【Objective】This research aims to resolve the limitations of existing forest fire recognition algorithms in temporal feature utilization and computational efficiency, this study proposes a video-based recognition model (C3D-ViT) to enhance both detection accuracy and operational efficiency in practical forest monitoring scenarios.【Method】We presented a hybrid architecture integrating 3D Convolutional Neural Networks (3DCNN) with Vision Transformer (ViT). The framework emploied 3D convolution kernels to extract spatiotemporal features from video sequences, which were subsequently tokenized into vector representations. Vision Transformer’s self-attention mechanism then globally models feature relationships across temporal and spatial dimensions, with final classification achieved through the MLP Head layer. Comprehensive ablation studied and comparative experiments were conducted against ResNet50, LSTM, YOLOv5, and baseline 3DCNN, ViT models.【Result】The C3D-ViT achieves 96.10% accuracy, outperforming ResNet50 (89.07%), LSTM (93.26%), and YOLOv5 (91.46%), and has improved compared to the accuracy of the original 3DCNN and Vision Transformer (93.91%, 90.43%). The improved C3D-ViT model performs better in recognition performance, with high recognition accuracy and stability under unfavorable conditions such as occlusion, long distance, and thin smoke. The demand for real-time detection can be realized.【Conclusion】The C3D-ViT framework effectively addresses spatiotemporal modeling challenges in wildfire detection through synergistic CNN-Transformer interaction, providing a technically viable solution for next-generation forest fire early warning systems.