Enhance the accuracy of rice yield prediction through an advanced preprocessing architecture for time series data obtained from a UAV multispectral remote sensing platform

High-resolution temporal spectral data captured by unmanned aerial vehicles (UAVs) have become increasingly important in predicting crop yields. Effective preprocessing of these temporal datasets is crucial for improving yield estimation accuracy and facilitating the broader application of predictive models. Despite its growing importance, a comprehensive guide detailing the preprocessing procedures for UAV temporal data is currently lacking. Consequently, this research is dedicated to constructing a robust preprocessing framework tailored to UAV time series spectral remote sensing data, with a particular emphasis on assessing its impact on the accuracy of yield predictions. We developed a multi-level threshold segmentation (MLT) method specifically for rice particle swarm optimization (ricePSO). Three field experiments were executed under diverse nutritional regimes to contrast the efficacy of yield predictions derived from UAV temporal dynamic threshold segmentation against those achieved through temporal data smoothing. Results showed that the ricePSO multi-level threshold segmentation outperformed the conventional Otsu threshold segmentation method, enhancing yield prediction accuracy by 1–11 %. Meanwhile, data smoothing effectively reduced errors in the temporal data acquisition process. Combining MLT, Gaussian smoothing, and the Bidirectional Long Short-Term Memory (Bi-LSTM) model resulted in the highest yield prediction accuracy, with an R² value of 87.52 %. Overall, this study achieved improvements in yield prediction accuracy through the use of multilevel dynamic threshold segmentation and data smoothing, providing new strategies for the preprocessing of temporal multispectral remote sensing data from UAV.