Analysis and prediction of urban ambient and surface temperatures using internet of things

The high temperature condition experienced in urban areas as compared to semi-urban and rural areas creates an urban heat Island (UHI). Formation of UHI in a city should be avoided as it results in higher surface and ambient temperatures, leading to an increased demand for energy, high carbon emission and effects community health in general. Traditional ways of collecting such data is limited through weather stations, and hence such collections are limited to few parts of urban regions. With the advent of new system-on-chip based devices, wireless sensors with mobile OS communication, modern smartphones with a variety of environmental sensors, different communication modules, cloud based storage with data analysis and visualization facilities, it is now possible to develop low cost novel IoT systems to monitor, assess and analyse urban temperatures and their subsequent effects, anywhere and anytime. In this work, a location aware, mobile-sensing enabled IoT system has been designed to monitor, measure and analyse the presence and intensity of UHI effect in various geographical regions. Atmospheric temperatures have been used to predict surface temperatures. Multiple linear, quantile, support vector regression and extreme learning machine based prediction models have been evaluated for forecasting and results have been compared to examine their efficacy. The mean absolute error and the mean absolute percentage error values for the SVR model are the lowest, indicating that the SVR technique has the maximum accuracy. The mean square and the root mean square error values are the smallest for the quantile regression model, showing that the quantile regression approach produces the fewest large errors. All prediction models contain positive biases, according to the MPE values, with the quantile regression technique having the least. For training and testing, the worst-case time complexity of the multiple linear, quantile regression, and ELM models is found to be O(n), indicating that these models are faster than SVR, which has the worst-case time complexity of $$O(n^{3})$$ O ( n 3 ) .