پیوندهای مفید
اخبار و اعلانات
آمار
| تعداد نشریات | 127 |
| تعداد شمارهها | 7,237 |
| تعداد مقالات | 77,625 |
| تعداد مشاهده مقاله | 161,094,069 |
| تعداد دریافت فایل اصل مقاله | 120,790,286 |
Dual Stream Attention U-Net for High Precision Methane Plume Segmentation from Sentinel 2 Imagery | ||
| Earth Observation and Geomatics Engineering | ||
| دوره 10، شماره 1، مهر 2026، صفحه 41-50 اصل مقاله (756.46 K) | ||
| نوع مقاله: Original Article | ||
| شناسه دیجیتال (DOI): 10.22059/eoge.2026.410055.1213 | ||
| نویسندگان | ||
| Arshia Jafari؛ Reza Shah-Hosseini* | ||
| School Surveying and Geospatial Engineering, College of Engineering, University of Tehran | ||
| چکیده | ||
| Methane (CH₄), a potent greenhouse gas with a global warming potential far exceeding that of CO₂, is a critical driver of near-term climate change. Accurate identification and monitoring of high-emission point sources, also known as super-emitters, are essential for effective mitigation. Although multispectral satellite systems, such as Sentinel-2, provide the spatial resolution and coverage required for plume detection, conventional spectral ratio and enhancement techniques suffer from low sensitivity and high false favorable rates, often requiring manual verification. To address these limitations, we present a novel Dual-Stream Attention U-Net designed for automated methane plume segmentation from Sentinel 2 imagery. Our framework integrates two complementary input streams: (1) atmospherically corrected 12-band Sentinel 2 reflectance data, and (2) pre-computed Multi-Band Multi-Pass (MBMP) enhancement maps. The dual-encoder architecture learns modality-specific features that are fused at the bottleneck, while strategically placed attention gates, applied only to skip connections from the Sentinel-2 stream, dynamically suppress noise and emphasize plume-relevant regions. The model is trained and evaluated on the public CH4Net dataset using a composite Dice+BCE loss, and its robustness is enhanced by test-time augmentation and post-processing. Experimental results show substantial gains over the CH4Net baseline, achieving an IoU of 74.09%, Balanced Accuracy of 96.56%, and markedly lower false favorable rates. Qualitative analyses confirm the model’s ability to detect faint and irregular plumes in complex backgrounds where previous methods fail. These findings demonstrate that the proposed approach delivers scalable, high-precision, fully automated methane plume monitoring, suitable for operational deployment, and advancing global efforts in greenhouse gas mitigation. | ||
| کلیدواژهها | ||
| Methane Plume Segmentation؛ Sentinel-2 Imagery؛ Deep Learning؛ Attention U-Net؛ Dual-Stream Architecture | ||
|
آمار تعداد مشاهده مقاله: 11 تعداد دریافت فایل اصل مقاله: 5 |
||
سامانه مدیریت نشریات علمی. قدرت گرفته از سیناوب