Abstract

This article examines the construction of image recognition algorithms for the classification of plant pathology problems. Rooted in science and technology studies research on the effects of agricultural big data and agricultural algorithms, the study ethnographically examines how algorithms for the recognition of plant pathology are made. To do this, the article looks at the case of a German agtech startup developing image recognition algorithms for an app that aims to help small-scale farmers diagnose plant damages based on digital images of their symptoms. The study posits that the construction of these algorithms can be grasped as a succession of layers, at each of which the startup’s employees carry out different selection practices. It is argued that these practices gradually inscribe a selective recognition of the phenomenon of plant pathology into the algorithms of the app. This selective recognition is reflected in the fact that the emerging algorithms are effective in identifying isolated plant damages on isolated crops, while ignoring most agroecosystem-related actors and relations through which plant damages arise. The article concludes that the app’s selective recognition of plant pathology is likely to perpetuate or even exacerbate the pesticide use of its users. This is because the app’s view of plant pathology as a phenomenon comprised of isolated plant damages on isolated crops is more compatible with the use of chemical pesticides than with approaches to crop protection that strive for a more moderate use of these substances (e.g., integrated pest management, agroecology, organic agriculture).