Time-Series Estimation of the Best Time to View Seasonal Organisms Using Geotagged Tweets and Co-occurring Words

The meteorological agency has reduced the scale of its biological seasonal observations because of requirements for maintaining observation targets and securing personnel. A low-cost, nationally applicable estimation method is being sought as a substitute. Earlier research has produced cherry blossom viewing estimates for a certain period in the future by predicting the transition of tweet counts on Twitter. One challenge of earlier research was the decline in prediction accuracy obtained in areas that produced few tweets. Additionally, it is difficult to estimate the viewing season in areas with few tweets, where previous methods cannot infer an increasing or decreasing trend. This paper therefore presents a proposal of a method to increase the number of data used for viewing season estimation for a certain period in the future using co-occurrence words. This method improves prediction accuracy compared to previous methods and demonstrates the possibility of estimating the viewing season in more areas than those estimated using previous methods.