Florian Teste

Several approaches based on statistical, mechanistic, agronomic, and economic models have been developed and are currently used to forecast agricultural commodity prices by market analysts and government agencies. Most of the time, these approaches rely on regional crop production values, either estimated before harvest or measured after, as well as on estimates of demand for food goods. In practice, however, obtaining reliable data on regional crop production early enough before harvest is difficult to predict price changes, partly because of the multiple environmental conditions that influence regional crop production and the associated uncertainties. Even after harvest, regional production data are often unreliable in many key producing regions. It is also difficult to accurately estimate the demand for food goods due to many factors involved.

Satellite data products (e.g. derived from MODIS) offer interesting alternative sources of information for estimating local environmental conditions at high resolution. As satellite images are linked to land and vegetation characteristics, they are potentially usable for forecasting changes in the prices of major agricultural commodities on a global scale without the need to estimate agricultural production. Similarly, satellite information can potentially be useful for estimating food goods’ demand. Thus, this thesis aims to develop and test a methodological framework for predicting global agricultural commodity price variations based directly on high-resolution satellite data. If successful, the proposed approach will allow the prediction of price variations in real-time, avoiding the use of regional agricultural production and demand values, which are often unavailable and uncertain. Our approach will be based on machine learning algorithms capable of identifying critical features of satellite images with solid predictive capabilities.