Journal of Artificial Intelligence, Virtual Reality, and Human-Centered Computing
Temperature Prediction with Satellite Imagery Features of New York City Using Machine Learning
Abstract
Junrong Wang, Gladys Elizabeth Kenyon, Lingbo Liu and Shuming Bao
This project applies satellite imagery to study the temperature changes in cities with an advanced machine learning approach. For extracting features, we employ the MOSAIKS (Multi-Task Observation Using Satellites and Kitchen Sinks) which comprises the RCF (Random Convolution Features) version which is an unsupervised learning algorithm, deciding which features to extract [1]. Then Xtreme Gradient Boosting or Random Forest are used to predict temperature from features. Specifically, two RCF architectures and multiple regression methods are compared systematically to evaluate overall performance and interpret-ability, contributing to Artificial Intelligence. Multiple methods performs differently in various image sizes, revealing a new relationship between methods and image sizes, and guiding method choices for specific temperature prediction. Existing satellite-based temperature studies rely on physics-based LST (Land Surface Temperature) retrieval algorithms (mono-/split-window) combined with empirical regressions and sensor data, unsuitable of prediction in design [2,3]. Some research on machine-learning (Random Forest, gradient boosting, Decision Tree, Artificial Neural Network, Convolutional Neural Networks, Fast Library for Approximate Nearest Neighbors) depend on complex sensor data inputs, unsuitable for prediction from just imagery [4-6]. Despite of improved accuracy, they are computationally demanding, offering limited interpretability, whereas more transparent regression models underperform and hardly scale globally. In contrast, our approach requires only imagery and achieves relatively high accuracy [7,8]. More importantly, using our methods, solely satellite imagery can yield reasonable results, useful for designers to generate temperature predictions from fake satellite imagery they design, benefiting engineering.