Drug abuse continues to accelerate towards becoming the most severe public health problem in the United States. The ability to detect drug-abuse risk behavior at a population scale, such as among the population of Twitter users, can help us to monitor the trend of drug-abuse incidents. Unfortunately, traditional methods do not effectively detect drug-abuse risk behavior, given tweets. This is because: (1) tweets usually are noisy and sparse and (2) the availability of labeled data is limited. To address these challenging problems, we propose a deep self-taught learning system to detect and monitor drug-abuse risk behaviors in the Twitter sphere, by leveraging a large amount of unlabeled data. Our models automatically augment annotated data: (i) to improve the classification performance and (ii) to capture the evolving picture of drug abuse on online social media. Our extensive experiments have been conducted on three million drug-abuse-related tweets with geo-location information. Results show that our approach is highly effective in detecting drug-abuse risk behaviors.