A compressive hyperspectral video imaging system using a single-pixel detector is introduced, enabling high-throughput hyperspectral video recording at low bandwidth. The system leverages the high compressibility of 4D hyperspectral videos compared to 2D images, using a joint spatial-spectral encoding scheme to capture compressed measurements while maintaining temporal correlation. A reconstruction method combining signal sparsity in 4D space and deep learning accelerates the recovery process. The system achieves a 900× data throughput compared to conventional imaging, reconstructing 128×128 hyperspectral images with 64 spectral bands at 4.3 frames per second. The system uses a single-pixel detector with a structured random STOne pattern sequence for spatial modulation and pseudo-random Walsh-Hadamard patterns for spectral modulation. An optical flow-assisted 4DTV regularization model enhances temporal-spatial-spectral data recovery from compressive measurements. The system also includes a deep learning approach for fast reconstruction, significantly reducing reconstruction time. Experimental results demonstrate the system's ability to reconstruct high-quality hyperspectral video with a 900:1 compression ratio, achieving 157 frames of 128×128×64 hyperspectral data from 184,000 single-pixel measurements. The system's performance is validated using both simulation and experimental data, showing improved spectral accuracy and faster reconstruction times compared to traditional methods. The system's unique design allows for multi-resolution reconstruction and efficient data acquisition, making it suitable for applications requiring high-resolution, high-speed hyperspectral imaging with limited bandwidth.A compressive hyperspectral video imaging system using a single-pixel detector is introduced, enabling high-throughput hyperspectral video recording at low bandwidth. The system leverages the high compressibility of 4D hyperspectral videos compared to 2D images, using a joint spatial-spectral encoding scheme to capture compressed measurements while maintaining temporal correlation. A reconstruction method combining signal sparsity in 4D space and deep learning accelerates the recovery process. The system achieves a 900× data throughput compared to conventional imaging, reconstructing 128×128 hyperspectral images with 64 spectral bands at 4.3 frames per second. The system uses a single-pixel detector with a structured random STOne pattern sequence for spatial modulation and pseudo-random Walsh-Hadamard patterns for spectral modulation. An optical flow-assisted 4DTV regularization model enhances temporal-spatial-spectral data recovery from compressive measurements. The system also includes a deep learning approach for fast reconstruction, significantly reducing reconstruction time. Experimental results demonstrate the system's ability to reconstruct high-quality hyperspectral video with a 900:1 compression ratio, achieving 157 frames of 128×128×64 hyperspectral data from 184,000 single-pixel measurements. The system's performance is validated using both simulation and experimental data, showing improved spectral accuracy and faster reconstruction times compared to traditional methods. The system's unique design allows for multi-resolution reconstruction and efficient data acquisition, making it suitable for applications requiring high-resolution, high-speed hyperspectral imaging with limited bandwidth.