With the intensification of global warming, the monitoring of methane
emission and detection of gas plumes from landfills have increasingly received
attention. We decompose methane emission monitoring into three sub-tasks:
methane concentration inversion, plume segmentation, and emission rate
estimation. Conventional algorithms have limitations: methane concentration
inversion usually uses the matched filter, which is sensitive to global
spectrum distribution and contains a large amount of noises. There is limited
research on plume segmentation, with many studies resorting to manual
segmentation that is likely to be subjective. The estimation of methane
emission rate often utilizes IME algorithm, which relies on obtaining
meteorological measurement data. Using the WENT landfill site in Hong Kong and
PRISMA hyperspectral satellite imagery, we propose a new deep learning-based
framework for quantitative monitoring of methane emissions from remote sensing
images based on physical simulation. We generate simulated methane plumes using
large eddy simulation (LES) and different concentration maps of fugitive
emission using the radiative transfer equation (RTE), while combining
augmentation techniques to create a simulated PRISMA dataset. We train a U-Net
network for methane concentration inversion, a Mask R-CNN network for methane
plume segmentation, and a ResNet-50 network for methane emission rate
estimation. All three deep networks achieve higher validation accuracy compared
to conventional algorithms. We further respectively combine the first two
sub-tasks and the last two sub-tasks to design the multi-task learning models -
MTL-01 and MTL-02, both of which achieve higher accuracy than single-task
models. Our research serves as a demonstration of applying multi-task deep
learning to quantitative methane monitoring and can be extended to a broad
range of methane monitoring tasks.

通过基于物理仿真的遥感影像监测方法，我们提出了一种新的基于深度学习的定量甲烷排放监测框架，该框架包括甲烷浓度反演、甲烷烟团分割和甲烷排放速率估计三个子任务，并引入多任务学习模型来提高准确性。