Efficacy of Season Prediction for Geo-locations Using Geo-tagged Images

Tens of thousands of pictures are taken at different locations throughout the year. People often visit places and take pictures to remember their visits. We believe that the seasonal travel patterns of people to specific locations will create a correlation between a location and the season of the images taken in that location. For example, fewer people visit Bear Valley, California during the summer than during the winter as it is a popular Ski Resort. Therefore, we believe we will find more pictures of Bear Valley taken during winter when compared to any other seasons. Today, most of the photographs have geo-location (Latitude and Longitude) and time when the photograph was taken included in their metadata. Given the distribution of photographs, correlations between locations and seasons found using this metadata could potentially be used to develop a system to predict the best time of year to visit particular locations or number of people who may visit a tourist destination in the next season allowing business establishments to prepare appropriately. In this work we evaluate the efficacy of using this metadata to predict the season given the location. In our work, using a dataset comprised of photograph metadata, we focused on approximately 1.1 million photographs taken in California. Using variations of the nearest neighbor algorithm, we were able to predict the season of a photograph with a maximum correctness of 80.9% with a sufficiently large training set. We experimented with both weighted K-Nearest Neighbor (K-NN) and Fixed Radius Nearest Neighbor (FRNN) using no-weight, inverse, logarithmic, and gaussian weight calculators. Using the K-NN model, we found that logarithmic weighted K-NN performed the best at 79.55% correctness. Using the Fixed Radius NN model, we found that the gaussian weighted model using standard deviation of 0.0001 performed the best at 80.9%.