Identifying Land Cover Change Via MAD-Based Outlier Detection in NDVI Time Series

Vegetation phenology-based studies have gained significant attention across research communities, particularly with increasing concerns about climate change and its impact on ecosystems. Remote sensing datasets provide valuable insights into vegetation’s spectral responses over time and space, reflecting climatic variations and environmental changes. To accurately analyze these datasets, effective data preprocessing is essential to reduce atmospheric disturbances, sensor noise, and eliminate data gaps. This study presents an enhanced Savitzky-Golay filter, termed as the Iterative Weighted Savitzky-Golay (IWSG) method, which reconstructs NDVI time series data. This method preserves the local fluctuations in vegetation patterns while filling data gaps and filtering out noise, ensuring precise data analysis. Following preprocessing, this study employs a simple algorithm for detecting shifts in NDVI patterns by identifying the number of anomaly years at pixel level, based on the median absolute deviation (MAD) score. By overlaying these results on satellite images, anomalies were contextualized, to understand land cover dynamics. Validation across various land cover classes confirmed the method’s consistency with observed NDVI signature changes, demonstrating that the MAD-based anomaly detection provides a robust framework for monitoring long-term land cover changes at pixel level using time series data. In contrast, the BFAST algorithm, though popular for detecting structural breaks, overestimated breaks for all sampled pixels, highlighting its limitations for large spatial datasets. Overall, MAD-based anomaly detection algorithm offers a simple and reliable framework for monitoring long-term land cover changes for extensive spatial area using time series data.