Phenology-preserving temporal reconstruction of satellite-derived NDVI via morphological operations for unsupervised vegetation clustering

• Morphology-based method is proposed for temporal NDVI reconstruction. • It preserves low-frequency phenological signals under cloud contamination. • It outperforms moving average, Savitzky-Golay, and HANTS in spectral fidelity. • It enhances vegetation clustering stability and separability in unsupervised analysis. Cloud contamination remains a major challenge in the analysis of satellite-derived Normalized Difference Vegetation Index (NDVI) time series, particularly in dryland and semi-arid ecosystems where phenological signals are sparse and irregular. This study investigates temporal reconstruction of NDVI under quality masking–induced data gaps, with a specific focus on preserving low-frequency phenological structure rather than maximizing pointwise accuracy. We propose a fully unsupervised reconstruction strategy based on one-dimensional flat morphological closing applied along the temporal dimension, and systematically compare it against common baseline methods, including moving average smoothing, Savitzky-Golay filtering, and harmonic analysis of time series (HANTS). Reconstruction fidelity is first evaluated under controlled cloud simulations using spectral-domain metrics derived from dominant annual and intra-annual harmonics. At a noise level of 0.3, morphological reconstruction achieves a spectral fidelity of 0.93 and an RMSE of 0.02, compared to spectral fidelity value of 0.81 and RMSE of 0.073 for the strongest competing method. The practical implications of reconstruction fidelity are then assessed through unsupervised clustering of real Sentinel-2 NDVI time series. Clustering performance is evaluated using F1-score and precision, both with and without spectral feature augmentation derived from low-order Fourier amplitudes. Morphological reconstruction achieves a mean F1-score of 0.68 compared to 0.59–0.62 for baseline methods and shows minimal improvement (<0.02) after spectral augmentation, indicating that dominant phenological information is already preserved. In contrast, competing methods gain 0.05–0.08 in F1-score after augmentation, suggesting compensation for spectral distortion. Together, these results demonstrate that morphological temporal reconstruction provides a simple, parameter-light, and phenologically consistent alternative for quality-based masking mitigation in NDVI time series, with measurable advantages for downstream unsupervised analysis