kilimoAI: Participatory Machine Learning for Low-Resource Crop Disease Detection in Tanzania

Smallholder farmers in low-resource settings often face crop disease pressure without timely access to diagnostic expertise, infrastructure or locally relevant machine learning tools. This paper presents kilimoAI, a community-engaged machine learning initiative for early detection of maize and common bean diseases in Tanzania. The project combined field-based image collection, expert-supported annotation, model development, mobile deployment and farmer and extension-officer training. From November 2023 to December 2025, the project collected and curated more than 500,000 leaf images across three regions, trained four models, selected LeViT for deployment after achieving 99.89% accuracy in project evaluation, and published open maize and common bean datasets for reuse. Beyond technical performance, the experience highlights practical strategies for inclusive ML including offline first data workflows, participatory engagement, digital-literacy support, extension-officer mediation and iterative monitoring in real agricultural environments.