The Normalized Difference Vegetation Index (NDVI) is widely used as an indicator of irrigation activity in arid and semi-arid agricultural regions. This study evaluates how NDVI extraction scale and statistical metric influence groundwater level prediction accuracy in the irrigated Tadla Plain, MoroccoA total of 96 Long Short-Term Memory (LSTM) models were developed by combining six NDVI extraction scales, from the well pixel to 20,000 m buffers, and four statistical metrics: mean, median, maximum, and minimum. Results demonstrate that extraction scale is a critical factor controlling model performance. Across the four monitored wells, larger buffers (≥2500 m) generally outperformed smaller ones (≤1000 m), with the optimal scale occurring at 15,000 m for three wells, while one well achieved its best performance at the 1000 m scale. The best models achieved RMSE values between 0.09 and 0.625 m and R² values ranging from 0.94 to 0.997. Maximum NDVI provided the highest predictive accuracy for three wells, whereas minimum NDVI performed best for one well. Statistical analyses further confirmed that extraction scale generally exerts a stronger influence on prediction performance than the choice of NDVI metric. Spatial validation revealed that irrigated areas more than doubled between 2001 and 2017, and model performance improved as NDVI captured this broader irrigation footprint. These findings suggest that groundwater level variations in the Tadla Plain are more strongly associated with NDVI signals extracted at broader spatial scales than with strictly local vegetation conditions around individual wells, highlighting the importance of optimizing both NDVI extraction scale and metric for groundwater forecasting in irrigated semi-arid regions.