Forecasting NDVI in Patagonian meadows using statistical and machine learning models

Patagonian pastoral systems rely heavily on meadows (mallines). These ecosystems are vital for livestock production and ecological resilience but face increasing threats from climatic variability and overgrazing. To support adaptive rangeland management, timely forecasts of vegetation productivity are essential. This study evaluates the potential of statistical and machine learning models to predict vegetation dynamics using the Normalized Difference Vegetation Index (NDVI) across ten meadows in northwestern Patagonia. We assess three main issues: (i) different predictive models, (ii) the role of preprocessing filters in improving model performance, and (iii) changes across four forecasting horizons: short-term (16days), seasonal-term (1–3 months), medium-term (4-7months), and long-term (8–12 months). Results showed that model performance varied with the prediction horizon. For long-term forecasts, simple autoregressive models outperformed more complex approaches, while for shorter horizons, the combination of a sinusoidal (SIN) filter with autoregressive models (AR) improved accuracy. At intermediate horizons (1–6 months), the Long Short-Term Memory (LSTM) recurrent neural network model delivered the most consistent performance, showing slower error accumulation and better adaptation to temporal dynamics. However, no single method was universally optimal; rather, model suitability depended on the interplay between temporal scale, preprocessing strategy, and site characteristics. We found that a SIN filter with Random Walk model was the best for short-term forecasts, while LSTM with raw data and SIN filter with AR offer reliable options for seasonal and mid-term applications. The proposed framework provides a foundation for operational NDVI forecasting in rangelands and supports adaptive management strategies under increasing climatic uncertainty.