Development of Remote Sensing Software Using a Boosted Tree Machine Learning Model Architecture for Professional and Citizen Science Applications

Abstract

Remote sensing has rapidly gained significance in environmental science research over the past twenty years, especially when it comes to assessing the urbanizing landscape. As urban sprawl continues to grow, the role of citizen science in environmental studies has also gained greater importance by providing a broader range of study routes. The current study aims to bridge the gap between remote sensing of urbanization and the ability of citizen scientists to contribute to our expanding knowledge base. We developed a machine learning model-based architecture that makes use of pre-trained models to estimate pervious and impervious surface percentages within a user-defined region. Pre-trained model-based architectures provide greater ease-of-use and can be made more accessible than softwares that require manual supervised learning (e.g. ArcGIS). Therefore, we have developed an iOS application called Tar Print that utilizes model-based architecture, with a boosted tree algorithm and per-pixel classification of publicly available satellite imagery. We optimize the iterative per-pixel design by extracting features about neighboring and contrasting pixels, which are conventionally under-utilized, to provide the machine learning framework with more features. The Tar Print version 1 model includes six training classes and nine features that are sorted into both pervious and impervious categories. We used 2,250 data points to train this model, which achieved a 100% training score, 98% testing score, and an 86% validation score. We are currently developing a version 2 model that is trained over 50,000 data points and nine training classes. This research expands the tools citizen scientists and professionals can use to accurately monitor urban development on a larger scale.