WEED CLASSIFICATION ON MOBILE DEVICES VIA OFFLINE INFERENCE: ASOFTWARE ENGINEERING APPROACH

Abstract

Computer vision techniques are a viable approach for the on-site evaluation of environmental conditions unfavorable to crop development, such as monitoring and identifying weeds. In this context, this work presents PlantScan, a system for Android platform designed for visual weed classification, adapted to the paradigm of edge computing to mitigate the effects of limited internet connectivity. The system integrates a deep learning model
based on the ResNet-50 architecture — with weights originally trained and made available in a previous study in the DeepWeeds dataset — converted to the TensorFlow Lite format, making it suitable for on-device inference on Android devices. The tool also incorporates georeferencing features for mapping the distribution of detected species. Consistent with the baseline study, the classifier recognizes eight weed classes from the DeepWeeds repository. Furthermore, when evaluated on the official test subset of the same repository, the model achieved an accuracy of 96.07% on the official test subset. These results demonstrate the feasibility of deploying complex machine learning models on mobile devices and integrating offline agronomic diagnostics into field workflows.