Abstract: For nowcasting models based on the Convolutional Neural Networks (CNNs) used in radar echo extrapolation, different strategies are commonly applied to forecasting the radar echo of future (usually within two hours) multiple time steps. In this work, using the nowcasting of the atmospheric vertical integrated liquid water content, the forecast performances of models with two types of strategies, namely Recursive Forecast Strategy (RFS) and Direct Forecast Strategies (DFSs), were compared. CNN-based models were constructed for spatiotemporal forecasting with the UNet architecture as the backbone. Results exhibited the significantly better performance of the two DFS models than the RFS model on the overall forecast horizon with a roughly 19% lower root-mean-square error. With increasing forecast time steps, the forecast estimation errors of the RFS model accumulated much faster than those of the two DFS models. For the DFS models, the multioutput DFS (DFS-M) was more robust and performed better than the single-output DFS (DFS-S) on the overall forecast horizon; for a short forecast horizon, DFS-S has a slightly better forecast (approximately the first two time steps in the future). Moreover, a neural network interpretation method (i.e., DeepLIFT) was applied to the two DFS models to find the relative importance of each input time step. It was revealed that for both DFS models, approximately 80% of the mean importance score was in the last two input time steps and early input time steps have an increasing impact on longer forecast times. The input importance of DFS-M at each output step was much more stable than that of DFS-S due to the self-constrain effect exerted by the stochastic dependencies between output steps. By combining the forecast of two direct models at different time steps, a more balanced forecast can be performed on the whole forecast horizon.