the researchers multiplied their spectral responses together. Where one loop might transmit several candidate modes, the second loop transmits only a subset, and the third a subset again. The product of these comb functions is a spectrum with a single dominant transmission peak at the desired oscillation frequency and deep, wide rejection windows everywhere else. In effect, the mode-selection burden is distributed across the cascade, so no individual element needs to perform near-impossible filtering on its own.

The payoff is a microwave signal that combines the close-in phase noise of a very long loop with the spurious-free spectrum of a much shorter one. In measurements reported by the team, the oscillator achieved phase noise levels at low frequency offsets that would normally require kilometer-scale fiber, while suppressing neighboring modes by margins that conventional long-loop designs cannot reach without elaborate external filtering. The spurious modes that plague standard long-loop OEOs—tones sitting tens of kilohertz or a few hundred kilohertz from the carrier—are pushed down below the noise floor, leaving a signal clean enough for the most demanding radar, navigation, and metrology applications.

The technical details reveal how much engineering sits beneath the concept. The zero-dispersion operating point must be found and held: fiber dispersion shifts with temperature, so the team implemented bias control that keeps each loop anchored near its zero-dispersion wavelength despite environmental drift. The gain and loss budget across the cascade had to be balanced so that the desired mode sees net gain while every competing mode sees net loss over a complete round trip through all loops. The photodetector, amplifier, and modulator in the feedback path were characterized to ensure that their own noise contributions—thermal noise, shot noise, and amplified spontaneous emission from any optical amplification—do not undermine the phase-noise advantage earned by the long effective delay. Each of these elements is individually familiar to photonicists; the achievement lies in integrating them into a self-stabilizing oscillator that behaves as a single, coherent system.

Why does zero dispersion help rather than hurt? In a dispersive loop, different wavelength components of the modulated optical carrier accumulate different phase shifts per round trip, which smears the interference that defines the loop’s transmission peaks and can destabilize mode selection. At the zero-dispersion wavelength, the leading-order phase distortion vanishes, so the loop behaves as if all spectral components travel together, while the residual higher-order dispersion still provides the wavelength-dependent phase structure that shapes the comb response. The researchers show that operating precisely at this point maximizes the contrast of the interference-based mode discrimination: the wanted mode accumulates constructive phase across circulations while unwanted modes fall onto transmission minima that deepen with every additional pass. The recirculating geometry thus turns a potential liability—long propagation—into the very mechanism that cleans the spectrum.

The implications reach well beyond the laboratory bench. Microwave oscillators with sub-femtosecond timing jitter are the hidden backbone of modern technology: they clock coherent radar arrays that resolve faint targets, they steer phased-antenna beams in 5G and future 6G networks, they provide the local oscillators for atomic clocks and very-long-baseline interferometry, and they define the repetition-rate stability of frequency combs used to count optical cycles. Any architecture that breaks the long-loop trade-off without resorting to bulky, temperature-controlled external cavities is a candidate to move out of the metrology lab and into deployed systems. Because the cascaded-loop scheme is built from standard telecom fiber, modulators, and detectors, it is inherently compatible with photonic integration and chip-scale packaging strategies that the field is actively pursuing.

There remain, of course, questions that further work must answer. Long-term frequency drift, vibration sensitivity, and the scaling of the cascade to even longer effective delays will need systematic study, as will the power consumption of the bias-stabilization electronics in field conditions. But the conceptual advance is clear and likely to be influential: the phase noise–mode purity trade-off, long treated as a fundamental constraint of loop-based oscillators, is not a law of nature but a consequence of design choices. By letting dispersion physics do the mode selection, the researchers have shown that the two most coveted properties of a microwave source can finally be purchased with the same currency. For engineers who spend their lives chasing quieter, cleaner signals, that is a trade-off worth breaking.

Subject of Research: Cascaded zero-dispersion recirculating loops for high-purity, low-phase-noise optoelectronic oscillators

Article Title: Breaking the phase noise–mode purity trade-off in long-loop OEOs via cascaded zero-dispersion recirculating loops

Article References: Wang, Z., Bernal, S., Plant, D. V., & Chen, L. R. (2026). Breaking the phase noise–mode purity trade-off in long-loop OEOs via cascaded zero-dispersion recirculating loops. Light: Science & Applications, 15(1), Article 379. https://doi.org/10.1038/s41377-026-02413-3

Image Credits: AI Generated

DOI: 10.1038/s41377-026-02413-3

Keywords: optoelectronic oscillator, phase noise, mode purity, microwave photonics, zero-dispersion loop, recirculating fiber loop, mode selection, spectral purity, fiber optics, radar, frequency combs, photonic integration