In some embodiments, data from multiple vehicle-based natural gas leak detection survey runs are used by computer-implemented machine learning systems to generate a list of natural gas leaks ranked by hazard level. A risk model embodies training data having known hazard levels, and is used to classify newly-discovered leaks. Hazard levels may be expressed by continuous variables, and/or probabilities that a given leak fits within a predefined category of hazard (e.g. Grades 1-3). Each leak is represented by a cluster of leak indications (peaks) originating from a common leak source. Hazard-predictive features may include maximum, minimum, mean, and/or median CH4/amplitude of aggregated leak indications; estimated leak flow rate, determined from an average of leak indications in a cluster; likelihood of leak being natural gas based on other indicator data (e.g. ethane concentration); probability the leak was detected on a given pass; and estimated distance to leak source.