Forecasting NDVI Variability Using SPI-Driven Hybrid Deep Learning in a Semi-Arid Environment.

Effective land management in semi-arid regions is contingent upon the accurate forecasting of vegetative alterations in response to climatic variations. This research utilize the CNN-LSTM model as a hybrid deep learning framework to predict fluctuations in Normalized Difference Vegetation Index (NDVI) using lagged Standardized Precipitation Index (SPI) and NDVI inputs. The objective of the model is to capture the enduring memory effects of vegetation that impact plant growth, as well as to account for short-term variations in precipitation. A dataset comprising MODIS NDVI and monthly SPI data from 2001–2022 was developed for the region of Semnan, Iran, which is characterized by its extreme aridity. After extensive preprocessing, various configurations of NDVI and SPI lags were systematically assessed. The optimal performance was obtained utilizing one-month SPI values with both 1- and 2-month time lags, in conjunction with a 1-month NDVI lag, resulting in notable accuracy (RMSE = 0.0038; r = 0.968).
The application of explainable artificial intelligence methodologies—including SHAP, LIME, and Random Forest feature importance—validated that NDVI lag-1 consistently emerged as the most significant predictor across all analytical approaches. Additionally, SPI lags made substantial contributions, with SPI-1 generally demonstrating a more pronounced impact than lags associated with longer precipitation durations. These results underscore the pivotal influence of short-term vegetative memory and recent precipitation anomalies in determining the dynamics of NDVI within dryland ecosystems.