Utilizing Land Surface Data for Air Temperature Prediction in Urban Regions through Machine Learning
Board Location: #69
Discipline: Computer Sciences and Information Management
Subcategory: Geosciences and Earth Sciences
Session: 3
Taseen Islam - CUNY Macaulay Honors College
Co-Author(s): Kip Nielsen, NASA GISS, University of Kansas, Lawrence, KS; Shaunak Sharma, NASA GISS, South Brunswick High School, South Brunswick, NJ; Ashley Grey, NASA GISS, Baruch College Campus High School, New York, NY; Audrey Lofthouse, NASA GISS, Brigham Young University, Provo, UT; Hamidreza Norouzi, CUNY New York City College of Technology, Brooklyn, NY; Reginald Blake, CUNY New York City College of Technology, Brooklyn, NY
Accurately forecasting air temperature is vital for gaining insights into atmospheric phenomena. However, the limited availability of air temperature data presents significant challenges for conducting comprehensive analyses and predictions. Land surface temperature (LST) and air temperature are often shown to be correlated, but air temperature data is not as widely available as LST data is, making coupled analysis and predictions difficult. NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS) satellite provides continuous global coverage of LST data. However, Automated Surface Observing Systems (ASOS) weather stations that capture air temperature data are only available at specific points, creating a lack of data between ASOS stations where only MODIS data is available, especially in urban areas. A supervised machine learning model was developed using the K-Nearest-Neighbors (KNN) regression algorithm to model air temperature trends and predict air temperature values when given land surface data from the MODIS satellite. 227 weather stations spanning multiple cities across 125 countries around the world were randomly selected. LST data was obtained from MODIS observations at 1 km resolution at the weather station locations. These stations were split into training and testing sets before a variety of feature engineering methods like interpolation and scaling were applied to them. The model effectively utilizes LST data from both Terra and Aqua MODIS satellites, considering both day and night observations, in conjunction with Normalized Difference Vegetation Index (NDVI) data from Terra and Aqua to produce air temperature value predictions. It performs with an average accuracy of 85%, RMSE of 4.27 °C, and a median absolute error of 3.11 °C. It is able to accurately predict and model air temperature data for the US and other countries and is even able to model air temperature during times where ASOS air temperature data is not available, filling data gaps in regions without recorded air temperature readings. This capability presents a unique opportunity to generate comprehensive air temperature predictions for urban regions with heterogeneous land cover types. The model has the potential to continue predicting future air temperature values across the world, fill in gaps in data, and provide spatial heat maps of air temperature estimates to create a more comprehensive understanding of air temperature in urban settings.
Funder Acknowledgement(s): National Science Foundation REU Program
Faculty Advisor: Dr. Hamidreza Norouzi, HNorouzi@citytech.cuny.edu
Role: I was the machine learning lead on the team. Since everyone else on the team is in environmental sciences or civil engineering, I wanted to incorporate machine learning into the project, taking the original idea and transforming it into a more ambitious project with ML. Our team worked on multiple projects, so I started and led the ML portion of the team. I created the problem statement, found and shared relevant literature with the team, planned and executed experiments, programmed the ML model, tested it, analyzed it, found all the results, and made all the images in the poster. I got help retrieving information from the satellites and received advice on which input parameters would be useful in my model, but model selection, data preprocessing, and all other ML techniques were researched and executed by me.