Recent article demonstrated that hyperspectral imagery can indirectly identify underground natural gas microleakage through the stress symptom of surface vegetation. Leveraging the abundant spectral and spatial information in hyperspectral imagery, a novel approach combining stacked autoencoder (SAE) and multiscale three-dimensional convolutional neural network (MS3D CNN) was proposed for stress identification of grass, soybean, corn, and wheat. The SAE was applied to learn low-dimensional representations of hyperspectral data with more discriminative ability. Subsequently, the MS3D CNN extracted spectral–spatial features from stressed vegetation. Results showed that the MS3D CNN achieved superior overall accuracy (OA) for the four vegetation species: grass (95.05%), soybean (95.18%), corn (97.68%), and wheat (93.75%). Compared with full-band hyperspectral data, low-dimensional data as input significantly reduced the total parameters of the MS3D CNN and its OA demonstrated a competitive performance (grass: 93.25%, soybean: 91.36%, corn: 94.02%, and wheat: 91.04%). Furthermore, the minimum circumscribed circle was fitted to analyze the temporal and spatial patterns of stress identification results based on the concentric spatial distribution of natural gas diffusion, after which the detection accuracy of leakage points was less than 0.5 m without false alarms. The proposed method is effective and applicable for intelligent underground natural gas microleakage identification.