CARM: Conformal Adaptive Risk Masking for Reliable Scientific Super-Resolution

Guangtao Zhang, Haonan Wu, Junxuan Gao, Zhengyang Liu, Zhaoxi Zeng, Zhihao Li, Jianguo Huang, Zhiqiang LiuarXiv 2026 · arXiv

TL;DR

CARM (Conformal Adaptive Risk Masking) is a new statistical framework that generates reliable "trust masks" for scientific super-resolution models, even when data distributions shift over time due to seasonality or extreme events. By dynamically aggregating threshold proposals from multiple adaptive experts, CARM provides provable guarantees that reconstruction errors will stay within user-defined limits without requiring the data to be independent or identically distributed.

What Problem Does This Paper Address?

Deep learning models are excellent at "super-resolution" (SR)—reconstructing high-resolution Earth observation data from coarse satellite imagery. However, for scientific applications like monitoring ocean health or wind patterns, reconstruction accuracy isn't enough. Researchers need to know exactly which parts of a generated image are reliable.

Existing methods like Conformal Mask Calibration (CMC) create "trust masks" by setting a threshold on uncertainty scores. If the data distribution is stable (exchangeable), these masks offer rigorous guarantees on the maximum expected error. But Earth is not stable. Seasonal cycles, mesoscale eddies, and marine heatwaves cause "distribution shifts," where the relationship between a model's uncertainty and its actual error changes. A threshold calibrated on last year's data might be dangerously overconfident or uselessly conservative today.

Methods

CARM is a model-agnostic, post-hoc framework. It doesn't require retraining the underlying SR model; instead, it sits on top of any model that provides a spatial uncertainty score.

Key components include:

  • Multi-rate Experts: Rather than committing to a single adaptation speed, CARM maintains a pool of "experts," each updating its threshold proposal at a different rate ($\eta$). Some experts react quickly to sudden changes, while others provide stability during calm periods.
  • Online Aggregation: The framework aggregates these proposals into a single deployed threshold ($\tau_t$) by upweighting experts that have recently kept their realized risk closest to the target level $\alpha$.
  • Asymmetric Surrogate Loss: To update weights, CARM uses a loss function that penalizes risk violations (being too risky) more heavily than unnecessary conservatism (being too safe).
  • Causal Delayed Feedback: In real-world Earth observation, ground-truth high-resolution data often arrives with a delay. CARM is designed to work only with "matured" feedback available after a fixed delay ($d$ days).
Fig. 2. Overview of CARM. Experts with different step sizes update their threshold proposals. The proposals are aggregated into the deployed threshold τt, which is applied to Ut to generate the trust mask.

Datasets

The authors validated CARM across three distinct geophysical variables and seven different super-resolution backbones (including FNO, SwinIR, and ResShift):

  1. South China Sea Chlorophyll-a (Chl-a): A daily stream of ocean color data.
  2. ERA5 Wind Fields: Global atmospheric reanalysis data.
  3. OSTIA Sea-Surface Temperature (SST): High-resolution satellite-derived temperature fields.

Key Insights and Ideas

  • No Single Rate Fits All: The authors demonstrate that the "optimal" adaptation rate for a threshold changes repeatedly over time (see Figure 1). By using an ensemble of experts, CARM avoids the need for manual hyperparameter tuning for different datasets.
  • Theoretical Robustness: The paper provides a formal proof (Theorem 1) that the time-averaged risk converges to the target level $\alpha$ at an $O(1/T)$ rate. Crucially, this guarantee does not require the data to be i.i.d. or exchangeable, making it highly suitable for non-stationary time series.
  • Reliability-Coverage Trade-off: Reliability isn't free. When the environment becomes more volatile, CARM maintains safety by shrinking the "trusted" region of the image. This ensures that the pixels that are marked as trusted actually meet the fidelity requirements.
Comparison of adaptation rates over time, showing that no single fixed rate is uniformly best.

Key Results

CARM showed consistent, significant improvements over static calibration (CMC) across all tested scenarios:

  • Violation Reduction: On the South China Sea Chl-a stream at a target risk $\alpha=0.10$, CARM reduced the empirical violation rate from 43.49% to 9.38%.
  • Sustained Reliability: The moving-average exceedance frequency (MAER), which measures how often the risk stays above the target for a week, dropped from 47.29% to just 1.22%.
  • Visual Precision: Qualitative analysis shows that the trust masks correctly identify and reject high-error regions, such as complex coastal structures and boundary layers, while retaining stable interior-ocean areas.
Fig. 4. Qualitative examples of deployed trust masks on the South China Sea test stream at α = 0.10. Each row corresponds to one representative test sample. From left to right: HR reference, ensemble SR prediction, absolute reconstruction error, and the deployed trust mask. Red regions are rejected as untrusted.

Limitations and Future Work

  • Delayed Feedback Limits: Because CARM relies on matured feedback, there is an inherent lag in its response to extremely abrupt, unprecedented distribution shifts.
  • Uncertainty Quality: The effectiveness of the mask depends on the underlying uncertainty scores being at least somewhat informative. If the uncertainty map is pure noise, the mask cannot effectively rank spatial errors.
  • Future Directions: The authors suggest developing faster adaptation mechanisms for sudden shifts and incorporating explicit "risk-coverage" objectives to maximize the size of the trusted region while maintaining safety.

Summary updated

Abstract

Deep learning-based super-resolution (SR) can reconstruct fine-scale Earth-observation fields from coarse observations, but reconstruction accuracy alone does not provide a calibrated rule for identifying reliable spatial regions. Conformal Mask Calibration (CMC), built on conformal risk control, converts a spatial uncertainty score into a trust mask and controls the expected worst-case local distortion under exchangeability. However, its fixed threshold can become miscalibrated in Earth-observation streams affected by seasonality and episodic events. We propose $\textit{CARM}$ (Conformal Adaptive Risk Masking), a statistical framework for adaptive spatial risk control under temporal distribution shift. CARM maintains experts with heterogeneous update rates and aggregates their threshold proposals online using an asymmetric surrogate loss, balancing rapid adaptation to distribution changes against stability during relatively stationary periods. All updates are causal and use only mature feedback available after a fixed delay, allowing CARM to operate with any SR model that provides a spatial uncertainty score. We prove that the time-averaged mask-conditioned risk converges to at most $α$ at an $O(1/T)$ rate, without distributional assumptions and for any choice of aggregation weights across experts. On a daily South China Sea chlorophyll-$\textit{a}$ SR stream, CARM reduces the empirical violation rate from $43.49\%$ to $9.38\%$ and the moving-average exceedance frequency from $47.29\%$ to $1.22\%$ at $α=0.10$, while retaining a non-trivial trusted region. Experiments on ERA5 wind and OSTIA sea-surface temperature further show consistent relative improvements across multiple SR backbones.
BibTeX
@misc{zhang2026carm,
  author = {Guangtao Zhang and Haonan Wu and Junxuan Gao and Zhengyang Liu and Zhaoxi Zeng and Zhihao Li and Jianguo Huang and Zhiqiang Liu},
  title = {{CARM: Conformal Adaptive Risk Masking for Reliable Scientific Super-Resolution}},
  year = {2026},
  eprint = {2611.06151},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.06151},
}
Figure