Image Classification to Determine the Level of Street Cleanliness: A Case Study

The street cleanliness of big cities has a high impact on urban environment and health, so cities invest lots of effort to make their streets clean. With the recent advances in technology, it is feasible to develop smart systems for monitoring street cleanliness at scale such as automatic classification of street images to identify the cleanliness level by utilizing conventional classifiers (e.g., Naive Bayes and AdaBoost). However, these baseline methods do not provide a desired classification accuracy in practice. Thus, we propose a geo-spatial classification approach to enhance the classification accuracy. In particular, since the crowdsourced images are tagged with geo-location (i.e., GPS coordinates), we devise a novel framework with multiple local trained models exploiting the similarity of local images so that the proposed models learn better street image classification for each geographical region. This paper also presents a case study of street cleanliness classification using a large real-world geo-tagged image dataset obtained from Los Angeles Sanitation Department (LASAN). Experimental results showed that our framework was able to achieve an F1 score of around 0.9.