In order to improve the accuracy of sampling and reduce resource consumption in network availability testing, a new kind of adaptive sampling method is proposed in this paper. It adjusts sampling probability over time according to network condition. Its basic sampling method is simple random sampling. The network availability is predicted based on exponential smoothing method. Comparing the measured value of availability with the predicted one, we can decide whether the sampling parameter now used is suitable for the network condition. A liner rule is advanced to decide how to adjust the sampling probability over time. Taking the tactical Internet as an example, efficiency of this sampling method is compared with that of simply random sampling. The result shows that the accuracy of the sampling method is higher than simply random sampling when the size of sample is equal, and the result would be more obvious when traffic burst occurs.