<p>Accurate lightning nowcasting is a critical and complex challenge in meteorological forecasting. Deep learning models have shown great potential in lightning nowcasting. However, the ability of deep learning to capture the high-dimensional and complex spatiotemporal correlations in meteorological data still needs to be improved. Furthermore, it is also necessary to enhance the interpretability of the deep learning model. This paper proposes a novel spatiotemporal sequence forecasting model, STFNet. The model adopts a structure that separately extracts spatial and temporal features and innovatively designs the TFBlock. By using strip depthwise convolutions to approximate large-kernel convolutions and incorporating adaptive channel-wise weight allocation, the TFBlock extracts long-range temporal dependencies and motion features in a parallelized manner, thereby enhancing the capability of STFNet to characterize the temporal evolution of lightning. Using satellite data and lightning location data, this study conducts lightning nowcasting experiments in southwest China. Compared with other deep learning models, STFNet achieves the highest TS score and ETS score of 0.1304 and 0.1296, respectively. This paper employs channel combination control experiments and permutation tests to conduct a feature importance analysis, aiming to explain the contribution of input features to lightning nowcasting. The channel combination control experiments reveal that brightness temperature differences between channels are more critical for prediction than single-channel brightness temperatures. And permutation tests prove that cloud optical thickness contributes significantly. STFNet not only exhibits strong forecasting performance but also provides insights into the key factors influencing lightning occurrence, offering an effective new approach for lightning nowcasting.</p>