Technical Summary: ContrailNet for Global Automated Contrail Monitoring

Problem Statement

Aviation-induced contrails contribute significantly to global climate warming, with their radiative forcing estimated to be three times that of aviation CO₂. Despite their impact, global, long-term, automated monitoring of contrails remains challenging due to the scarcity of high-quality, manually annotated datasets. While Global Climate Models (GCMs) can simulate contrails, they suffer from significant uncertainties related to microphysical parameterizations and a dependence on high-resolution emission inventories. Traditional satellite-based detection methods using polar-orbiting sensors (e.g., MODIS) often rely on fixed thresholds or geometric features, making them vulnerable to noise from complex backgrounds, terrain, and tropical cirrus clouds. Although deep learning has shown promise, previous attempts often failed to achieve satisfactory results without massive annotated datasets. Furthermore, existing high-performance deep learning models, such as those developed by Google for the GOES-16 geostationary satellite, are optimized for North America and cannot be directly applied to global polar-orbiting data due to sensor discrepancies and the lack of global training labels.

Methodology

The study introduces ContrailNet, a UNet-based computer vision model designed for the detection of slender, linear contrail features. The core innovation is a three-stage transfer learning strategy that migrates expertise from a large, expert-annotated GOES-16 dataset (North America) to the global MODIS dataset (Terra and Aqua satellites).

1. Model Architecture

ContrailNet is built upon the U-Net architecture but integrates specific modules to address the unique challenges of contrail detection:

• Atrous Spatial Pyramid Pooling (ASPP): Embedded to expand the receptive field, allowing the model to capture long-distance linear geometric features that standard U-Net architectures often fragment.
• Squeeze-and-Excitation (SE) Attention: Integrated to automatically recalibrate channel weights, enhancing the representation of contrail-sensitive spectral bands while suppressing background noise.
• Focal Loss: Replaces traditional cross-entropy loss to address the extreme class imbalance where contrail pixels constitute less than 1% of the image, forcing the model to focus on hard-to-classify edge pixels.

2. Three-Stage Training Pipeline

• Stage 1 (Pre-training): The model is pre-trained on the Google OpenContrail dataset (GOES-16), which contains over 20,000 expert-annotated samples. This stage allows the model to learn fundamental linear morphological and spectral characteristics of contrails.
• Stage 2 (Cross-Sensor Transfer): The pre-trained weights are fine-tuned using a small subset of hand-marked MODIS images (~100 samples). This stage adapts the model to the spectral response differences between the GOES-16 and MODIS sensors.
• Stage 3 (Noise-Adaptive Enhancement): To improve robustness against complex global backgrounds (terrain, cirrus, sensor noise), the model undergoes further fine-tuning. This stage utilizes a "Copy-Paste" data augmentation strategy, where pure contrail bodies extracted from oceanic backgrounds are synthetically combined with complex interference scenes (e.g., rivers, ice sheets, broken clouds) to create a diverse training set without requiring extensive new manual labeling.

3. Data Strategy

The study constructs a lightweight, manually labeled MODIS dataset comprising 179 samples (133 for fine-tuning, 18 for validation, 28 for testing). The input data utilizes specific infrared bands (MODIS Bands 29, 31, 32) to construct RGB false-color images that highlight the distinct thermal and spectral properties of ice crystals in contrails, mirroring the approach used for GOES-16.

Key Results

Performance and Efficiency

• Accuracy: ContrailNet achieves an F1-score of approximately 0.63 on the global MODIS dataset (Stage 3), which is comparable to the benchmark DeepLab model's performance on GOES-16 images. At a recall of 0.7, the precision reaches 0.68.
• Efficiency: ContrailNet converges significantly faster than the baseline DeepLab model. It achieves peak performance in 71 iterations compared to 2000 for the benchmark, consuming only 1/30 of the computational resources (TFLOPs) while maintaining comparable accuracy.
• Sample Efficiency: The study demonstrates that performance gains plateau after approximately 100 annotated MODIS samples, proving that massive global datasets are not required for effective cross-sensor transfer.

Generalization and Robustness

• Cross-Regional Validation: The model was validated across four major flight corridors: North Atlantic (NAFC), Asia-North America (ANAFC), Australia-Asia (AAFC), and North American Regional (NARC). Performance remained stable (precision 0.50–0.65 at 0.7 recall) across these diverse regions.
• Noise Suppression: The Stage 3 noise-adaptive training effectively suppressed false positives caused by sensor stripes, river contours, ice sheet edges, and tropical cirrus clouds, which previously caused significant errors in Stage 1 models.
• Environmental Sensitivity: The model showed slightly lower performance over land (F1=0.58) and polar regions (F1=0.53) compared to oceans (F1=0.71) due to complex emissivity and high-albedo interference, but maintained robust detection capabilities globally.

Global Monitoring Application

Using ContrailNet, the authors reconstructed global annual mean contrail coverage (AMCC) for 2013, 2019, 2020, and 2022.

• Trend Analysis: Global contrail coverage increased by 18.7% from 2013 to 2019, aligning with aviation industry growth.
• Pandemic Impact: In 2020, global contrail occurrence dropped by 37% due to COVID-19 lockdowns.
• Radiative Forcing Impact: This reduction in contrails corresponds to a decrease in warming radiative forcing of approximately 21 mW m⁻² (equivalent to a reduction of ~0.24 Gt CO₂-eq yr⁻¹).
• Recovery: By 2022, contrail occurrence recovered by 27%, though it remained below pre-pandemic levels.

Significance and Claims

The paper claims that ContrailNet provides a cost-efficient, real-time applicable tool for the automated monitoring of aviation non-CO₂ climate effects. Its primary significance lies in:

1. Regulatory Compliance: It offers a technical solution for the European Union's Emissions Trading System (EU ETS) under Directive (EU) 2023/958, which mandates the monitoring of non-CO₂ climate impacts, including contrails.
2. Scalability: By demonstrating that a model trained on a geostationary satellite (GOES-16) can be effectively transferred to polar-orbiting satellites (MODIS) with minimal manual annotation (~100 samples), the study enables the construction of global, long-term monitoring systems without the prohibitive cost of manual labeling.
3. Scientific Validation: The model successfully captured the "trough" in contrail coverage during the 2020 pandemic, validating its ability to process complex global heterogeneous scenes and providing reliable evidence for assessing the climatic impact of aviation activities.
4. Future Adaptability: The transfer learning framework establishes a basis for future adaptation to next-generation polar-orbiting sensors and supports the development of near-real-time global contrail monitoring systems.