Abstract: The mutual constraints between temporal and spatial resolution are common challenges in the quantitative remote sensing monitoring of resources and the environment. High-temporal and high-spatial resolution NDVI time series data play a vital role in supporting decision-making across a range of domains, including surface vegetation monitoring, agricultural management, phenology change analysis, disaster early warning, and land use change mapping. In the field of agricultural engineering in particular, the spatiotemporal continuity of NDVI data is critical for improving the accuracy of crop yield prediction and enhancing the efficiency of farmland resource management. However, in mountainous regions, optical satellite imagery is often affected by cloud and fog, leading to incomplete data coverage and results in spatiotemporal discontinuities that challenge accurate vegetation monitoring. Traditional spatiotemporal fusion algorithms for remote sensing face limitations, including reduced accuracy in areas with complex terrain and heterogeneous landscapes, as well as difficulties in scaling to regional levels. To address these challenges, this study proposes a high spatiotemporal resolution NDVI reconstruction method (object-level gap filling and savitzky-golay filtering method, OLF-SG) based on an object-level gap filling strategy. This method fuses MODIS and Sentinel-2 data to reconstruct high-quality Sentinel-2 NDVI time series. A weighted Savitzky-Golay filter is applied to smooth the reconstructed NDVI time series, eliminating noise errors and generating high-resolution NDVI products with an 8-day temporal resolution and 10-meter spatial resolution. By using MODIS NDVI as a reference and introducing an object-level gap filling strategy to replace the traditional similar pixel filling method, it not only effectively reduces the computational complexity, but also avoids the memory limit problem caused by high computational load in the GEE environment. Two heterogeneous underlying surfaces, referred to as Area A (typical piedmont fragmented landform, rich vegetation types and obvious differentiation) and Area B (agricultural land-dominated area, complex planting structure), were selected within the Erhai Lake Basin to evaluate the reconstruction performance under different surface characteristics. Compared with the traditional Gap Filling and Savitzky-Golay filtering method (GF-SG) and the Object-Level Spatial and Temporal Adaptive Reflectance Fusion Model (OL-STARFM), the proposed OLF-SG method provides the reconstructed NDVI images closest to the reference images, reconstruction accuracy and efficiency. It also avoids the problem of base image pair selection and makes full use of the cloud-free observation data of pixels, thus improving the utilization rate of data. The average deviation (AD) of the reconstructed Sentinel-2 NDVI image and the reference image was as low as 0.039 and 0.006, respectively, the boundary clarity was optimal (Edge was -0.130 and -0.094, respectively), indicating the optimal stability of the model. The experiments were qualitatively and quantitatively verified the effectiveness of the model. In Area A, the OLF-SG achieved a root mean square error (RMSE) of 0.060 and a coefficient of determination (R²) of 0.984. In Area B, the RMSE was 0.068 and the R² reached 0.952. The high spatiotemporal resolution NDVI time series generated by the OLF-SG effectively captured the phenology patterns across complex landscapes. The strong potential can also be offered for a wide range of applications, including crop growth monitoring, vegetation phenology analysis, soil erosion assessment, and agricultural management.