Abstract: Infrared remote-sensing spectroscopy is an important technology for the rapid detection of chemical warfare agents on battlefields. Traditional spectral recognition algorithms, such as error backpropagation neural networks and support vector machines, have difficulty in learning the infrared spectral features of target objects on a global scale. Convolutional neural networks have strong feature-extraction capabilities for sample data and are widely used in fields such as object recognition. This study is based on a small amount of measured spectral data, Gaussian function fitting simulations were conducted to obtain pure spectral data of four chemical warfare agents and 28 toxic and harmful gases. By adding baseline and random noise, a large number of samples with significant differences were obtained and subsequently divided into training and validation sets. The convolutional neural network was trained, and the network parameters were optimized. The results indicate that without the need for data preprocessing, the recognition accuracy of the model for the test set reached 98.83%. Therefore, combining data augmentation methods and convolutional neural networks is effective in qualitatively recognizing chemical warfare agents and toxic and harmful gases, thereby providing a new approach for the recognition of infrared remote sensing spectra.