A modified pi-sigma fuzzy neural network error calibration model is designed to decrease measurement error of radiosonde humicap in upper atmosphere especially in low temperature environment.K-means clustering algorithm and weights direct determination method is implemented to improve network performance.Practical test result is compared with BP neural network

it shows that the maximum relative error of 144 groups of train samples at temperature of-30

4

0 ℃ are 4.774 %

15.2 % respectively

convergence time are 0.01

2s respectively.Result of 4 groups of test samples demonstrates that pi-sigma fuzzy neural network effectively realizes temperature compensation and nonlinear correction

and is prior to BP network in predicting precision

generalization

ability and training speed.