Deep learning‑based super‑resolution (SR) can reconstruct fine‑scale Earth‑observation fields from coarse data, yet reconstruction error alone does not yield a calibrated rule for trustworthy spatial regions. Conformal Mask Calibration (CMC), built on conformal risk control, converts a spatial uncertainty score into a trust mask and controls the worst‑case local distortion under exchangeability. Its fixed threshold, however, becomes miscalibrated when observation streams are affected by seasonality or episodic events.\
We introduce CARM (Conformal Adaptive Risk Masking), a statistical framework for adaptive spatial risk control under temporal distribution shift. CARM maintains a set of experts with heterogeneous update rates; each proposes a threshold. An asymmetric surrogate loss aggregates these proposals online, balancing rapid adaptation during distribution changes with stability during relatively stationary periods. All updates are causal and rely only on mature feedback available after a fixed delay, allowing CARM to work with any SR model that outputs a spatial uncertainty score.\
We prove that, without distributional assumptions and for any aggregation weights, the time‑averaged mask‑conditioned risk converges to at most $\alpha$ at an $O(1/T)$ rate. On a daily South China Sea chlorophyll‑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 $\alpha=0.10$, while preserving a non‑trivial trusted region. Experiments on ERA5 wind and OSTIA sea‑surface temperature further demonstrate consistent relative improvements across multiple SR backbones.\
Review: By leveraging adaptive threshold aggregation and causal delayed feedback, CARM addresses the calibration drift of conformal risk masks in temporally varying environments, offering both theoretical guarantees and substantial empirical gains. It constitutes a robust tool for enhancing reliability in scientific super‑resolution applications.