Abstract: During the simulation of thermal infrared remote sensing, the high spatial resolution scene of land surface temperature can be estimated by moderate and lower resolution thermal infrared data. The GA SOFM (Genetic Algorithms & Self Organizing Feature Maps) Artificial Neural Network (ANN) can be used to construct the relation between the inverted land surface parameters based on VNIR data and lower resolution data, which is also considered the unmixing process of mixed pixel. Finally, the high resolution land surface scene can be generated by this method. In this paper, the discussion and analysis the accuracy, applicability and prospect about this method are carried out. It is easy to put into operation with higher accuracy. Utilizing the ASTER data to test it, conclusions show that subpixel land surface temperature under different land cover types can be retrieved based on a pair of remote sensing data if we don't directly invert high resolution land surface temperature or run short of experienced knowledge about land surface. Also, it is a new approach to quickly estimate and simulate high resolution land surface temperature.

Key words: subpixel, land surface temperature, land surface parameters, GA SOFM(genetic algorithms & self organizing feature maps), artificial neural network

摘要： 在热红外遥感成像模拟中，高空间分辨率的地表温度场景可以由中、低分辨率的热红外遥感数据估算得出。基于可见光 近红外数据反演的若干地表参量和低分辨率的地表温度数据，在二者间引入遗传自组织神经元网络，建立非线性像元分解方法，最终获得高空间分辨率的地表温度场景。利用ASTER卫星产品数据对该方法进行了验证，结果表明： 对于无法直接进行高分辨率地表温度反演，或缺少大量地表先验知识情况下，该方法只需利用两组遥感数据即可估算出不同地表覆盖下子像元地表温度，方法简便易行，精度较高，为快速模拟和估算高分辨率地表温度分布提供了一条新途径。最后对方法的估算精度、适用性及应用前景进行了探讨。

关键词: 子像元, 地表温度, 地表参量, 遗传自组织特征映射, 神经元网络