Menu Close
 

Machine Learning to Detect Fertilizer Adulteration in Developing Countries

Project Team

  • Dr. Hope Michelson, Assistant Professor, Dept. of Agricultural and Consumer Economics
  • Dr. Andrew Margenot, Assistant Professor, Dept. of Crop Sciences
  • Dr. Ranjitha Kumar, Assistant Professor, Dept. of Computer Science

Abstract

The research objective of this proposal is to develop, deploy, and evaluate the effects of a new tool that allows small farmers in Africa to detect fertilizer adulteration at the point of purchase using cell phone app that processes phone images with a machine learning driven, image processing algorithm.

Fertilizer use remains below recommended rates in most of Sub-Saharan Africa, contributing to low agricultural productivity, pervasive poverty, and food insecurity. Smallholder farmers have voiced suspicion that fertilizer is often adulterated, and evidence suggests that these suspicions lead to inefficient fertilizer use: too much in some cases, and too little in others. These suspicions arise because the agronomic quality and efficacy of mineral fertilizer—its nutrient content—is unobservable to the untrained eye at the point of purchase. For example: though urea fertilizer should contain 46% nitrogen (by mass), that content cannot be quantified with the naked eye. This problem, compounded by weak regulation and poor enforcement of product standards, provides opportunities for cheating in the market by adulteration; and awareness of that possibility drives suspicions and thus decision-making among farmers. Recent research confirms that farmers believe that adulterated fertilizer is common in the market in Tanzania, Côte d’Ivoire, Ghana, Nigeria, Senegal and Togo (Michelson et al. 2018; Sanabria et al. 2013). Michelson et al. (2018) find that 24% of interviewed farmers in the Morogoro region of Tanzania listed purchasing high quality fertilizer among their top two concerns as they prepared for the start of a typical agricultural season; 36% of these farmers reported concerns about adulterated fertilizer in markets, with 18% believing that more than half of the fertilizer for sale was likely adulterated.

This project develops, deploys and tests the farmer-level effects of a new tool that farmers in sub-Saharan Africa can use to accurately assess the nutrient content of fertilizer at the point of purchase using only a cell phone. Our intention is to develop and implement a tool that exploits rapidly-advancing techniques in supervised machine learning to eliminate subjective fertilizer quality judgments, allowing farmers to robustly determine fertilizer authenticity solely from cell-phone images. Profitable adulteration of fertilizer requires use of fillers in quantities substantial enough for image-based detection, enabling phone image app-based detection.