The Ghost Rain Over Chile: Teaching AI to Catch Satellites' Invisible Errors

Somewhere off the southern coasts of Chile and Argentina, a satellite looked down and saw rain that was never there. The pixels weren't impossible — individual values fell comfortably within the range of real rainfall. But together they formed something no meteorologist would ever believe: rigid, parallel bands of precipitation marching in lockstep, a pattern that looks like a printing error, not weather. This particular glitch, photographed from the Special Sensor Microwave Imager/Sounder in September 2006, is a ghost. And ghosts like it are multiplying.

For decades, only a handful of space agencies read the sky. Now anyone can. The rise of cheap, compact "SmallSats" has flung open the doors to rainfall observation, drawing in a wave of private companies that build, launch, and sell their own precipitation data. Yet validation has not kept pace. A single undetected glitch in a brightness temperature measurement — a reading softened by a layer of ice, or blurred by a coastline's shifting emissivity — can propagate through a retrieval algorithm into rain fields that look plausible but are fiction. And in the new, decentralized satellite economy, those fictions can reach the public with no one checking.

The fix, according to researchers Andres F. Monsalve at the University of Texas at El Paso, Hernan A. Moreno (also at UTEP), and Christian D. Kummerow at Colorado State University, is a computer vision model that learns what bad data looks like — and can be taught new kinds of bad data as they appear. Their framework, tested on two legacy microwave sensors, uses a pre-trained neural network that meteorologists incrementally retrain by showing it human-labeled examples of six distinct artifact types. It doesn't just flag an orbit as suspicious; it tells you where the problem is, what kind it is, and can be retrained on the spot when a brand-new failure mode emerges (Monsalve et al., 2026).

The Science

To understand what these researchers built, you first have to understand how we measure rain from space, and why it keeps going wrong.

Passive microwave (PMW) radiometers work by sensing the faint microwave radiation naturally emitted by the Earth and its atmosphere. Their key advantage over infrared sensors is that microwaves "see through" the upper layer of clouds — they can look into a storm rather than at its lid — and they operate identically in day and night. The flagship of this endeavor is the Global Precipitation Measurement (GPM) mission, a collaboration led by NASA and the Japanese space agency JAXA, whose "core" satellite anchors a constellation of instruments from agencies including NOAA, EUMETSAT, and India's ISRO. Together, that fleet revisits any given location roughly every 90 minutes.

But the constellation is changing under the pressure of a new economics. CubeSat technology — so named for its standardized cube-shaped form factor, assembled like Legos and launched as ride-along cargo on rockets — has made it dramatically cheaper to put a radiometer in orbit. The result is a burst of private satellite-launch initiatives and surface precipitation products. This is, in principle, a wonderful democratization of Earth observation. But it comes with a mismatch: quality control that took government agencies years to develop is thin or absent on the newest spacecraft.

"While these platforms incorporate Quality Assessment procedures," the authors write, "even minor, undetected anomalies in brightness temperature measurements can propagate into significant errors in the calculated precipitation fields." The retrieval schemes that turn raw microwave readings into rain rates are Bayesian algorithms that compare measured values against a database of simulated rain profiles matched to known temperatures. These methods are "highly non-linear," the paper notes, and they demand well-calibrated, stable inputs. Feed them a glitch and they will confidently convert it into weather.

Existing quality control leans on fixed physical thresholds — temperature limits, voltage stability, orbital maneuvers — and mostly operates at the pixel level, before the rain product is finalized. A newer tool, SPEEDe (the Satellite Precipitation Estimate Error Detector), trains a convolutional autoencoder on nominal, artifact-free precipitation fields; when shown anomalous data it fails to reconstruct it, generating a score. But SPEEDe and similar unsupervised detectors have a critical blind spot: interpretability. "They provide overall scores or binary indicators," the authors note, "but do not identify the underlying characteristics of the anomaly." They can't tell you where in an image the error lives, or what it looks like.

The manual alternative is brutal. Meteorologists at NASA's Precipitation Processing System (PPS) have, by eye, hunted for and archived these artifacts across sensors including SSMIS (on satellites F16, F17, and F18), the MHS on Metop-B and Metop-C, the ATMS on NOAA-19 and NOAA-20, and the legacy SSMI on F08. It is slow and labor-intensive — which is precisely why the recorded inventory of these glitches stays small even though they are everywhere.

What They Found

The researchers worked with two instruments: the SSMI (using data from the DMSP F-8 satellite, 1987–1991) and its successor the SSMIS (from DMSP F-16, F-17, and F-18, with orbits sampled from 2006 to 2024). All data came from the GPROF V7 precipitation products, with the artifact-containing orbits curated and supplied by NASA's PPS personnel and the full labeled dataset publicly deposited on Zenodo.

From studying these orbits, the team codified six distinct classes of artifact:

• Lines — structured cross-track streaks of moderate-to-low intensity (found on both SSMI and SSMIS).
• Spots — scattered single-pixel dots, often faint (both sensors).
• Bands — parallel along-track structures alternating high and low intensity (SSMIS only).
• Landmask — rain features that bizarrely follow coastlines with varying intensity (SSMIS).
• Smoothness — rain with unnaturally smooth geometry and sharp boundaries (SSMIS).
• Mosaic — discontinuous features with the look of a broken tile floor (SSMIS).

Categorizing them matters because a one-size-fits-all detector is blunt. The framework assigns each anomalous patch a class, giving scientists a concrete starting point for diagnosis, and the paper's Table I reads like a field guide — an image dictionary of the failures a radiometer can suffer.

The core innovation is an incremental learning strategy that mirrors real operational life. A newly launched satellite's data is presumed nominal; only over time do systematic errors surface and get catalogued. The framework was trained in 15 sequential "cycles," each cycle introducing a new batch of anomalies in their real chronological order of appearance, then testing performance against a fixed hold-out validation set of six orbits per satellite. This is deliberately not a one-shot training run. It simulates deployment, discovery, and adaptation.

Three training regimes were compared. The "Baseline" model never saw retraining beyond its initial setup. The "Joint" model aggregated all data at once, the standard but operationally unrealistic approach. And the "Warm-Start" model — the heart of the proposal — began from a pre-trained backbone and was incrementally updated, ingesting each new orbit's labeled patches and continuously improving.

SSMI: Maximum artifact-detection F2-score by training approach

Label	Value
Baseline	0.62
Warm-Start	0.88

SSMIS: Maximum artifact-detection F2-score by training approach

Label	Value
Baseline	0.78
Warm-Start	0.87

For the SSMI data, the Warm-Start approach beat the Baseline on every cycle, climbing to a maximum F2-score of 0.88 while the Baseline stalled near 0.62. The F2-score is a precision-recall metric that weights recall (catching actual artifacts) more heavily than precision (avoiding false alarms) — the sensible bias for a screening tool, where missing an artifact is costlier than double-checking a borderline case. The SSMIS results told a similar story, with the Warm-Start model reaching an F2-score of 0.87 against the Baseline's 0.78: good out of the box, but decisively better once it had absorbed the sensor's idiosyncrasies.

The pipeline behind these numbers is a study in obsessive care. A 70×70-pixel window — roughly 875 by 875 kilometers on the surface — slides across each orbit with a stride of 23 pixels, creating 66% overlap between adjacent patches. This sub-windowing transforms sparse, orbit-level data into a dense set of localized training samples: one anomalous orbit yields tens of positive patches overcoming the severe class imbalance. And the labeling is meticulous: windows with unequivocal artifacts go into the positive pool; windows with only marginal traces at their borders go into a separate "IDK" category, excluded rather than risk injecting label noise that could derail training.

The augmentation is equally deliberate. Patches are reflected, rotated in rigid 90-degree increments, and translated by 15 pixels — but never sheared, scaled, or rotated by arbitrary angles. The authors insist on this not out of caution but physics. Continuous transformations require interpolation, which mathematically alters raw precipitation intensities and blurs sharp boundary gradients — the very gradients that define an artifact's signature. Rigid transformations preserve intensities exactly. "The selected transformations ensure that the augmented dataset represents spatial variations," they write, "without corrupting the unique morphological signature of the target artifacts."

Why This Changes Things

The significance here is not the novelty of a good detector — it's what the detector is designed for, and what it signals about the future of Earth observation.

Note what the framework is not. A traditional unsupervised anomaly detector treats "normal" and "anomalous" as a fixed binary. This system treats them as a moving target, and it is built to keep moving. The pre-trained backbone — MobileNetV3-Large, chosen after experiments showed smaller MobileNet was too weak serious for the task and larger ResNets overfit, ironically performing worse despite more parameters — provides a foundation of general visual understanding. Retraining ingests sensor-specific quirks. "By establishing a baseline through transfer learning and subsequently ingesting sensor-specific data," the authors explain, "the network implicitly learns the spatial biases of the target instrument."

This is crucial for adapting to the actual geometry of the new satellite era. Legacy SSMI and SSMIS instruments use conical scanning, which keeps a constant field of view. Current CubeSat radiometers mostly use cross-track scanning, which stretches the field of view at high scan angles. The two families produce structurally different data, yet the patch-based convolutional approach transfers cleanly between them. "Despite this geometric divergence," the authors note, "the proposed patch-based convolutional approach remains applicable." The framework doesn't depend on any static assumption about a swath's shape; it learns on the fly.

And importantly, explainability.

The model produces pixel-level probability maps, not just binary flags. The paper's Figure 5 shows a striking demonstration: an SSMIS orbit over southern South America carrying both Mosaic and Smooth artifacts, next to the framework's heat map of artifact probabilities across the domain, with transparent regions where confidence drops below 0.5. For a meteorologist trying to figure out what went wrong — or whether the whole orbit should be discarded — a map that says "here, and it looks like this" is transformative versus a score that just says "suspicious."

This technological shift arrives just in time. The constellation of PMW radiometers is growing faster than any validation bureaucracy can track, and the authors frame their work as a contribution to "the ever-growing availability of surface precipitation data." The value of rainfall data is not abstract. Accurate precipitation measurement underpins flood warnings, drought monitoring, agricultural forecasts, and hydrological modeling — decisions with real human cost. A decentralized fleet that distributes the sensing must also distribute the quality control. This framework is an attempt to make trust scalable.

What's Next

The researchers are candid about the limits of what they've built)Skip. "The specific causes of the artifacts presented in this research remain unidentified," they freely admit. The glitches could come from atmospheric ice layers, ocean or surface-type conditions, emissivity transitions between surface types, or purely mechanical instrument problems. Getting at root causes "would require extensive independent verification and is beyond the scope of this work." Diagnosis — the point where a detected "Mosaic" becomes a known, preventable failure mode — remains open territory.

The scarcity of anomalous data is both a feature of the design and a persistent constraint. The curators at NASA's PPS have done heroic manual work, but the total inventory of confirmed artifact orbits is small, and heavily skewed: SSMIS yielded the Bands, Landmask, Smoothness, and Mosaic classes, while the older SSMI contributed only Lines and Spots across its 1987–1991 window. When truly novel artifact types emerge on a brand-new CubeSat, the framework has a strategy — human spot-checks catch a problem, label some patches, and the system learns — but each first encounter starts from scratch.

There is also the honest challenge of class imbalance. The artifact orbits are statistically rare, and the team's solution is brute-force clever: spatial decomposition into overlapping windows, aggressive undersampling of the trivial "no-rain" patches (only 20% retained) to keep the model from learning the lazy answer "it's not raining, so it's fine," and a validation set of six orbits per satellite. It works, but it is a tightrope, and the F2-scores of 0.87–0.88, while strong, are not perfect.

What would make this a genuinely operational tool is deployment in near-real time — catching an artifact in the stream before dissemination, rather than reconstructing it years later. The framework's design, with its chronological cycles and incremental updates, is pointed exactly there abstract public. The team has open-sourced both the code and the full labeled dataset, inviting the community to stress-test it, retrain it, and extend it to new sensors.

The deeper significance is subtler. In an era where the tools to observe our planet are being democratized at a staggering pace, this paper makes a quiet but sturdy argument: that rigor, too, can be democratized. The question of who gets to say what the sky is doing is being answered by a new generation of satellites and companies. The harder question — how to make sure what they tell us is true — now has a template. The ghost rains off the coast of Chile may be growing, but so is the machine learning to hunt them.