Low-light remote sensing is a remote sensing imaging technology for remote sensing observation in low-light conditions. Compared with traditional remote sensing technology, low-light remote sensing can be observed in low-light environments such as night, dawn, or dusk, and has extensive applications in urban lighting monitoring, night navigation, disaster monitoring, and other fields. However, in the case of weak lighting conditions at night, remote sensing cameras need a longer exposure gaze to ensure the desired signal-to-noise ratio (SNR). During long-time gaze exposure, the relative motion of low-orbit satellites and ground objects brings serious geometric distortion and motion blur to the image, which results in an imaging resolution decrease. To this end, we combine video snapshot compressive imaging (SCI) technology with remote sensing imaging technology and propose a low-light remote sensing imaging technology based on video SCI. This technology slices and reconstructs the exposure process of remote sensing sensors in low-light conditions, and shortens the exposure time of a single time frame. Additionally, the imaging SNR is ensured, and the geometric distortion and motion blur of the image are reduced.

According to the characteristics of low-light remote sensing imaging, we transform the low-light imaging problem into a low SNR imaging problem. Meanwhile, we add different levels of noise to the ideal image to simulate the low-light image in different lighting conditions. Based on the image motion theory and single-exposure compressive imaging principle, we propose a remote sensing imaging model for video SCI and an evaluation index for remote sensing images. 32 consecutive frames of remote sensing video in the VISO dataset are selected as the simulation original image, and Poisson noise corresponding to the given imaging conditions is obtained according to the low-light remote sensing imaging model. The simulated low-light images are then encoded and compressed according to the video single-exposure compressive imaging model. After the compression observation, the deep learning algorithm EfficientSCI with the best reconstruction effect currently is adopted for reconstruction, and the proposed remote sensing image evaluation index is employed to evaluate the imaging SNR.

According to the simulation reconstruction results and the fitting model, the conclusions are drawn as follows. In the illumination condition of equivalent entry pupil radiance of 1×10^-3 W/(sr·m²), the image quality after coding modulation and reconstruction increases with the rising camera integration time, and the relationship is basically a logarithmic function (Fig. 5). As the compression ratio rises, the overall reconstruction effect increases first and then decreases, and the increase in single frame integration time increases the amount of information in a single frame, with the peak reconstruction quality appearing at a lower compression ratio. Under the single frame integration time of 3.9, 9.8, and 19.6 ms, the peak SNR will appear at the compression ratio of 16, 12, and 10 respectively (Fig. 7). The acceleration of the target motion speed will cause a more serious degree of motion blur in remote sensing images. The signal in the image will be dispersed due to motion blur, which will reduce the signal intensity, thereby resulting in a higher level of noise relative to the signal, and a lower SNR in imaging (Fig. 8). The experimental results of the real prototype show that the integration time increase of the image sensor under the low-light environment is conducive to the target information accumulation, and the overall image brightening is accompanied by the improvement of image detail (Fig. 10). Additionally, the image quality increases first and then decreases with the rising compression ratio (Fig. 11), which is consistent with the simulated experimental results.

Based on the principle of video SCI and the theory of low-light remote sensing imaging, we propose a novel remote sensing imaging technology for video single-exposure compression, which can decompose a long exposure process into multiple short time frames, thus alleviating the image motion problem caused by long exposure imaging. Meanwhile, a remote sensing imaging model based on video compressive imaging is built, and the effects of the camera integration time, compression ratio, and target motion speed on imaging quality are explored via simulation experiments. The results show that the SNR of 8.45 dB, 8.31 dB, and 8.29 dB can be improved at 31.38, 78.46, and 313.86 ms respectively in the illumination condition of 1×10^-3 W/(sr·m²) equivalent entry pupil radiance of the remote sensing camera. Additionally, the resolution of remote sensing images increases with the rising camera integration time. When the integration time increases from 31.38 to 313.86 ms, the SNR rises by 5.88 dB. Under the single frame integration time of 3.9 ms, the increase in the compression ratio from 4 to 16 is conducive to the improvement of the reconstructed SNR, and under the compression ratio higher than 16, the reconstruction quality gradually decreases with the increasing compression ratio. Based on the simulation experiments, we verify the correctness of the imaging model and simulation results by conducting the principal prototype imaging experiment. These results can provide not only a solution to the image motion in low-light remote sensing but also theoretical guidance and engineering references for the design of new optical remote sensing.