遥感的本质在于反演。传统的遥感反演通常假设回归参数、模型误差服从正态分布，这个假设比较主观。 本文首次提出了由Bootstrap方法寻找回归参数和模型误差的先验分布的反演方案。同时，对按照地物分类后的先验数据作统计假设检验，说明了将先验知识分类的合理性。最后，以核驱动模型RossThick-LiTransit组合为例，用NOAA-AVHRR观测数据对结合该方法的反演结果与传统的Tikhonov正则化反演，Bayes反演结果做比较，表明了对先验知识分类和结合Bootstrap方法的遥感反演能明显减少反演结果的不确定性。 同时，鉴于通常的定量遥感着重于寻找反演算法来解决遥感模型中参数 欠定问题，而对遥感模型本身的合理性探讨较少。本文基于 73组观测数据对遥感领域中广泛使用的一类线性模型进行合 理性检验，结果表明，在模型中考虑交互效应比传统的线性模型更合理。

It is usually assumedthat the prior distributions of parameters and error are Gaussiandistribution in remote sensing inversion, This assumption seems tobe impractical in many cases. Prior distribution of parameters anderror are very important in remote sensing inversion since manyremote sensing inversion strategies are based on prior knowledge. Wepresent a bootstrap method for estimating the prior distributions ofparameters and error in this paper. This method relaxes thedistribution assumption of parameters and error, and obtains thosedistribution by means of priori data. Moreover, we classify priordata since they are collected from different vegetational zone, andimplement statistical test for classified prior data, results showthat proper classification of priori data is reasonable. Finally, wetake RossThick-LiTransit linear kernel-driven model as an example,and make a comparison with usual Tikhonov regularizing inversion andBayes inversion under normal hypothesis with NOAA-AVHRR observationdata. The result shows that classifying prior data and using theprior distribution obtained by bootstrap method can significantlydecrease uncertainty of parameters.