Reconstruction of Daily MODIS 250-m Two-Band Enhanced Vegetation Index Time Series Using Gaussian Process Regression

High temporal resolution vegetation index (VI) time series provide detailed insights into vegetation growth and significantly enhance the reliability of environmental applications such as vegetation phenology metrics extraction and dynamic analysis. However, using high temporal resolution often pose greater reconstruction challenges compared to lower resolution, such as Maximum Value Composite (MVC) processed data, as the complexity and prevalence of contaminated or invalid data points in daily VI time series can substantially reduce the robustness of reconstruction algorithms. In this study, we introduce Gaussian Process Regression (GPR), a non-parametric temporal dynamic model, to reconstruct daily MODIS 250 m two-band Enhanced Vegetation Index (EVI2) time series for the Yangtze River Basin in 2020. To evaluate the method's performance, we compare it with widely used methods such as the Savitzky-Golay (SG) filter, Best Index Slope Extraction (BISE), Double Logistic (DL) function fitting, and Harmonic Analysis of Time Series (HANTS) in both real and simulation reconstruction experiments. In the spatial performance evaluation based on MOD09GQ EVI2 time series reconstruction, GPR has better Variance of Laplacian (VL = 0.0011) and Structural Similarity Index Measure (SSIM = 0.7491). In the temporal performance evaluation in the sample set that consists of 1774 time series, GPR achieves the smallest Roughness Index (RI = 3.221 × 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−5</sup>). In simulation experiments, the reconstructed time series of GPR has smallest Root Mean Square Error (RMSE = 0.0315) and Spectral Angle Mapper (SAM = 0.0698) to the synthetic ground truth. Additionally, we examined the methods' robustness to randomly distributed and consecutive gaps through simulation experiments, with GPR consistently showing the smallest RMSE and SAM in all conditions. In the ground validation where the transition dates derived from reconstructed EVI2 are compared with green-chromatic coordinate (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G_CC</i>) transition dates, GPR achieves highest R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> as well. These findings indicate that GPR is an ideal algorithm for daily time series reconstruction.