Adaptive uncertainty quantification for maritime classification under cloud cover in satellite imagery

Classifying vessels in satellite imagery is important for naval intelligence, search-and-rescue, and the monitoring of illegal activities. Performance, however, degrades under adverse meteorological conditions (e.g., clouds), which obscure discriminative features and increase ambiguity of class membership. Conventional deep learning methodologies fail to directly address this ambiguity, sometimes generating overly confident predictions for a specific erroneous class. To address this gap, we present WAVES (Weather-Aware Visual Estimation with Sets), a conformal prediction framework that makes uncertainty quality-aware. WAVES learns a temperature map that adjusts classifier logits as a function of predicted cloud coverage. A single global conformal threshold is then computed using the standard order statistic, preserving split-conformal marginal coverage while producing prediction sets that adapt across quality conditions. We evaluate WAVES using a publicly accessible dataset comprising various ship classes, which we augment with realistic synthetic clouds. Across three relevant classification models from the literature and compared against standard split–conformal prediction, WAVES typically maintains or improves coverage while reducing prediction–set size; across all three models, it reduces average set size by 6–7% on average (up to 15% in the best setting) while keeping coverage within ±0.1 percentage points of the global split–conformal prediction baseline. These results indicate that quality-aware calibration combined with a global conformal threshold provides reliable, compact prediction sets for maritime surveillance under variable cloud cover, especially in operational scenarios where classification outcomes directly influence critical human or autonomous decisions.