with Joel Fergurson and Jay Sayre
Work in progress
We develop a procedure to feasibly produce remotely sensed agricultural outcome measures (such as crop yields) using publicly available survey data and large scale satellite imagery. Our approach builds upon the methodology of You et al. (2017) to introduce a dimensionality reduction technique that allows us to train a convolutional neural network in a setting in which the training data are much more aggregated than the level of the satellite imagery, yet maintains the temporal and spatial features necessary for accurate crop discrimination. Additionally, we develop a methodological toolkit that can be used by other researchers to select, download, query, and run analyses on large scale sets of imagery where such images must be stored locally on computing clusters. The primary application of our procedure is to predict maize yields in Mexico, using high resolution satellite imagery from Planet Laboratories obtained for all of the western states in Mexico. We show that our technique performs well at predicting yields at the close-to-farm level using validation microdata, despite being trained only on aggregated agricultural survey data.