A deep learning model to predict food losses at the national and sub-national levels

This study explores the use of FAO food balance sheet and world bank data to train a feedforward neural network model to predict national and sub-national food loss (FL), with rice as a proof-of-concept example. The model demonstrates strong predictive performance, highlighting the utility of combining economic, climatic, and agricultural predictors. An ablation study reveals the importance of factors such as "Year," "Precipitation," and "GDP per capita" in capturing essential patterns, while features like "Population" and "FAO Region" show limited relevance. Validation with other crops such as wheat and apple affirms the model’s adaptability, though with reduced accuracy. The study identifies limitations tied to data dependencies, geographic scope, and omitted variables like infrastructure and supply chain factors, which may influence FL predictions. Despite these constraints, the proposed AI-driven approach provide us with a potentially scalable and cost-effective alternative to traditional data collection methods, which can particularly benefit resource-constrained regions. Future enhancements, including model finetuning and the integration of domain knowledge, will help to further improve predictive accuracy and broadening applicability.