Applying Machine Learning Algorithms on Urban Heat Island (UHI) Dataset

Climate change worldwide is a huge challenge, and urban heat island (UHI) is being explored as one of the contributors to this challenge. UHI is an urban or rural area with a temperature variance than its neighbouring areas. Researchers can model the UHI data and predict the temperature change using various relative parameters of UHI. The land surface temperature (LST) data and its co-related parameter of the study area, i.e. Srinagar City, JK, India, has been extracted from satellite imageries. LST data of the study area is assessed to understand the evolution to help analyse the UHI effect and its variance. The LST data was extracted through MODIS Satellite, from 2001 to 2020, with an 8-day revisit time/peak month of the season. In having a voluminous dataset, i.e. 16 sampled LST data/each km²/year measured in Kelvin(k), various machine learning algorithms were applied on LST data to establish relations for UHI modelling. Unsupervised machine learning algorithms were used on continuous LST data to define clusters and further standardized/compared with existing scientific classifications of the study area. The number of clusters was tweaked to determine the best-case scenario. Additionally, correlation and regression were applied to determine if there is multicollinearity amongst the LST data. The outcome of two analyses was used to build a UHI framework on a structured UHI dataset. Performance of algorithms in predicting UHI parameters like urban, vegetation and wetlands zones varied considerably. Naive Bayes and support vector machine did considerably well in predicting wetlands but failed to perform impressive accuracy for urban and vegetation zones. Random forest, gradient boost tree and probabilistic neural networks failed in predicting wetlands. Neural networks have performed worst in predicting wetlands, having a prediction accuracy of around meagre 5%, while the decision tree algorithm has performed well in all three zones.