Study on the applicability of the CH4Net model and Sentinel-2 imagery in identifying the spatial distribution related to CH4 emissions in the Red River delta

Abstract

Methane (CH₄) is one of the important greenhouse gases, playing a significant role in global climate change due to its high heat-trapping capacity in the atmosphere [1]. Monitoring and predicting the spatial distribution of CH₄ remains challenging, stemming from the complex interaction between surface factors and atmospheric dynamics. In this study, a multi-source data integration methodology framework is proposed, combining Sentinel-2 optical imagery, CH₄ data from Sentinel-5P, and ERA5-Land meteorological data to classify CH₄ levels according to a spatial grid at the monthly scale. To ensure that the model evaluation accurately reflects the generalization capability, the study applies a spatially independent data partitioning strategy, in which the grid cells in the training and test sets do not overlap. The CH4Net-RRD deep learning model was constructed and used to extract spatial features from Sentinel-2 images, while meteorological variables were processed using a multilayer perceptron network. The two feature sources were combined to perform methane binary classification. Experimental results showed that the model using only Sentinel-2 achieved a Macro F1-score of 0.8100, reflecting average surface information extraction capabilities. When meteorological data were integrated, the model performance increased to 0.9128, indicating a significant supporting role of atmospheric variables. Notably, the model using only ERA5 achieved perfect performance (Macro F1 = 1.0000), suggesting that at the monthly grid scale, CH₄ distribution is primarily governed by meteorological conditions. Spatial analysis revealed that prediction deviations are not randomly distributed but concentrated in transitional regions. The research results clarify the dominant role of atmospheric factors in CH₄ distribution, while also emphasizing the importance of spatially independent assessment methods in environmental modeling problems using remote sensing.