Lightning is formed in the atmosphere through the combination of intricate dynamic and microphysical processes. Mainly due to this complexity, attempts to solve the important problem of lightning prediction have generally failed to yield accurate results. Furthermore, current prediction systems are slow and very complex, and they require expensive external data acquired by radar or satellites. In this thesis, we propose a machine learning approach to provide early lightning warnings using data that can be obtained from any weather station. We focused on a small area (usually around a critical infrastructure) targeting a higher spatiotemporal resolution and accuracy. The model is customized for each area of interest to account for the variation of the lightning activity pattern, driving mechanisms, and local conditions from one site to another, hence providing tailor-made, site-specific lightning warnings. The algorithm uses four local atmospheric parameters that can be acquired with readily available sensors to produce excellent prediction of the occurrence of lightning during three subsequent ten-minute intervals and within a radius of 30 km. It allowed to successfully hindcast lightning hazards using single-site ground-based observations from one of 12 stations selected in Switzerland. Being independent of external data sources such as radar, satellite, and weather model outputs, the algorithm can be used with commodity weather stations to scale up their functionality as an early lightning warning system. That means we can co-ver remote regions that are out of radar and satellite range and where communication networks are unavailable. In addition to predicting "when" lightning will occur, determining "where" it struck (localization) is also important in a wide range of research and application domains, including geophysical research, lightning warning, aviation/air traffic, weather services, insurance claims and power transmission and distribution, etc. For example, some lightning warning systems rely on such data to indicate approaching thunderstorms and thus to prevent catastrophic effects of lightning strikes to critical infrastructure, sensitive equipment or systems, and outdoor facilities. Localization is key not only to lightning safety but also to data collection for high-resolution nowcasting with Ma-chine Learning. Over the years, researchers have explored different approaches to lightning localization. However, all current approaches require the presence of multiple RF sensors and, hence, are not applicable in many practical scenarios where the size and power consumption of the device matter. We broke this barrier by developing two machine learning-based lightning location systems. The first one uses the lightning-induced voltage measurements from two preinstalled sensors on transmission lines to localize nearby lightning. The second works on a data from a single electric field sensor to estimate the 2D geolocation of the lightning strike point. The key to reducing the sensor array size to one is hybridizing a powerful RF imaging technique called Electromagnetic Time Reversal (EMTR) with advanced techniques in computer vision and machine learning. Both localization models have been trained and optimized using fully synthetic data.