In the explosive scene investigation, the rapid detection and accurate identification of fireworks and explosives play a vital role in preventing, controlling, and rapidly disposing of major explosions. However, the current rapid detection methods for fireworks and explosives mostly have problems such as low recognition speed and difficulty in visualization. Because of this, this paper proposes a method based on hyperspectral imaging technology combined with one-class support vector machine (OCSVM) for rapid detection and recognition of fireworks. Firstly, the hyperspectral data of the sample in the 400~720 nm band were collected with a hyperspectral camera. Principal component analysis (PCA) was used to reduce the dimension of the data, multiplicative scattering correction (MSc) was used to eliminate the baseline offset caused by particle scattering on the sample surface, and Savitzky-Golay (SG) was used to smooth the high-frequency noise and improve the spectral signal-to-noise ratio. Secondly, to reduce the complexity of the model and improve the efficiency, the representative pyrotechnic samples were selected from the spectral data by Kennard stone (K-S) method as the data set, which was divided into the training set and the test set at the ratio of 4∶1. On this basis, the OCSVM model was established. Thirdly, to verify the recognition ability of the model to the pyrotechnic composition, the isolated forest (iforest) and self-encoder (AE) models were established using the same training set, and the recognition ability of the three models to the pyrotechnic composition was compared. Finally, the recognition result is mapped to the RGB image of the test material, and the recognition image is obtained by marking the target pixels with the operation of the mask to realize the visual recognition effect of the pyrotechnic composition. The results show that the overall accuracy of the OCSVM method is higher than 0.95, the F1 score and AUC value are more than 0.8, and the recognition time is less than 2 seconds. The performance of OCSVM in classification accuracy, running speed, F1 score, and AUC is better than that of the isolated forest and self-encoder models. In terms of visual recognition, the recognition image based on the OCSVM model after mapping and mask operation can more accurately reflect the distribution of smoke and powder in all samples. At the same time, the recognition image based on an isolated forest and a self-encoder model can not well reflect the distribution of smoke and powder on yellow paper and black polyester cloth. The research shows that the pyrotechnic identification method based on hyperspectral imaging combined with OCSVM proposed in this paper has the characteristics of high recognition accuracy, fast response speed, and strong generalization ability, and can quickly, accurately, and nondestructively identify pyrotechnics in the test material. Its recognition accuracy, recognition speed-, and visualization effect can be well applied to the rapid discovery and on-site detection of pyrotechnic and explosive at the explosion scene, and provide an effective method for searching for pyrotechnic and explosive in the scene investigation.