Lightning is a high-impact meteorological hazard associated with severe convective weather, posing substantial threats to socioeconomic activities and public safety. Timely and accurate lightning forecasting and warning with high spatiotemporal resolution are therefore of great importance for disaster prevention and mitigation. This paper systematically reviews existing research on observation-based lightning forecasting methods. First, from a physical-mechanisms perspective, it summarizes the characteristic relationships between lightning activity and various meteorological factors. Regarding the dynamical environment, previous findings are reviewed on how updrafts, Convective Available Potential Energy (CAPE), and vertical wind shear affect charge separation processes and modulate lightning frequency. Regarding microphysical characteristics, existing studies on collision-induced charging mechanisms between ice crystals and graupel particles are synthesized, along with conclusions on the roles of hydrometeor content and precipitation intensity in characterizing lightning intensity. The spatial distribution patterns of lightning activity along the margins of convective systems and within tropical cyclones are also discussed. Second, this paper classifies and reviews technical approaches for constructing lightning nowcasting models based on single-source observational data. For radar observations, it summarizes applications of traditional centroid tracking, optical flow algorithms, and deep learning methods—including temporal networks, Generative Adversarial Networks (GANs), and conditional flow matching—in radar echo extrapolation and lightning retrieval. For satellite observations, it reviews progress in identifying convective initiation and estimating lightning occurrence probability based on cloud-top evolution features. In addition, it summarizes the effectiveness of ground-based lightning location networks and atmospheric electric field mills in local ultra-short-term warning applications. On this basis, the paper further reviews studies on cross-modal fusion techniques that integrate multi-source heterogeneous observations with Numerical Weather Prediction (NWP) models, three-dimensional semantic segmentation models, and next-generation meteorological foundation models. Existing findings indicate that multi-source fusion approaches can compensate for the limitations of single-source observations, effectively extend forecast lead time, and improve the capability of refined probabilistic forecasting. Finally, this paper summarizes current challenges in lightning forecasting, including sample imbalance, weak regional adaptability of models, and insufficient physical interpretability. It also provides an outlook on future development directions for integrated intelligent probabilistic forecasting technologies that combine physical constraints with data-driven approaches.