Full Citation
Title: Transfer Learning with CNNs: Predicting Income per Capita and Population Levels in England at Fine Spatial Scales using Satellite Imagery and US-trained Models
Citation Type: Dissertation/Thesis
Publication Year: 2023
ISBN:
ISSN:
DOI:
NSFID:
PMCID:
PMID:
Abstract: Deep learning has proven to be useful for predicting economic variables at fine spatial scales using satellite imagery and can be used to predict data that is challenging to obtain, coarse, or released infrequently. However, training and tuning these models have substantial computational and time requirements. If transfer learning could be effectively applied to these deep learning methods, this would allow various stakeholders to obtain accurate predictions without the need for training models themselves. This paper explores this possibility, where predictions of income per capita and population levels at fine spatial scales of England are made using satellite images and a convolutional neural network trained on US data. The predictions achieve negative R 2 values, however, this paper further employs a robust model stacking approach using an ordinary least squares regression, which results in R 2 values of 0.051 and 0.460 for income per capita and population levels respectively. Furthermore, this paper finds that the predictions are heteroskedastic, where the residuals increase as the actual values increase. The implications of these results are that transfer learning has the potential to be successfully applied to deep learning methods that predict economic variables, however , further research is required in finding improved model stacking approaches and more adaptable deep learning methods for predicting economic variables in different geographic contexts.
Url: https://thesis.eur.nl/pub/67917
User Submitted?: No
Authors: Merwe, Michael van der
Institution: Erasmus University Rotterdam
Department: Erasmus School of Economics
Advisor:
Degree:
Publisher Location:
Pages: 1-26
Data Collections: IPUMS NHGIS
Topics: Methodology and Data Collection, Population Data Science
Countries: