A new method for forecasting non stationary series is developed. Its steps are as follows: Step 1. Data delaminating. Non stationary series is delaminated into several multi scale steady data layers and one trend layer. Step 2. Modeling and forecasting each stationary data layer. Step 3. Imitating trend layer using polynomial. Step 4. Combining the forecasting layers and imitating layer into one series. The EMD (Empirical Mode Decomposition) method suitable to process non stationary series is selected to delaminate data, while ARMA (Auto Regressive Moving Average) model is employed to model and forecast stationary data layer and least square error method for trend layer regression. Aiming at forecasting length, forecasting orientation and selective method, experiments are performed for SAR (Synthetic Aperture Radar) images. Finally, an example is provided, in which the whole SAR image is restored via the method proposed by this paper. A new method for forecasting non stationary series is developed. Step 2. Data delaminating. Non stationary series is delaminated into several multi scale steady data layers and one trend layer. Step 2. Modeling and forecasting each stationary data layer Step 3. Imitating trend layer using polynomial. Step 4. Combining the forecasting layers and imitating layer into one series. The EMD (Empirical Mode Decomposition) method suitable to process non-stationary series is selected to delaminate data, while ARMA (Auto Regressive Moving Aiming at forecasting length, forecasting orientation and selective method, experiments performed for SAR (Synthetic Aperture Radar) images. Finally, an example is provided, in which the whole SAR image is restored via the method proposed by this paper.