How AI Uncovers Contrails in Thermal Satellite Images?

Contrails, the trails left by aircraft, affect climate by reflecting heat back to space. Detecting them in thermal satellite images is tough: they are thin, have blurry edges, and blend with cirrus clouds. To tackle this, researchers built MFcontrail, a deep‑learning model that mixes multi‑axis attention with frequency‑domain enhancement. Its core is a MaxViT encoder that captures long‑range spatial relationships, while the FreqFusion decoder preserves high‑frequency details, i.e., sharp edges. An “edge‑aware” loss further pushes the network to refine boundaries. Tests on two datasets – OpenContrails and Landsat‑8 – show MFcontrail beating DeepLabV3+: +5.03 % F1‑score and +5.91 % IoU on OpenContrails, and +3.43 % F1‑score and +4.07 % IoU on Landsat‑8. Ablation studies reveal that almost 70 % of the IoU gain comes from frequency enhancement. This high‑precision tool helps researchers better assess aviation’s climate impact, proving frequency‑domain strategies are valuable for satellite cloud analysis.

MFcontrail leverages a hybrid architecture that merges the strengths of Vision Transformers (ViT) with multi‑axis attention mechanisms. The MaxViT encoder splits the image into patches and applies a local attention window followed by global fusion, enabling capture of patterns at both fine (contrail edges) and coarse (cloud structures) scales. The FreqFusion decoder operates in the frequency domain via 2‑D Fourier transforms, separating low‑frequency components (gradual temperature variations) from high‑frequency ones (sharp contours). High frequencies are reintegrated during up‑sampling, sharpening edge definition. The edge‑aware loss combines standard cross‑entropy with a Sobel‑gradient term, penalizing boundary errors. In tests, IoU improved by 5.91 % on OpenContrails due to better separation of contrails from cirrus clouds, which share similar frequency spectra. Ablation studies revealed that removing the frequency module loses nearly 70 % of the IoU gain, confirming the critical role of high‑frequency preservation. Moreover, MFcontrail uses fewer parameters than DeepLabV3+, lowering memory usage and speeding up inference on onboard hardware. This approach paves the way for more responsive climate monitoring systems capable of processing large volumes of thermal data in real time, providing essential inputs for aviation radiative forcing models.