Name of the new projectEstimating agricultural development indicators over large areas from satellite images – an approach using convolutional neural networks and transfer learning
|Title||Name of the new projectEstimating agricultural development indicators over large areas from satellite images – an approach using convolutional neural networks and transfer learning|
|Summary||In this project you will use deep learning models, specifically convolutional neural networks (CNN), to analyze satellite imagery to estimate agricultural development indicators.|
|References|| Xie, M., N. Jean, M. Burke, D. Lobell & S. Ermon (2015) Transfer learning from deep features for remote sensing and poverty mapping. arXiv preprint arXiv:1510.00098.
Jean, N., M. Burke, et al (2016) Combining satellite imagery and machine learning to predict poverty. Science, 353, 790-794.
|Prerequisites||Good knowledge of machine learning, convolutional neural networks and programming skills for implementing machine learning algorithms|
|Supervisor||Mattias Ohlsson, Thorsteinn Rögnvaldsson|
In this project you will use deep learning models, specifically convolutional neural networks (CNN), to analyze satellite imagery to estimate agricultural development indicators. This project is connected to an overall aim of enhancing our understanding of the pace of agricultural and rural transformation in contemporary sub-Saharan Africa, its poverty and distributional impacts and drivers.
Due to the lack of large labeled data sets, the approach will build upon transfer learning where day-time satellite images will be used to estimate the distribution of light during night (nighttime lights) or vegetation indices (VI). This idea builds upon the connection between nighttime lights and economic activity, and the fact that VIs can be used as proxy for biomass (and thus yield). The features extracted from the trained CNNs can then be used to estimate indicators related to e.g. poverty, structural transformation, and settlements. In this project we will focus on indicators that measure the presence of a settlement, eg. are there buildings or not?
This project build upon a collaboration between Halmstad University and the Department of Human Geography, Lund University (LU). LU will provide all necessary data (satellite imagery and indicators).