Abstract: Passive thermal infrared imaging provides a non-contact sensing approach for robotic detection of shallowly buried targets. However, background objects such as rocks and vegetation can produce target-like thermal anomalies, resulting in increased false alarms across different terrains. In this study, thermal infrared mine detection is formulated as a conservative confirmation task following high-recall candidate generation, and a method termed Mine-RCCNet is proposed. The method preserves the central region of each candidate while replacing the surrounding background to construct intervention views. It then combines stability aggregation, risk-head regularization, and training-domain quantile-based threshold calibration to determine whether the candidate response consistently relies on thermal evidence from the central region. When trained on the grassland and sandpit domains and evaluated on the held-out bareground domain, MatchedBI with K = 2 achieved the lowest false positive rate (FPR) of 3.65% and the highest F1 score of 58.72%, while exhibiting smaller recall fluctuations than temporal baselines. These results demonstrate that the proposed mechanism can reduce candidate-level false alarms in unseen terrains and improve confirmation stability. Nevertheless, cross-terrain threshold calibration remains an important direction for future research.