Drought constitutes a major threat to food security and the sustainability of agricultural systems in tropical developing countries. These systems are highly vulnerable due to their strong dependence on rainfall, their low resilience to climatic hazards, and the scarcity of reliable data. This study proposes a methodological framework for assessing and forecasting agricultural drought exposure in semi-arid tropical regions, using Haute Matsiatra in Madagascar as a case study. The framework integrates multi-source geospatial data with artificial intelligence to construct and predict a drought exposure index. Climate data are retrieved from the NASA POWER LARC platform and local institutions, while Landsat satellite imagery (30 m resolution) is sourced from the USGS platform to derive radiometric indices. The approach consists of three main steps: (i) selection and preprocessing of climate variables, satellite imagery, and radiometric indices such as NDWI and SPEI; (ii) development of a composite exposure index through weighted aggregation, incorporating local expertise using Saaty’s Analytic Hierarchy Process (AHP); and (iii) forecasting the index with a Long Short-Term Memory (LSTM) recurrent neural network, using horizons of 6 to 36 months aligned with the local agricultural calendar. To adapt to these horizons, daily data are aggregated to reduce noise due to data granularity. The outputs will be presented as thematic maps illustrating exposure levels under different horizons and climate scenarios. This framework offers a decision-support tool for climate risk management and can be replicated or adapted to other regions facing similar vulnerabilities.