This study proposes a data-efficient framework for the automatic detection of methane ( \(CH_4\) ) emissions in urban landfill environments using hyperspectral satellite imagery and machine learning. Given the limited availability of labeled data, restricted to two hyperspectral cubes acquired by the EnMAP mission, a synthetic data augmentation strategy was developed to expand the training manifold by combining soil spectral signatures with controlled methane absorption patterns. Shallow Artificial Neural Networks (ANNs), optimized through Bayesian hyperparameter tuning, were employed to model the spectral response of methane while mitigating overfitting risks associated with deep architectures. The methodology was evaluated under both inter-scene generalization and intra-scene validation scenarios using repeated cross-validation. Results demonstrate that the proposed approach achieves high predictive accuracy and strong generalization across different temporal acquisitions, with robust detection of methane-related spectral features. Spatial prediction maps further confirm the model’s capability to localize methane plumes and distinguish them from background materials. The findings highlight the effectiveness of combining spectral feature engineering, data augmentation, and shallow learning models for environmental monitoring under data-scarce conditions, offering a scalable and computationally efficient solution for methane detection in real-world scenarios.