Predicting Postharvest Food Losses at National and Sub-National Levels Using Data-Driven and Knowledge-Based Neural Networks

Abstract

Food loss is a major challenge for global food security, resource use efficiency, and sustainability. However, collecting primary food loss data is costly. This study explores a neural network-based approach to estimate food loss in the postharvest stage using the FAO’s food balance sheets for proof of concept. We investigated both traditional data-driven feedforward neural networks (FNNs) and knowledge-informed neural networks (KiNNs) using rice, wheat, and apple data from the FAO’s food balance sheets. The results show relatively high prediction accuracy with the data-driven approach when a larger amount of data is available. It also demonstrates the high potential of using KiNNs to improve the prediction accuracy when data availability is relatively limited. In general, the proposed approach shows great potential to be developed into an effective supplementary tool that can partially replace costly primary food loss data collection at the postharvest stage, which is particularly valuable when resources for primary data collection are limited.