Abstract: Cloud cover interferes with the ground information acquisition by remote sensing satellites, while remote sensing data missing caused by thick cloud cover can be effectively addressed using developed cloud removal algorithms. However, existing algorithms, which select similar pixels by computing differences between the central pixel and its spatiotemporal neighbors, suffer from logical defects when applied to the land surface with abrupt changes. Therefore, this study proposed an improved method for selecting similar pixels while accounting for temporal variations, aiming to enhance the performance of existing cloud removal algorithms. The improved method is detailed as follows: time-varying cluster centers of time-series clean pixels are extracted using the K-means clustering method; these cluster centers are used for supervised classification of cloud-free remote sensing images, followed by the inference of the temporal variation patterns beneath clouds; since it is considered that pixels with variation patterns consistent with those of the central pixel share the identical temporal variations, the intersections of these pixels and similar pixels identified using existing algorithms are the optimal similar pixels; finally, cloud removal is performed using the original algorithm based on the determined optimal similar pixels. Using the agricultural region in north Henan as a case study, the proposed method for selecting the optimal similar pixels was applied to two cloud removal algorithms: modified neighborhood similar pixel interpolator (MNSPI) and weighted linear regression (WLR). The results indicate that the improved method can accurately extract similar pixels from rapidly changing agricultural scenarios. Both cloud removal algorithms based on the selected optimal similar pixels outperformed their original algorithms in terms of both spectral differences and structural similarity. Specifically, the improved MNSPI algorithm reduced the average root mean squared error (RMSE) from 0.023 4 to 0.015 2 and increased the structural similarity index measure (SSIM) from 0.602 3 to 0.716 6 compared to its original algorithm. Meanwhile, the improved WLR algorithm decreased the average RMSE from 0.037 5 to 0.016 8 and increased the SSIM from 0.589 7 to 0.646 1 compared to its original algorithm. Applications to real remote sensing images demonstrate that both algorithms based on the proposed method can accurately reconstruct information beneath clouds, achieving encouraging visual effects. The results of this study provide a scientific basis for removing clouds in remote sensing images of rapidly changing land surfaces, thereby offering data support for high-quality time-series remote sensing monitoring.