Identifying food insecurity situations timely and accurately is a complex challenge. To prevent food crisis and design appropriate interventions, several food security warning and monitoring systems are very active in food-insecure countries. However, the limited types of data selected and the limitations of data processing methods used make it difficult to apprehend food security in all its complexity. In this work, we propose models that aim to predict two key indicators of food security: the food consumption score and the household dietary diversity score. These indicators are time consuming and costly to obtain. We propose using heterogeneous data as explanatory variables that are more convenient to collect. These indicators are calculated using data from the permanent agricultural survey conducted by the Burkinabe government and available since 2009. The proposed models use deep and machine learning methods to obtain an approximation of food security indicators from heterogeneous explanatory data. The explanatory data are rasters (population densities, rainfall estimates, land use, etc.), GPS points (of hospitals, schools, violent events), quantitative economic variables (maize prices, World Bank variables), meteorological and demographic variables. A basic research issue is to perform pre-processing adapted to each type of data and then to find the right methods and spatio-temporal scale to combine them. This work may also be useful in an operational approach, as the methods discussed could be used by food security warning and monitoring systems to complement their methods to obtain estimates of key indicators a few weeks in advance and to react more quickly in case of famine.