Small Area Population estimation via Random Forest and Dasymetric Methods

Keywords: gridded population, dasymetric mapping, small area estimation, California, random forest
Session Type: Virtual Paper
Day: Thursday
Session Start / End Time: 4/8/2021 08:00 AM (Pacific Time (US & Canada)) - 4/8/2021 09:15 AM (Pacific Time (US & Canada))
Room: Virtual 19

Abstract

The quality of top-down gridded population models positively correlates with detailed administrative boundaries and high-resolution supporting data. Because highly detailed ancillary data is not consistently available at geographies of fine scale, gridded population models rarely exceed 100m resolution. However, the increasing availability of remotely sensed building footprints, high-quality group quarters data, and rapidly improving image sharpening techniques make sub-100m estimates more feasible. This research adapts a Random Forest model and GQ inclusive Cadastral Expert Dasymetric System (CEDS) to estimate small area human population distribution for case studies in California. It was found that random forest based models could be used to enhance the CEDS methodology, particularly in areas of recent construction and rural areas where metadata is lacking. Maps are produced at 30m resolution adjusted to E1 City/County estimates, and are made available at increasing scale via the USNG.