Computer Science Perspectives on Spatial Analysis: Using the Segment Anything Model on High-Resolution Satellite Imagery Collections

Keywords: Computer Science, Image Processing
Abstract Type: Poster Abstract

Abstract

Spatial analysis is extremely reliant on complex computation and task automation. Studies of land cover change already utilize complex modeling and analysis, but often with labor-intensive human input, particularly for training data collection. Here, we explore the viability of Segment Anything Model (SAM), a foundational model by Meta to semi-automate the task of training data collection. We report on lessons learned from deploying computer science expertise in order to deploy SAM for imagery training data collection. Using identification of planted tree cover as a test, we report on lessons learned that allowed us to: 1) utilize computing power effectively, 2) perform version control for better reproducibility, and 3) customize analysis to minimize manual work for research. Using local servers and graphics processors on computing clusters, the time-efficiency and processing reliability of this process was greatly improved. Additionally, version control of this code in Git allowed for reproducibility of this process, avoiding more common human errors in survey research. Using Computer Science principles in this experiment allowed XXX images to be processed, adding up to a total memory size of XXX gigabytes of data, processed in XXX minutes. This approach significantly reduced the time of manual satellite data gathering and processing to get tree plot hectare counts.