Post-Harvest Loss Prediction

Problem

Post-harvest loss along smallholder tomato supply chains is both an economic and a food-security failure. It is also uneven: some nodes lose far more than others, and interventions need to know where.

Context

A web-based real-time application for predicting post-harvest losses along the tomato supply chain in Kibwezi West sub-county, combining statistical modelling with an operational interface.

Dataset

Supply-chain observations, spatial units and loss indicators from the Kibwezi West tomato value chain.

Methodology

Predictive modelling of loss along the chain, with spatial structure taken seriously rather than treated as a nuisance. The application layer made forecasts available in near real time.

Model / architecture

Statistical and machine-learning models for loss prediction along supply-chain nodes.

Results

A decision-support surface identifying where loss concentrates along the tomato supply chain.

Evaluation metrics

Geography

Kibwezi West

Commodity

Tomato

Challenges

Sparse observation at some chain nodes, confounding between weather, handling and market delay, and the need for a tool extension officers would actually open.

Impact

A decision-support surface for where loss concentrates, typical of Jeff's preference for models that leave the notebook.

Tools / technology

• R
• Python
• Spatial analytics
• Web application