Land surface dynamics exhibit strong periodic patterns, yet their interannual variations pose challenges for accurate forecasting. This paper proposes a POD-AM-HAR framework that shifts prediction from point-wise to cycle-to-cycle forecasting. The framework integrates Proper Orthogonal Decomposition for dimensionality reduction, Harmonic analysis Ridge regularization for small-sample stability. Validation on 10-year MODIS NDVI data from Ningxia, China, shows the framework effectively captures interannual evolution and maintains robust performance with limited training cycles.