S2MetNet: A novel dataset and deep learning benchmark for methane point source quantification using Sentinel-2 satellite imagery

Methane, as a crucial greenhouse gas, plays a significant role in global warming, contributing to approximately one-quarter of the observed climate change since pre-industrial times. Consequently, the detection and quantification of major methane emitters are vital in addressing this issue effectively. Satellite sensors with shortwave infrared (SWIR) spectral bands provide valuable information for monitoring methane emissions. For example, Sentinel-2 multispectral data have the capability to detect methane plumes of large point sources. As such, a wide range of quantification approaches have been developed to quantify methane source rates based on this dataset. Most of the existing methods, however, require auxiliary data, such as wind speed, and have large uncertainties. In this study, we introduce a novel approach based on deep learning models to enhance the precision of methane quantification using Sentinel-2 data without the reliance on external data sources. To train the proposed deep learning model, a comprehensive benchmark dataset has been generated, using Sentinel-2 data. This dataset is created by integrating simulated plumes and background noise extracted from real Sentinel-2 images. This approach ensures the integration of realistic environmental conditions within the simulated data, enhancing the robustness and reliability of our proposed model. The generated benchmark dataset is utilized in different deep learning architectures, namely VGG-19, ResNet-50, Inception-v3, DenseNet-121, Swin-T, and EfficientNet-V2L, to estimate methane source rate. The performance of deep models has been evaluated in three learning strategies, namely from scratch, transfer-learning, and fine-tuning. The fine-tuned EfficientNet-V2L achieves the highest accuracy with root-mean-square error (RMSE), mean absolute percentage error (MAPE), and Pearson R of 2101 kg h−1, 10.05%, and 95.70%, respectively. More importantly, the proposed model demonstrates superior performance compared to conventional physical-based quantification methods (e.g., integrated mass enhancement) and recently developed deep learning model techniques (e.g., MethaNet). In particular, the proposed model exhibits an improvement of approximately 1287 kg h−1 in terms of RMSE, a 3.92% reduction in MAPE, and a 5.01% enhancement in R compared to the IME method. These results highlight the advancements achieved by the proposed approach in accurately quantifying methane emissions using Sentinel-2 imagery. The generated benchmark dataset and the developed deep learning model presented in this study serve as a fundamental resource and constructive framework for future research, promoting extensive implementation across various methane monitoring scenarios on different satellites and in distinct geographic regions, which delivering greater effectiveness to global methane emission monitoring initiatives.