Facilitating Climate-Friendly Aviation: Spatial-Frequency Synergy for Contrail Detection in Remote Sensing Imagery

Aircraft condensation trails (contrails), formed in ice supersaturation regions, are human-induced clouds with significant climate warming effects. Mitigating their climate impact through feasible flight rerouting requires precise contrail detection, tracking and prediction from satellite imagery. Given the potential divergent feature distributions in the frequency domain of comparable contrail observations, this study proposes to integrate frequency-oriented learning and regularisation into encoder-decoder-based semantic segmentation models. It includes a lightweight separable frequency-aware (LSFA) block to capture complementary patterns in feature extraction and an auxiliary regularisation for frequency-domain discrepancy of active samples (FDA) in the loss function. Evaluated on two humanlabelled contrail imagery datasets (GOES-16 and Landsat-8), our approach achieves consistent improvements over the state-of-theart methods across a suite of metrics with a minimal decrease in inference speed. Ablation studies confirm the effectiveness of incorporating frequency-domain knowledge into contrail detection. This study also evaluates contrail detection performance across contrail coverage variations, significantly enhancing the detection of short and fragmented contrail segments. The generalisation of our method across two datasets with varying resolutions and label granularity suggests its potential for real-world deployment in aviation-induced climate impact mitigation. Furthermore, the benchmarks established in this study provide a foundation for future research in the field.