Overview: This paper transfer an ImageNet pertained CNN to predict regional poverty level based on the satellite images.
Transfer learning: There are two steps in the transfer learning process:
a) from ImageNet to nighttime lights: The authors use NOAA annual nighttime images as the intermediate training target . To accommodate the label (light intensity from 0 to 63), two methods are used: random crop and FCN + linear transform.
b) from nighttime lights to poverty estimation: The same set of feature representation will be used to predict the poverty rate (no new features are trained)
Evaluation and discussion: The biggest take-away of this paper (from my point of view) is the evaluation and discussion. The paper evaluate the results from following perspectives:
Comparison with non-transfer models: The transferred models beats raw ImageNet features, light features and the combination. It is not beating the survey data (a survey that provides the feasible features from remote sensor such as roof type, house type, etc), which is the upside of the satellite image models.
Comparison with nightlights data: The paper compares the performance of transferred model and nightlights data only model in across the poverty rate quantile. In the poorest region, the nightlights are almost 0 and does not provide much information.
Random experiments: To illustrate the result is significant, a random experiment was carried out where wrong inputs are used for training. The true R-squared way out of its empirical distribution.
Across country transferring: The authors also shows the model transfer well across country boundaries.