Wildfires have significant effects on human populations worldwide. Smoke pollution, in particular, from either prescribed burns or uncontrolled wildfires, can have profound health impacts, such as reducing birth weight in children and aggravating respiratory and cardiovascular conditions. Scarcity in the measurements of particulate matter responsible for these public health issues makes addressing the problem of smoke dispersion challenging, especially when fires occur in remote regions. Previous research has shown that in the case of the 2014 King fire in California, crowdsourced data can be useful in estimating particulate pollution from wildfire smoke. In this paper, we show that the previous model continues to provide good estimates when extended statewide to cover several wildfires over an entire season in California. Moreover, adding the semantic information contained in the social media data to the predictive model significantly increases model accuracy, indicating a confluence of social and spatio-temporal data.