A resource efficient neural network based gas classifier using the 1.5-bit quantization of sensing channel difference as the feature extraction is proposed in this paper, which is designated for unattended electronic noses for long-term surveillance. The feature rate of the proposed method is as low as 48 bits per second (bps), significantly reducing the computational complexity of the classifier compared with state-of-the-art works. Based on the simple and identifiable features, a slight fully connected neural network (SFCN) is proposed as a gas classifier. The parameters and floating-point operations per second (FLOPs) are reduced 41 and 624 than state-of-the-art gas classifiers, respectively. A self-recorded gas dataset with 9 types of gases and 64413 samples is used to validate the performance of the proposed feature extraction method and classifier. Simulation results show that this method can classify 9 different gases with an accuracy of 97.2%, which is comparable to the state-of-the-art works. Meanwhile, the parameters and the computation complexity are greatly reduced, which makes it suitable for long-term electronic noses with computation resource constraint.