Abstract: This paper proposes a fusion approach, which combines radius-basis-function artificial neural network model learning from global data with spatial interpolation model sampling from local data. The integration model minimizes the mean squared prediction errors by optimizing weights of their outputs and correcting the errors. It evaluates the fusion approach by cross-year-validation based on multi-angle imaging spectra radiometer (MISR) collected from January 2002 to December 2003 over 26 local regions in the continental US. It is noticed that fusion model can significantly reduce prediction error on land and desert.

Key words: Spatial target prediction, Fusion model prediction, Global model prediction, Local prediction model, Aerosol optical thickness

摘要： 提出了一种融合预测模型，把基于全局数据的径向基函数、人工神经网络与基于区域数据的空间采样插值相结合，并通过优化权重组合和修正误差，使得预测误差最小化。利用了多角度成像光谱辐射仪MISR，采集了从2002年—2003年美国大陆26个局部区域的辐射数据，对模型进行了2组气溶胶光学厚度的预测实验。地表特征因为反射能力的不同，导致了预测模型的不同复杂度。结果显示，融合模型能显著地减少陆地上的预测均方差。

关键词: 空间目标, 融合预测模型, 全局预测模型, 局部预测模型, 气溶胶光学厚度

CLC Number: 

• TP311.5