A light use efficiency principle guided framework for detecting underground natural gas micro-leakage in multi-crop rotation area using hyperspectral imagery

Accurate detection of underground natural gas micro-leakage is essential for operational safety. However, existing spectral-spatial models based on hyperspectral imagery (HSI) face challenges of limited integration with crop physiology and low interpretability. To address this, we propose LUEGNet, a novel framework that integrates the SAFY-NG crop growth model with deep learning through light use efficiency (LUE) theory, enabling physics-guided detection. Specifically, the LUEGNet framework comprises two modules: (1) A natural gas stress detection module, which extracts vegetation stress features from HSI and meteorological data for effective detection of natural gas stress. By innovatively integrating the driving principle of the SAFY-NG crop model, this module employs a hybrid CNN-GRU architecture to combine high-level spatial feature extraction with temporal analysis to accurately simulate the natural gas stress factor (Ka). (2) A feature interpretation module, which facilitates post-model interpretability analysis. This module innovatively combines regression activation mapping with stress-stage analysis to decode feature contributions. Results demonstrate that LUEGNet achieved high-accuracy simulation of the natural gas stress factor Ka, with an R2 of 0.96 and an nRMSE of 6.66 % on the combined wheat–maize dataset. Interpretability analysis further revealed that spectral features dominated stress detection, contributing up to 83.79 % in wheat and 73.71 % in maize. By effectively integrating physiological knowledge with deep learning, LUEGNet not only improves detection accuracy but also provides mechanistic transparency. These findings validate LUEGNet as a robust, interpretable solution for long-term detection of natural gas micro-leakage using near-surface HSI, significantly enhancing environmental monitoring and operational safety in gas storage facilities.