Real-time on-device weed identification using a hardware-efficient lightweight CNN

Abstract

Accurate and timely weed identification is fundamental to sustainable crop management, particularly for autonomous agricultural systems operating under strict energy and hardware constraints. While deep learning has significantly advanced image-based weed recognition, most existing models rely on GPU-based inference and therefore cannot be deployed directly in low-power field devices. In this study, we propose a hardware-efficient lightweight convolutional neural network (CNN), named TinyWeedNet, designed specifically for real-time on-device weed identification in precision agriculture. The model integrates multi-scale feature extraction, depthwise separable inverted residual blocks, and compact channel attention to enhance discriminative ability while maintaining a minimal computational footprint. To evaluate its suitability for field deployment, TinyWeedNet was trained and tested on the public DeepWeeds dataset and implemented on an STM32H7 microcontroller via the TinyML workflow. Experimental results demonstrate that the model achieves 97.26% classification accuracy with only 0.48 M parameters, supporting sub-90 ms inference and low energy consumption during fully embedded execution. A comprehensive analysis, including benchmark comparisons, hyperparameter sensitivity tests, and ablation studies, demonstrates that TinyWeedNet provides a good balance of accuracy, speed, and energy efficiency for resource-constrained agricultural platforms. Overall, this work demonstrates a practical pathway for integrating real-time, low-power weed identification into field robots, UAVs, and distributed sensing nodes, contributing to more autonomous and energy-aware weed management strategies in precision agriculture.