The classification of the line-of-sight (LOS) and non-line-of-sight (NLOS) signals is one of the major problems for robust GNSS positioning and the shadow matching in urban environments. The existing techniques include the multi-sensor integration, 3D map aid, using a dual-polarized antenna, an omnidirectional camera aid are proposed to solve the classification problem. However, they all require external hardware or up-to-date map, which is expensive or impractical for mass-market applications. Consistency checking with the receiver autonomous integrity monitoring (RAIM) is widely used for the detection of NLOS signals, but it is efficient only when the majority of the received signals are LOS signals. Machine learning methods, including the decision tree, the support vector machine (SVM) have been explored to classify LOS and NLOS with good accuracy. However, all current machine learning based method only utilize information within one epoch, all the inter-epoch information and data features in time series are lost, and the information of signal propagation in the complex urban environments is not fully manifested in the Rinex level observation and NMEA level observations in one single epoch. In this paper, a multivariate Long Short Term Memory Fully Convolutional Network (MLSTM-FCN) based signal classification method is proposed. With the aid of the convolution layer and long short term memory block, this method handles the data features in both time domain and value domain. Six time series features of GNSS signal, including differenced C/N0, time differenced ambiguity, double difference phase and pseudorange, phase and pseudorange consistency are analyzed and used as the input of the MLSTM-FCN. Datasets from two locations in the urban Calgary are collected, each of which is used for training and testing purposes respectively. The results reveal that, compared to the SVM classification method, the overall testing accuracy of the newly proposed classifier is improved from 93.00% to 95.97% for the Rinex level observation, and from 92.99% to 93.83 for the NMEA level observation. This improved classification accuracy brought by the proposed classifier is encouraging since it will enhance the robustness of the conventional GNSS positioning and the shadow matching based navigation system by reducing unbounded NLOS signal errors in urban environments and result in significant improvement in positioning accuracy. Compared to the SVM classifier aided single point positioning (SPP) test, the accuracy in the form of RMS of the MLSTM-FCN aided SPP test can be improved by 24.3%, 17.8% and 24.4% in the East, North and Up directions respectively, and the rate of the valid solution can be increased from 99.02% to 99.94%. The new method has the potential to be widely applied by various receiver types with the output of the raw observation or only with the NMEA observations output.