GranoScan: an AI-powered mobile app for in-field identification of biotic threats of wheat.

Abstract

Capitalizing on the widespread adoption of smartphones among farmers and the application of artificial intelligence in computer vision, a variety of mobile applications have recently emerged in the agricultural domain. This paper introduces GranoScan, a freely available mobile app accessible on major online platforms, specifically designed for the real-time detection and identification of over 80 threats affecting wheat in the Mediterranean region. Developed through a co-design methodology involving direct collaboration with Italian farmers, this participatory approach resulted in an app featuring: (i) a graphical interface optimized for diverse in-field lighting conditions, (ii) a user-friendly interface allowing swift selection from a predefined menu, (iii) operability even in low or no connectivity, (iv) a straightforward operational guide, and (v) the ability to specify an area of interest in the photo for targeted threat identification. Underpinning GranoScan is a deep learning architecture named efficient minimal adaptive ensembling that was used to obtain accurate and robust artificial intelligence models. The method is based on an ensembling strategy that uses as core models two instances of the EfficientNet-b0 architecture, selected through the weighted F1-score. In this phase a very good precision is reached with peaks of 100% for pests, as well as in leaf damage and root disease tasks, and in some classes of spike and stem disease tasks. For weeds in the post-germination phase, the precision values range between 80% and 100%, while 100% is reached in all the classes for pre-flowering weeds, except one. Regarding recognition accuracy towards end-users in-field photos, GranoScan achieved good performances, with a mean accuracy of 77% and 95% for leaf diseases and for spike, stem and root diseases, respectively. Pests gained an accuracy of up to 94%, while for weeds the app shows a great ability (100% accuracy) in recognizing whether the target weed is a dicot or monocot and 60% accuracy for distinguishing species in both the post-germination and pre-flowering stage. Our precision and accuracy results conform to or outperform those of other studies deploying artificial intelligence models on mobile devices, confirming that GranoScan is a valuable tool also in challenging outdoor conditions.

Plant phenotyping relevance

小麦の葉・穂・茎・根の病害や損傷を画像から認識するAIモバイルアプリの開発・性能評価が中心であり、植物の病害状態を直接推定するフェノタイピング手法に該当する。害虫・雑草識別も含むが、病害認識の技術的評価が明示されている。

Code and data availability

The article's data availability statement explicitly states that the authors' weed phenotyping image dataset is publicly available on Zenodo (DOI 10.5281/zenodo.7598372), a paper-specific public asset. No author analysis code or trained model checkpoints are described with a public URL.