People's olfactory perception potential mapping using a machine learning algorithm: A Spatio-temporal approach

Abstract

This research aimed to conduct Spatio-temporal modeling of people's olfactory perception in the Tehran city using a machine learning-based approach considering the importance of people's perception of the urban environment and the effect of the sense of smell in strengthening or weakening the environmental and urban spaces. Therefore, a spatial database containing dependent and independent data was established to model and prepare a potential map of people's olfactory perception in Tehran. The dependent data were collected through a questionnaire among 687 people, which contains three states of sense of smell, including neutral (350 cases), pleasant (247 cases), and unpleasant (90 cases). The independent data consisted of distance to public transport stations, livestock and poultry production centers, streets, fuel stations, fruit and vegetable fields, restaurants, industrial areas, garbage volume, traffic volume, population density, normalized difference vegetation index (NDVI), meteorological data (wind speed and direction, and humidity), and air quality index (AQI). The last three criteria were provided temporarily in four seasons. A random forest (RF) algorithm was utilized to model and prepare the potential map of people's olfactory perception in four seasons. Generally, 70% of the information was dedicated to modeling, and the remained 30% was applied for validation. Results of the Gini index revealed the more significant impact of the criteria of distance from fruit and vegetable fields and distance from restaurants in neutral smells, distance from streets and distance from restaurants in pleasant smells, and distance from livestock and poultry production centers and garbage volume in unpleasant smells among spatial criteria. One of the essential spatial-temporal criteria was wind direction in all seasons for neutral smells, in summer, fall, and winter for pleasant smells, and in fall for unpleasant smells. The RF algorithm modeling results indicated the greater effect of distance to public transport stations in spring, garbage volume in summer and winter, and population density in autumn on the sense of smell. The potential maps of olfactory perception were evaluated using the receiver operating characteristic (ROC) curve and area under the curve (AUC). AUC values were calculated for the spring, summer, autumn, and winter seasons as much as 0.956, 0.955, 0.957, and 0.958 for neutral, 0.944, 0.928, 0.946, and 0.942 for pleasant, and 0.922, 0.931, 0.963, and 0.981 for unpleasant states, respectively. The potential maps of this study help present a more pleasant olfactory space to urban users by urban planners and managers.