CELNet: A comprehensive efficient learning network for atmospheric plume identification from remotely sensed methane concentration images

Abstract

Methane is an important greenhouse gas contributing to global warming and climate change. The effective identification of atmospheric plumes in spatial images of methane concentration data retrieved from remote sensing is a critical step in quantifying emissions and ultimately helping to mitigate climate change by reducing large methane emission sources. In this paper, we propose a comprehensive efficient learning network (CELNet) for atmospheric plume detection, which is constructed with several deep neural modules and detects the shape of plumes in methane concentration images effectively. Specifically, to conduct an efficient plume identification, a generative module is constructed, which is tasked to generate feature maps for the characterisation of potential plumes in remotely sensed methane concentration data. This helps to reduce the search space in the detection implementation. Methane plumes in remotely sensed image data normally exhibit complex morphological structures with high background noise, which can interfere with the delineation of the shapes of plumes. Thus, the generative module alone cannot guarantee an accurate identification. To conduct high quality methane plume delineation, an extractor module is introduced to extract features that intrinsically characterise methane plumes in remotely sensed image data. The extracted intrinsic features are encoded using an encoder module for compact representation, which convey important information for implementing a better methane plume delineation. In particular, to enhance the capability of the generative module for generating accurate features, we structurally pair it with a discriminative module. In the training process, the discriminative module takes the generated features and the intrinsic features as inputs and improves its capability to discriminate the generated features from the intrinsic ones, whereas the generative module strives to generate accurate features that the discriminative module is unable to identify. They thus build an adversarial game which is beneficial for enhancing the feature generation capability of the generative module during the training process. The generated features along with the intrinsic features are then fed into the decoder module to produce accurate methane plume detection maps, where the intrinsic features incorporated provide additional supervision information that enables the CELNet to perform a more effective methane plume identification. We validate the proposed technique with different types of remote sensing image datasets (e.g., Landsat 5, AVIRIS-NG), and the accuracy achieved by CELNet outperforms the other comparison methods over 6%. This highlights its applicability for different sourced images with high performance, making it valuable for remote sensing community.